Hey all — recently I wrote a three-part brain dump on everything I remember from building viral loops in the Web 2.0 days and some of the ways in which all these ideas have evolved in the mobile era and beyond. I posted it as three parts on X and brought it together as one essay with some minor formatting changes here and there but otherwise preserving it in its entirety. Hope you enjoy! -A

BRAINDUMP ON VIRAL LOOPS #1

The golden age of Web 2.0 (~2005-2010) was a special time for viral products, which were systematically engineered to reach millions of people. Back then, people were building the first versions of things we take for granted: Social networks, user-generated platforms, collaborative workplace products, messaging apps, etc.

During this time, the industry developed a comprehensive and systematic understanding of creating viral loops. There was measurement, A/B testing, and equations that needed to be satisfied. Products optimized and optimized, engineering virality. And some of the most successful of these products eventually grew to billions of users and became recognizable names.

Then all of a sudden, it ended.

And funny enough, the people who were successful at creating viral products all became billionaires, or if they didn’t create successful companies, many became FAANG executives or VC investors. Eventually, all the knowledge around building virality was basically lost. This happened as Web 2.0 ended and the mobile era began.

I thought it might be nice to write down a full brain dump of everything that I learned about viral loops during this period, because it does continue to be relevant, although the mechanics and tactics around it keep shifting and changing. Yet the overall theory is still the same, and I think it can be extended into future technology platforms and marketing channels. It also has a high degree of applicability to this era’s Product-Led Growth, all the sharing flows of gen AI apps, and referral programs that still exist within every marketplace/ecommerce product.

With these notes, I plan to cover a bunch of different ideas and topics:

• Simple viral loops that work, but then quickly disintegrate.

• Viral factor and how to calculate it.

• How to actually break down a viral loop step by step and optimize

• What increases the viral factor, and what decreases it?

• How to improve viral factor

• Case studies like content sharing, invite flows, referral loops, collaboration loops, etc.

• How retention drives viral growth.

• How do you incorporate retention into your viral factor calculation

• Why new users invite more people than existing ones.

• How to think about word-of-mouth vs. engineered virality.

• How to think about paid marketing, SEO, social media, and the effects of other channels on viral factor.

• Why mobile eventually killed the classic viral loops.

• What kinds of viral loops work in the contemporary era

• ... and much more

If you have thoughts and questions please ask. Hopefully I’ll eventually turn this into a big PDF or something that people can read if they’re interested. Or maybe a deck if we have some kind of long-form text to PowerPoint deck gen AI tool one day that’s actually good :)

Viral growth as an equation
Usually when most people talk about viral growth, they mean something silly (and non-durable), like tweeting a really cool video that then gets shared a bunch of times and drives traffic to your product. That’s not what I’m talking about. Instead, the viral loop I’m talking about is designed into a product, with invite features, tagging, referral links, and so on.

These types of viral loops have interesting properties:

• it is measurable and can be tracked

• it can be improved with product decisions

• the math applies to any form of product-driven virality (invites, shared content, referral programs, etc)

The first thing is trying to measure viral factor. The viral factor is a simple ratio -- the usual thought experiment is that if you bring a 100 users into your product, and those users invite and eventually sign up 150 users, then those will sign up 225 users and so on. This is a viral invite flow, and your viral factor in this case would be 1.5 (rarely seen in the real world). On the other hand, if your 100 users end up signing up 50 of their friends, who then sign up 25, then your viral factor is 0.5. When it’s <1 then your viral loop eventually stops working.

Very precisely, the viral factor is a RATIO. The denominator is the # of users within a time-bound cohort (like everyone who signed up from X date to Y date), and the numerator is how many users they eventually signed up via the viral loops from that cohort.

Let’s use an example to make it concrete.

Case study: Content sharing
A very common form of viral loop has an app that lets a user create something really cool (maybe that’s with AI or photo filters or otherwise). After the user creates something, they want to share it, so naturally a link is provided. Some new users receive this link, view the content, and some smaller % of these folks sign up to make their own content. At its core, this is what the new generative video and photo apps do (like Sora). But this is also the same viral loop that made Instagram successful with photo filters or what’s made blogging popular as well. After all, when you create some type of content, you naturally want to share it with people, but those people naturally may want to participate as well.

So conceptually, you might understand the smile loop, but the question is, what can you do with it to optimize?

To make it systematic, you want to track it. When someone makes a generative video, you want to track things via the URL structure:

productdotcom/vid/[video ID]?sharer_id=[sharer]

(of course, your userID is encoded as the sharer id whenever you share) Now when you share it out via email, or messaging or whatever, if an invitee gets the link, watches the video, and then signs up, you should then store the sharer’s ID alongside the user’s signup row. So imagine you’d store the usual stuff:

• id

• email

• password

• name

• sharer_id <-- this is who caused this invite to happen

Now when you look at your rows and rows of users in this table, sometimes the share ID will be blank if that user just showed up and signed up on their own. However, if they are part of a viral loop, then their share ID will point to some other user that already exists in the table.

To calculate viral factor, you then take some cohort of users, for example, the ones that joined 3 months ago. Let’s say that there are 100 of these guys. Then, what you do is you grab that list of IDs and you ask the database how many times do those IDs show up as share IDs in new sign-ups that happen afterwards? Let’s say that’s 50. Then your viral factor at that moment in time is 50/100=0.5.

What about all those users with blank sharer IDs? I sometimes think of the users who just show up as “Gen 1 users” or “onramp users” and you should just discard these guys as part of your viral factor calculation. Instead, what’s interesting is to compare the ratio of Gen 3 and Gen 2 users. Or really, Gen N+1 and N, as long as N isn’t 1. It turns out this ratio actually tends to stabilize pretty nicely over time.

The big questions
Do this with your product, then you’re bound to ask, “Wow, this is cool, but how do I make this number go up?” and in particular, “How do I make this number go over 1?”

This is the right question to ask. The minute you realize these dynamics, you’ll realize that you can make product changes to increase that viral factor so that your product becomes more viral. For example, maybe you should ask new users to invite when they first come in, or you should make it easier to invite people because there’s a link that you can copy and paste or some other mechanical change that increases viral factor. This is one of the core insights that you can measure viral factor and that you can actually make product changes to increase it.

Because you can track and calculate this ratio, this means you can stick it onto a dashboard as well. As I described, it’s easy to do track, as long as you encode the user id into the URL that’s shared. Of course, as soon as you’re able to calculate it, the next thing is then to be able optimize it because then you can run A/B tests to see if more people share, and do they do more of them? Do they share with more people? And for the people who receive a growth link, what percentage of them convert, and how you make that conversion higher. This is sort of exists in the cookbook of all the different ways that you might be able to optimize your viral factor.

The final aspect of the viral loop idea that’s powerful is that it’s actually very generalizable. Which is why this is all so powerful for the modern era of Product-Led Growth. As long as you’re talking about being able for a group of users to somehow generate another group of users, then of course that growth process might be because they’re inviting other people. It might be because they are sharing a content link that then causes someone else to sign up. Or it might be a referral program or many other tactics that cause one user to invite another. This means that at the core, there ends up becoming a lot of applicability across marketing channels and strategies.

Alternative ways to calculate viral factor
When you ask the Internet how to calculate viral factor, it gives you something slightly different, which looks more like this:

VIRAL FACTOR = # INVITES x % CONVERSION RATE

This definition has been floating out there for a long time. However, it’s flawed because not all viral loops exist as a type of invite. Instead, we see sharing flows, collaboration flows, referrals, and many other variations. I also think that although it’s mechanically correct, because increasing the number of invites and increasing the conversion rate naturally does help your viral factor. It does not capture the fact that really what you want to know is this ratio between two cohorts of users.

Viral loops and spam loops
There is also a dark side of simply trying to increase these variables as well:

The more simplistic version of this equation pushes you towards thinking about how to be as spamming as possible. How do we get users to invite as many of their friends as possible? This sort of incentivizes pushing users to invite pushing them to invite a lot of people, and then for the invites to be spamming in such a way that there is a high conversion rate.

Going back to the Web 2.0 era, this actually was the way that people tended to think about virality. If you’ve built the bare bones of a social network as folks at Bebo, Tagged, Bebo, Hi5, MySpace did, you were generally growing your network based on getting people to invite their friends via email. In the early days, you’d actually ask people to type in their friends’ email addresses, and usually they’d type in like 5-10 emails since it was tedious. It turns out that if you could get them to import their address book from Hotmail or Yahoo Mail, by asking them for an email/password combo, then crawling their inbox, you could push the number of invites to 200+. Of course whatever viral factor you had, a 20x increase would often put you over the top.

The downside is that you would generally drive down your conversion rate because you were emailing dead emails. Eventually, all of the email providers would start labeling you as spam. But for ~10 years, the glorious email invite loop worked, and many large products were built -- including ultimately the big winners, Facebook and Linkedin.

Chain letters and why spam loops fail
You might know that over a hundred years ago, there was a concept of a chain letter where people would write actual snail mail letters to each other. Each chain letter would have a list of other addresses, and it would tell you that if you mailed a nickel over to them, and then also added yourself to the list, then eventually everyone would get rich because you would be receiving an exponentially higher amount of nickels from people over time. This was actually such a big deal that eventually the post office had to make it illegal since it swamped all of their delivery. But of course, mathematically, chain letters are going to fail.

The reason is simple, which is saturation. If you have a viral loop and you invite 200 people on each go, eventually after a bunch of generations, you’re going to completely saturate your target market. If you do the math, once you pass 5 rounds of 200+ invites, then you’ve already hit every human on the planet and more. So eventually, naturally, you’re just hitting the same people over and over again. Naturally, the response rate is going to go down. Why? 1) You’ve already signed them up, in which case that invite can’t convert. 2) Or, alternatively, they got invited before, but they’re not interested and they don’t interact. Either way, the response rate tends to decrease over time.

If the response rate goes down and your product is not retentive, then mathematically all you’ve done is created a huge spike of new users which looks good for a while, but then that lack of retention means that your total number of active users ends up being a tiny fraction of new users.

This is why some highly viable products end up jumping the shark. In the past, I’ve listed out some of these metrics to indicate that a product will be sticky:

1) cohort retention curves that flatten (stickiness)

2) actives/reg > 25% (validates TAM)

3) power user curve showing a smile -- with a big concentration of engaged users (you grow out from this strong core)

3) viral factor >0.5 (enough to amplify other channels)

4) dau/mau > 50% (it’s part of a daily habit)

5) market-by-market (or logo-by-logo, if SaaS) comparison where denser/older networks have higher engagement over time (network effects)

6) D1/D7/D30 that exceeds 60/30/15 (daily frequency)

7) revenue or activity expansion on a per user basis over time -- indicates deeper engagement / habit formation

8) >60% organic acquisition with real scale (better to have zero CAC)

9) For subscription, >65% annual retention (paying users are sticking)

10) >4x annual growth rate across topline metrics

These aren’t meant to be exhaustive, but they are all strong indicators that if a product can generate a lot of viral users, there’s enough there that it’ll stick. If not, then eventually the spike will go away.

The fact that you need product-market fit ends up being one of the reasons why sometimes you hear about highly viral apps that then disappear from public consciousness. You can actually have a very high viral factor or high word of mouth and sometimes you can even engineer it. But it’s almost like having a big bang launch or having a Super Bowl commercial or something like that. You get a huge spike and it’s great for a while, but inevitably, unless your ratio of active users to sign-ups is very high, it means that you’ll eventually lose them.

This was ultimately the outcome of many of the apps that were created during Web 2.0 or during the Facebook platform era. You heard about a lot of virality and potentially millions of users or even hundreds of millions of users, and sometimes this happened quickly. But a lot of this were people engineering viral loops. And as those users left the app because of a lack of stickiness, these viral loops could be engineered to try and reacquire them over and over again. The vast majority of these viral apps did not actually become successful businesses, which is why- Although I remain very interested in the ability for teams to create growth hacks. Of course, I also care a lot about their ability to actually retain users.

BRAINDUMP ON VIRAL LOOPS #2

You’ll often notice that very viral products tend to fall into one of two buckets:

Category 1:

very simple apps that do one thing, and that one thing is highly shareable or pushes out a lot of invites. Overnight successes. (think gen AI creative tools, but also Instagram/YouTube back in the day)

Category 2:

highly sticky products with deep functionality, and some sharing capability -- tends to grow slowly, but consistently. (think Figma, Slack, early Facebook, etc)

These two types of loops are very different.

Apps in category 1 have tight, simple viral loops with high conversions at every step, so that there’s almost nothing to do in the app other than creating viral content. The successful version of this is something like Instagram in its initial form with photo filters and sharing to Facebook, and not much else. The spammy bad version of this includes the litany of quiz apps and anonymous/secret apps that failed over the years. However, they have a huge advantage of being easy to create, and often in the early days of a viral platform, these are often the most successful ones.

The second category of apps ends up being complex products to build with high retention and a little bit of sharing functionality. The plus with these is that every time they acquire a user, they’re very sticky, and so, if you put the sharing functionality in front of users over a long period of time, eventually the viral factor will add up. However the growth doesn’t look quite as explosive as category one.

This type of loop is interesting to examine in the current era of AI apps because they days have viral loops very similar to Instagram, YouTube -- category 1. They have some of the same strengths (simple, high growth) but also some of the same weaknesses (spiky, low aggregate user retention). Maybe this is history repeating itself?

Viral content creation loops, step by step
These types of simple content creation loops work like this:

• you see something cool online

• then you watch

• click a link back to the creation tool

• then you try out the tool yourself. Cool!

• you share it on social media (or chat or whatever)

• more people see it

• ... and the loop repeats

I’m describing this very generically because you can imagine this being a video hosting service like YouTube, or it could be a cool photo with photo filters, or it could be an anime-style profile photo, or an AI-generated video with your friend’s face in it. In any of these cases, ultimately that takes advantage of the psychology of being creative. That is, after you create something, you want to share it with other people.

This type of loop is easy to understand, but it’s not just qualitative. You can quantify it too. Just break down each of the steps into percentages. For example, after someone views a video, there will be a clickthrough rate on the percentage of people end up clicking to your website. Then, after they go through the website, you’ll see what percentage end up creating content themselves. In fact, there are drop-offs at every single stage of this except for one stage: once you share it on social media or messaging, how many people end up actually seeing the whole thing. You might visualize the whole thing like this:

• you see something cool online

• then you watch (50%)

• click a link back to the creation tool (10%)

• then you try out the tool yourself. Cool! (20%)

• you share it (50%)

• more people see it (X people see it)

• ... and the loop repeats

The basic math on this viral loop is that if 50% * 50% * 20% * X > 1, then the loop will go viral. Of course this is hard because the first few terms together multiply to 0.005, so you’d need at least 200+ people to see it to go viral. But all of these numbers are highly sensitive.

When these loops work, they are magic. But funny enough, when a product is only slightly viral, you can barely tell. If you are attracting 100 users/day based on organic word of mouth, and you have a viral factor of 0.1, then the multiplier effect is only 1.11x. So your 100 users end up inviting 10 friends who invite 1 more, so that’s only a small bump. Can you really tell a small bump like that if your organic signups go up and down 10% each day as well? Not really. You really need high viral factors, like >0.5 to feel the difference. At 0.5, then 100 users will sign up 50 that will sign up 25, and so on, and that gets you a 2x boost. Now we’re cooking! This ends up being valuable because it means whatever you are spending on ads, you get 1 user free for each one that you buy. So even if you’re not viral, it’s helpful.

You can see what the multiplier effect is at different viral factors:

• viral factor of 0.1 is a multiplier of 1.11x

• 0.25: 1.33x

• 0.5: 2x

• 0.75: 4x

• 0.9: 10x

For the nerds, the math here is that the infinite expansion of gen 1 inviting gen 2 inviting gen 3 etc is 1/(1-v) where v is the viral factor. So if you stick 0.5 into that, you get 2x, which tells you that if you sign up 100 users with a viral factor of 0.5, you’ll end up with 200 total at the end -- aka 100 are free viral users.

As a result of the above, you are very much incentivized to get your viral factor above 0.5. Everything underneath that number almost isn’t worth it. Thus, building your app to be highly shareable, very simple, such that there’s almost nothing to do besides creating+sharing, is a clear way to go about it.

And it can be glorious when it works. But we’ve seen that there’s often spikes. One day it’s trending on social, the next day your traffic has crashed. (Let’s hope you pushed enough people through a subscription flow in the meantime, so that at least your ARR is sticky!)

Why the numbers go down over time
The natural direction for these forms of viral loops is that they degrade in performance over time. There’s a bunch of reasons:

• novelty effect creates higher metrics early, worse metrics late

• market saturation is inevitable

• virality runs on pre-existing platforms, and the platforms might get mad

There’s just a novelty effect. When there’s a new class of products like content creation tools with generative AI, users tend to respond to them at very high rates of adoption. A higher percentage want to watch them, a higher percentage want to try them more. And when they’re shared, a higher number of people want or are receptive to seeing the content. We saw this in the early days of 2D image generative AI where people would share amazing AI photos (even when people had six fingers). But when was the last time you shared a 2D image generated by AI? The novelty is worn off, and so you need a much higher bar before you’ll engage in that same behavior. (I won’t rehash The Law of Shitty Clickthroughs here, but it’s all the same idea)

And of course, as the viral loop becomes more successful, it may burn through the entire market of target users and the saturation rates will put pressure on all of these metrics again decreasing viral factor. At the beginning of a loop’s life, the entire market is addressable, so if you are inviting 10 people into the app all 10 are valid targets. But if you’ve saturated your network, maybe out of the top 10 people you’d invite, 5 people have already joined, so that only the 5 are valid, which drives down your viral factor by half. And these final users tend to be lower quality as well, since they are late adopters. Back in the email invite loop days, you would find that if people imported their email addressbooks, you might initially get ~200 contacts out of it. If you emailed them, you might have like a 10-15% open rate, and a 5% clickthrough rate. But as your product grew to many tens of millions of users, then peoples’ contacts typically got smaller -- eventually 150 then 100 -- simply because you were reaching into the edges of the global social graph. Popular, well-connected people tended to be onboarded first and then by the end you were bringing on the introverts, late adopters, and luddites :) Not only are they less viral, but their usage/engagement tended to be that much worse.

Every viral loop runs on a pre-existing platform. Email invites run on email. Instagram’s early sharing ran on the Facebook platform. Today’s generative AI video apps publish content to YouTube/Tiktok/social. So if it turns out that people are sharing a lot of these kinds of videos with certain watermarks and link backs to the creation tool, and eventually the video platform that you’re publishing on doesn’t like this, one of these conversion numbers might drop really, really low. And then your viral factor might go from being over 1 to being under 1, and the whole thing will decelerate. The platforms often regulate this behavior tightly because they are themselves sometimes building competitive solutions, or they don’t want their platform taken over. (Just read about Zynga/Facebook platform wars back during the Web 2.0 days)

The weaknesses of hypersimple apps
Even given these factors, the hypersimple hyperviral Category 1 products can win. I think of YouTube and Instagram as examples, where at the end of the day the entire app can be described with 3-4 screens of UI. Of course in subsequent years a lot has been built out, but the simple core is still there. It turns out the minimum product to generate a ton of retention is just a few bits of UI after all, and the network effects of having zillions of pieces of content means that a small app with a deep content base can remain endlessly engaging. This is sort of the magic of networked products like messaging apps, UGC, social networks, and so on.

BRAINDUMP ON VIRAL LOOPS #3
ok -- so here’s the case against all the fun/dumb viral techniques you’re seeing on social media right now. When people talk about “going viral” these days the state of the art -- if you can call it that -- is a hodgepodge of shitpoasting and social media videos. This includes:

• ragebaiting and shitpoasting by startups

• beautifully shot cinematic launch trailers

• tiktok video clipping

• billboards for social media coverage

• influencers astroturfing your app

• founders turning themselves into influencers

• etc etc

These are fun, but I predict they will not stand the test of the time -- not the way that classic viral techniques like referral programs, sharing links, invites, and so on have proven themselves over multiple eras.

One quantitative way to talk about this is -- do these techniques scale signups as a function of active users?

That is, if you tracked this as a ratio:

new users / daily active users

... would you be able to scale new users as you scaled daily active users? If not, naturally, new users will stay constant and even with strong retention, your DAUs will grow linearly but not exponentially.

One core reason is that these techniques generate one-time traffic. Good for creating a single spike of new users in your app. However as you grow and need a repeatable source of user acquisition, you simply can’t repeat these over and over again. If you drop in an amazing cinematic launch once a year, people may pay attention but if you start doing it monthly and then weekly, there’s just diminishing returns. Similarly there’s no defensibility in shitposting. If you shitpost and other founders shitpost, eventually you just won’t stand out as much anymore. As your customers acclimate and become less responsive to the novelty, these techniques will fade over time.

But! But! I admit they are still useful

Here’s why. (But allow me to diverge to a short history lesson here from the Web 2.0 days)

The end of Web 2.0’s virality
In my previous brain dumps on the Viral Loop topic, I talked about how during the Web 2.0 days (~2005-2010) there was insane innovation in email invite loops, content sharing, virality, Facebook apps that went from zero to millions of users overnight, etc. Then it ended.

What killed it? Mobile did.

The “golden age” of viral loops was possible because for the first time during the Web 2.0 era, there was so much novelty in getting an email or notification from a friend who is inviting you to something or sharing a piece of content to you via an app, etc. This made response rates very high, viral factors were able to go >1, and you saw a ton of products grow overnight -- including Facebook, Linkedin, YouTube, Spotify, Pinterest, and many other products that got their start with clever viral loops.

In the golden era this is where you saw hyper-simple hyper-viral products succeed. In the earlier part of the era, you had apps like BirthdayAlarm in 2001 (h/t Michael Birch, later the creator of Bebo) which would email you about a friend wanting to keep track of your birthday (how kind!), driving you to sign up, and then flipping that over to asking if you wanted to keep track of your friends’ birthdays. Put in a bunch of emails or eventually import your email address book and you can now ask hundreds of friends. Repeat that over and over and you have an app that might be used by millions of people.

Another example -- there were loops like Plaxo in 2002 (h/t Sean Parker, later the president of Facebook among many other adventures) that started with a friend asking if you wanted to maintain an up-to-date contact about yourself. Update the contact and then it asks you to add your friends and maybe a bunch of your other friends’ contacts kicking off the loop again. Many of these were hyper simple but when they were paired with social profiles and eventually a feed, these were the underlying mechanics that eventually birthed the category of social networking apps. Importantly the Social Network category had real utility and retention and wasn’t just a viral app on its own. We’ll talk more about these dynamics in a second.

But the golden era of Web 2.0 virality eventually ended. Consumers got used to these techniques, response rates went down, spam filters kicked in, and importantly the whole world moved to mobile. When email virality was dominant, you had the ability to ask users to import their email contacts and often you’d get 200+ emails that could then be invited to the service. Before the Law of Shitty Clickthrough kicked in, you might get something like the following:

• 50% of people would connect their contacts

• 50% of people would invite friends

• 200+ invite emails would be sent

• 10% would open/click through

• 50% of people would sign up

• ... then repeat

This would often generate viral factors of >2 or more (just multiple all the numbers above).

But mobile was a whole other thing. Apple made mobile contacts widely available, but you have to invite numbers one by one. Who’s going to do that? Response rates were high, but the # of invites was small. Some folks tried to build around this, using services like Twilio to deliver invites from the server. But that led to SMS spam, and the CAN-SPAM Act and resulting millions of dollars of fines convinced people to avoid all this. Between the platform change from email virality to SMS virality, and the declining novelty of invites, the response rates (and viral factors cratered)

The era of hypersimple hyperviral apps ended here

And to this day, it’s basically impossible to create loops where the viral factor is >1 in the first session

Retention is king for virality
The modern app isn’t obnoxiously hyperviral -- prompting to invite and invite, as prior generations of products did. Instead it’s driven on a few components:

1) multiple top of funnel channels

2) great retention that drives virality

Here’s what I mean:

First you need multiple acquisition channels that are diversified and can be as spiky as you want. This is where social media, video launches, press mentions, SEO, and even paid marketing techniques can all feed into your signups. It’s fine if this doesn’t scale or grow as long as you get some kind of consistent top-of-funnel dripping users into your product at all times.

For a product like the Uber app -- when I was there -- it was something like 50% of first trips might be from paid marketing. Then 10-20% from referral programs, and the remainder from word of mouth, SEO, and so on. We’d buy activated mobile users, with broad base and untargeted advertising, for $10-20 and the math all worked. For other products, you might have a different acquisition mix depending on how aggressively you pushed paid versus SEO and so on. A high-intent product category like travel might have a lot more from SEO and referrals, since you need to be near transactions. Either way, you just need some set of top of funnel sources that work.

Second you need the product to generate a bunch of user sessions, so that you can grow virality over time -- that’s another way to say, you need strong retention. Oftentimes the viral factor is described simplistically, as:

invites x conversion rate

... but that implies that all the virality is generated in one session. In a highly retentive product instead you get a whole bunch of sessions and you might be able to ask users to share, or invite, or refer every single time. So instead you can think of it as the following:

the sum of:

session 1’s viral factor +

session 2’s viral factor +

...

And I’ll leave it up to an exercise to the reader but you can think of this as the infinite sum of each of the points on the retention curve of which each session can generate a little bit of viral factor. Each session’s viral factor is then based on the % of users that interact with the viral features, multiplied by the resulting shares or invites, conversion rate, etc.

I like to use a rule of thumb where about half the viral factor is generated in the first session and then the rest is generated in all of the sessions after. The reason for this is that in the first session the user is in the mindset of “setting up” their account. This is where you can ask them to set up their workspace and invite their colleagues or to invite their friends, etc. They’re often coming in with a high degree of intent and those set up steps can generate a ton of reality. In the second or third sessions, the problem is not only has have many of the users dropped off but also they are sort of in a different mindset where they’re expecting value from the product at that point. So while you can take them into viral flows, you don’t have the excuse that it’s for setting up.

The other thing that you’re seeing is that in reality apps will have multiple loops that are all generating different types of viral factor. A product like Dropbox might have several, such as:

• sharing a folder with a coworker

• inviting people to share

• referrals program

• using other Dropbox apps that have their own viral loops

Each one might work at varying levels of performance but the important thing is that you might be able to convince a user to engage in all three at varying levels across multiple sessions

This was the case at Uber, which I saw first-hand. People were introduced to Uber in many ways. You had loops like the referral program or you could earn dollars by inviting friends and they would receive dollars too. But you also recruited users naturally simply because you would often invite friends IRL to come take a ride in your Uber. Or engage in functionality like “Share ETA” that would expose friends to Uber, driving them to eventually download it.

A product might have three or four major loops that all work in concert and are ideally all completely useful based on the context of the usage. That way you’re not asking people to invite over and over again but instead there’s a loop for inviting on your first session, then maybe a referral program that you introduce to them in session number two, then you might have some embedded functionality where they share some content or otherwise with friends in a later session and so on.

To nerd out a bit: The viral factor is not generated by how many invites you send in one session. The viral factor is generated by the sum of all the viral features that you engage with over all of your sessions. The more loops you have, the more useful across more sessions because of great retention, the more viral your product is over time. And the more novel and exciting your product is to users (as AI products are now), the easier it is for the entire system to work together.

Low retention apps need to be spammy. Sticky apps do not.
Naturally if your product is highly retentive then you get many sessions to be able to ask the user to share or invite. You can have a small and unobtrusive viral sharing feature. As long as you get a large number of sessions, you’ll eventually have a viral loop that’s >1. On the other hand if you have a low retention product where you only get two or three sessions on average, then you’ll need to be very prominent and spammy to push the user towards viral features. This is why sticky, high retention apps can be more viral over time.

I remember in Facebook’s early days it was amazing how less spammy they were compared to other social networking tools. While they had email invite capabilities, they existed just on the right railing of the website. Small and unobtrusive. If you didn’t have any friends, it would be quite prominent but for most users it really was not a big thing to push people to invite as many friends as possible. I think this is because the product was very sticky and well made from day one and so the viral factor was gained over a really long period of time rather than some of the spammier social network competitors that had to use invites to grow but ultimately a lower retention rate and user frustration given the spamminess caused Facebook to eventually win.

Is viral factor useful when it’s <1?
The funny thing about viral factors is that although people often love the idea of exceeding 1.0, as far as I can tell those unique windows only last in very short periods of time when there’s a new platform or some new novel mechanic that ultimately allows hyper simple viral apps to succeed. Most of the time, you see viral factors that are 0.2 or 0.3 or below.

This is still valuable! A viral factor of 0.2 means that when you have 1000 users sign up, you get 200 users “for free” and that discounts CAC in a meaningful way. In this way viral loops have ended up playing more of a support role because they help stretch the dollars that you’re putting into marketing. A highly-retentive product will over time grow signups in proportion to the active user base even if it’s not doing it in a super quick way.

The “speed” of a viral loop is an important concept here. Virality may not manifest itself quickly, particularly when it’s a high-retention/low-spamminess product. When viral factor is generated slowly over many sessions, you can understand why viral loops have a speed of how fast they generate invites. If you have a product that users engage with every day like a social network app, then they may send off a bunch of invites every day. This leads to an accumulating viral factor that is going to generate more sign-ups quickly. Contrast that to the referral program of a product like Dropbox, which might be high utility and you often use it in the background but you may not mess with the referral functionality more than once a month. This means that while the product is super sticky (by the way, it is -- Dropbox has signed up hundreds of millions of users), it might take years to generate viral signups simply because the loop speed is slow.

In that way, viral loops can be slow, powerful, and make big userbases even bigger. And this is important particularly in later stages when you really don’t want to use paid marketing extensively.

In markets like consumer or prosumer where the eventual goal of an app is to sign up hundreds of millions of users, you’re not really wanting to use paid marketing to buy an audience of that size. After all, 200 million users * $10/install = crazy money. Instead you’d like to get there because you spend millions on marketing but you also get a lot of organic spread. You get a discount that comes from all of your viral loops, you have SEO and ASO, etc. (And GEO, or whatever comes next in the AI era)

Long live ragebaiting and shitpoasting
OK so let’s wrap this up -- what’s going on in today’s viral marketing landscape?

If you believe my framework so far, it means that shitpoasting and ragebaiting and cinematic videos and so on, just serve to help generate random spikes of signups. They’re not repeatable, but that’s OK. But those signups are then amplified due to the nature of many of the AI tools out there.

The current generation of AI tools often have a “create and share” viral loops that can amplify whatever top of funnel comes in. If you give your users a novel way to generate music, or video, or anything else using AI generative models, then naturally a high percent of users will do that. And then a high percent of users that interact with the generated result will want to share it with their friends and so that percentage will also be high.

I think that’s why we’re seeing so many amazing highly visual AI tools becoming super viral because they not only use this “create and share” of our loop. Further, we’re in the age of highly visual social media -- short form videos, embedded clips on posts, etc. And so the nature of gen AI output and what’s working on social platforms is highly resonant. This means it’ll spread far and wide.

The optimist has the Pitch, and the skeptic has the Anti-Pitch.
The Pitch — and sometimes we call it the elevator pitch — is sacred inside the tech industry. We pithily describe a new idea, how it works, and why it’ll work. It should only take 30 seconds. Our technology industry is built on the elevator pitch, because it’s the beginning of trying something new.

The skeptic (boo!) has something else: The Anti-Pitch.

It’s the quick one-liner that’s meant to shut you down. It latches onto one thing that you say in the pitch, quickly puts you down. And if you’re not careful, it works.

Here’s how it appears in a conversation:

• New productivity idea? “Ugh these things always take 4 years to build, it’ll take so much funding. You don’t have the enterprise skills to sell this to big customers”

• New AI idea: “Isn’t this space too busy? I think I just saw this exact idea last week. Isn’t this just a GPT wrapper and Open AI will add this in a few months?”

• New non-AI idea: “Have you thought about adding AI?”

• New social app? “The network effects for Instagram/YT/whatever are too strong. Invite channels are dead. If you get traction they’ll just add a tab in their app to copy you. It’s so hard to monetize with ads these days…”

• New dating app idea? “Dating apps are so hard. Churn constantly happens as people pair off. Best case, Match Group buys you for a small amount…”

• New consumer hardware idea? “The holiday cycle is awful. You’ll either build too much inventory and bankrupt the company, or you won’t sell enough and you’ll miss your shot. If you win, China will rip you off…”

• New Cursor for X competitor? “We’ve just seen 10 of these in the past few months. It’s such a busy space, it’ll be really hard to break out. We’ve already seen one in XYZ category fail, it’s not as easy as it looks”

• New VR idea: “Apple Vision isn’t doing well. Do people want to wear goggles on their face? I’ve seen the stat that millions of people have bought VR, but do they use them?”

• New online edtech idea: “There’s too much free content already, isn’t YouTube your biggest competitor? Isn’t AI going to disrupt education completely?”

• New subscription idea: “Another subscription? Doesn’t everyone just cancel anyway? Won’t this just be part of Prime/whatever”

• Or, a few funny pairs:

  • Really brand new idea: “Seems like a lot of market risk? Have you done customer validation?”

  • Existing idea with a new innovation: “Really, yet another XYZ app?”

I could go on and on :) And you could too, you’ve certainly heard these. It’s truly the skeptic’s favorite tool. We can also admit that occasionally, we are the ones to think (though hopefully not say) these things.

The anti-pitch is lazy
Of course, the anti-pitch is the lazy way to think. After all, new things are very hard and are often bound to fail, and if you work in the tech world, you know that perhaps 10,000s of startups are funded each year and maybe 1-2% end up being anything. So 99% of the time, the skeptic would be right.

And yet we wouldn’t have new things if we all thought this way. Thank god the world has amazing new startup founders that still persist even when confronted with the difficulties. Plus, you often read about people being skeptical of the internet or Uber or whatever, trashing it with an anti-pitch, and years later it sounds real dumb.

So what's an optimistic founder to do?

The choices
When confronted with the Anti-Pitch, an optimist has a few choices:

• Ignore (eh, who cares what they think?)

• Accept, but then address the criticism

• Deny deny deny

These are all fine reactions in many cases, but they serve different purposes:

Ignore. Now, if the criticism is coming from your ideal customer, a candidate you're trying to recruit, or an investor, that's one thing. But often, the skepticism comes from someone whose opinion just doesn't matter. For example, it could just be some rando you met at a dinner party who's closed-minded and working in a dinosaur industry. You don't need to convince them, and there's very little upside or downside in going either way. Just ignore them. Maybe worse yet, there are some people in the world whose job *is* to criticize. Some of these people are called mainstream journalists. In the past decade, a whole industry has formed around trashing new technology products. And whether they're right or wrong, they're able to get clicks and traffic and create advertising revenue. So you should definitely ignore those people.

Sometimes the anti-pitch comes from someone whose opinion actually matters, like your ideal customers. If they don't get it, that's a really good question. On one hand, they simply might not know what they want. The often-quoted Henry Ford quip about faster horses makes exactly that point. This is especially true for new ideas with a creative or aesthetic component. It's tough to argue that you’re building, say, an email app that is just super well-designed and that you’ll win based on better design. It's hard to imagine the emotions and feelings a product will evoke. It's like describing the plot of a movie—you don't really know if it's good or bad until you actually watch it.

Criticism really stings when it comes from an ideal customer who gets what you're trying to do but just doesn't like it. In that case, they might just be a later adopter than you thought. Or maybe they care a lot about the problem you're solving, which is good, but they don't like *how* you're solving it. There might be a grain of truth in there, so listen.

Accept, then address the criticism. One of the really hard things about the Anti-Pitch is that it's often meme-like in its simplicity. It just ties onto one thing that you've said and tries to deposition every aspect of your idea. You might then decide to accept that. You might say, "Yes, this *is* the 10th investor-backed AI app for dog trainers, but..." But then you better have something interesting or clever to say after that.

This strategy works best when you're bringing something genuinely new to the table. Timing and innovation become critical. That's why, in this era of AI, it's easier to get a "yes." You can say, "We might be the tenth product in this category, but we're the first AI-native version. Here are three things we can do that no one else can." The problem, of course, is that you need to have an innovation or an insight. If someone's criticizing your consumer hardware, and then you go off on some new tangent about why your new gadget is so innovative, but you don't ultimately address the underlying concerns about the holiday sales cycle, it just sounds like you don't know what you're talking about.

The real danger is accepting the anti-pitch framing of your new product. By accepting their framing, you're already putting yourself in a hole, and you have to dig yourself out. Unfortunately, that is very hard.

Deny. Sometimes you end up in arguments where, if you describe an amazing new AI-powered product for finding restaurants, the "anti-pitch" becomes, "Well, here's why Yelp's business is so hard." You might want to deny that association by saying, "We're nothing like Yelp." The problem is that by saying you're *not* like something, you still connect yourself to that concept. It's the kind of conundrum a politician might face. An interviewer might ask a politician, "Do you think it's a criminal act to do XYZ?" If the politician angrily replies, "I am not a criminal!" then everyone in the audience is just going to think, "Hey, that guy's a criminal."

Pepsi also used this very cleverly to position themselves as the "un-Coke." Of course, by saying they aren't Coca-Cola, they're implicitly saying they *are* like Coca-Cola, just different. They're in the same category. In fact, it might be more clever to reply by saying “Yelp isn't the right comparison. The right comparison is how we're very different than [Booking.com or ChatGPT, or some other more successful thing you want to be compared to].” Again, contrast creates similarity.

Thus, I think “deny” is an option but the right move is to judo into a whole different direction.

The Best Anti-Anti-Pitch is a Good Pitch
What a founder quickly finds is that if their new pitch invites the same criticisms, you can use a little game theory and work backwards. Maybe come up with a better way to describe what you're doing that leads to a different and more positive outcome. For example, a new startup might begin by calling themselves a gambling app, inviting a bunch of criticism about regulatory risk. It might then make sense to iterate to a new version of the pitch that starts with the idea that this is a new predictive markets product. That at least gets you going in a different direction. You might get a different kind of criticism, but it may be that you have better answers.

In this way, you're navigating a maze of potential pitches, and as you hit dead ends, you find new, different ways to pitch your idea. I'm going to call this the "Pitch Maze" and copyright it lol.

To go further, almost every new product idea can be pitched in five or ten different ways. You might have one pitch that's really focused on the team and the people. Another might be focused on a customer need. If you have a dozen different features in your product, you might have different pitches that emphasize each of those. The point is that your ability to pick a pitch out of all the different permutations gives you a lot of leeway to go in whatever direction you want. Each direction might lead to a different universe of responses, criticisms, and doubts, but you can pick the right path that gives you the most advantageous options for responses.

Learning by Pitching
I don’t want to lose the cold, hard fact that there’s a shred or reality in the anti-pitch. For example, it's true that a VR idea might be hard right now because the VR platforms aren't doing well. If you completely ignore that type of consensus statement and try to work around it, you might win the rhetorical battle but lose some real data. This is especially true if that data point comes from someone in the industry or a customer who knows what they're talking about.

So I encourage you to listen to the Anti-Pitch. Not just pitch pitch pitch. A long time ago, someone taught me the phrase “show up and throw up” as a typical way that a junior sales person might act on a customer call. They show up, talk endlessly without listening, then leave, without understanding what the customer needs. Sometimes there’s truth in the Anti-Pitch, and founders need to hear it.

I say this, yet I root for the optimists. The world certainly needs more of them. So for the skeptics, you guys listen more too :)

I’ve been staring at retention curve data for 15-plus years now.

I was a founder, a product manager, and now a VC. And at Andreessen Horowitz, I end up meeting hundreds of startups each year, many of them through our a16z speedrun program, where we invest up to $1M in brand new startups. (And yes, we just announced the 2026 program and you can apply now).

But back to retention — I’ve seen thousands of curves, and it’s among the first things I ask for when evaluating a new startups. I've looked through thousands of data rooms, analyzed retention curves sliced across many segments and denominators. I've also seen it from the other side, as someone building products. I’ve run hundreds of A/B tests and drafted countless variations of onboarding and notification emails in attempts to bend the curve.

There are patterns.

Just as there’s the laws of physics, weirdly there are some constant patterns that keep cropping up over time. Here are a few that I’ll share:

• You can’t fix bad retention. No, adding more notifications will not fix your retention curve. You can’t A/B test your way to good retention

• Retention goes down, it doesn’t go up. And weirdly, it decays (oh, does it decay) at a predictable half life. Early retention predicts later retention.

• Revenue retention expands, while usage retention shrinks. Good news: You lose people over over time, but the ones that remain sometimes spend more more money!

• Retention is relative to your product category. There’s nature, and there’s nurture. Sorry, you’ll never make a hotel booking app a daily use product

• Retention gets worse as users expand and grow. The best users are early and organic. The worst users come after that

• Churn is asymmetric. It’s far easier to lose a user forever than to re-win them back

• Retention is weirdly hard to measure. Seasonality is a real thing. New tests throw things off. Bugs happen. D365 is a real metric but you can’t wait

• Crazy viral growth with shitty retention fails. We’ve run this experiment many many times already, across multiple platforms and categories

• Great retention is magic. When you see it out in the wild, it’s amazing.

We’ll dive into each of these.

You can’t fix bad retention. You've seen this happen before: You spend months developing a new product, and then you launch it. Bad news hits. The initial retention stats come in, and it’s terrible. You're already months into the product development, and it's hard to turn back now. How do we improve retention? I know, let’s add notifications and remind people to come back. Let's add a bunch of new features. Maybe we can A/B test the landing page and increase conversion.

I think we know how this ends. Unfortunately, it seems to be the case that when you have poorer retention, it's very, very hard to fix - nearly impossible. Yes, you might be able to make a marginal improvement. Let's say that your D1 retention is 40% and you'd like it to be 50%. This is great and potentially workable. If the D1 is 10% on the other hand, well, you probably just have built something that nobody wants, and all the local optimizations around A/B tests and notifications aren't actually enough to bend the curve enough for it to work. When there's been months of development and sunk cost, it's hard to not just give it a college try. But I think in many cases new products are better off pivoting right away.

The type of pivot that fixes retention involves a complete new redesign of the app's home screen. If it looks like a feed, maybe it needs to be a structured step-by-step flow. If the product is about sharing, maybe it needs to be mostly about creating and saving. You might need to describe the product in a totally different way and position it against a different product. It needs to be a big pivot in many ways, the bigger the better, in order to have a chance at changing retention.

Retention goes down, it doesn’t go up. Retention curves often follow very geometric curves that you see. For instance many curves I see resemble the following: Whatever the D1, it drops by 50% on D7. Whatever the D7, it drops by another 50% at D30. Months out you might end up at roughly zero, or if you’re lucky you might retain 10% overall. There’s just a predictable decay.

What you never see is a curve that starts high, then goes low, then becomes high again. That’s not possible. In other words, if your early retention isn't incredibly good, then it means that your late retention also probably isn't any good. You need to start strong in order to end strong.

There are some interesting exceptions to this rule that are worth calling out:

• Some products are extremely hardcore (e.g. online poker). You might have relatively low retention, but those who stay are extremely sticky and spend a ton of money. It turns out that this can work.

• A product that has network effects where new users might start out strong, then drift for a bit. But if the product (which might be a social network or a collaboration tool or something else that has network effects) is able to use more and more users to reactivate older users, you often see a small curve where retention comes up. This is a very rare type of situation but amazing when possible.

Revenue retention expands, while usage retention shrinks. One of the best and most important dynamics with retention curves is that you can apply them to users, but then also revenue. Thus far, we've been talking about user retention, and unfortunately, it has the undesirable dynamic of always going down. Revenue retention on the other hand is really interesting because people often end up spending more money over time with you, at least the ones that remain.

This is one of the biggest strengths of B2B SaaS products. Take a product like Slack. If you look at the user cohorts, what you'd likely find is that the retention curves go down just like any other product. Some people take to it, and some people don't. However, for the companies where people spend time adopting Slack, what happens is it will start to organically grow, and the amount of revenue you earn from that company starts to increase dramatically over time. The revenue retention curves start to grow rather than shrink. This is amazing, and unfortunately doesn’t apply to most consumer products. It’s one of the biggest ways in which B2B products have easier business model dynamics than consumer.

The consumer version of this looks more like Amazon where you might have started by buying books and music and over time as the product grows in its capabilities you start to use it to buy more and more things. Because of that your LTV in the product is essentially unbounded. We also saw this at Uber as well where user cohorts would decay over time but the amount of money that people would spend initially on Uber rides to the airport would grow into rides to restaurants or for commuting purposes to work. So the user retention curves go down but the revenue retention curves go up.

Retention is relative to your product category. I've written about this in the past on the concept of nature vs. nurture for retention. The reality is that there's just a natural use case for many products - for example with collaboration tools or coding apps, you might use them every day at work, capping your usage to 5 active days out of 7. Contrast that to something like a bug alert system - hopefully you don’t use it often! Same with consumer products, where people check news, messaging, and social apps daily, but generally don’t use medical reference guides frequently. Some apps have great retention but infrequent usage, like weather or banking apps. And some categories like gaming are highly addictive and frequent, but people usually quit after a few weeks of use once the content is played out.

Nature vs. nurture is important because it tells you that many new products simply don't have a chance. If you're developing a travel app, but it's meant to be social, the reality is people don't travel that much. It'll be hard to create a product that sole mission is interaction with friends. Instead, it would be better to accept its infrequent nature and figure out how to monetize it better by owning part of the transaction or to have a more frequent use case like a restaurant and nightlife app like Yelp but also be able to use travel features as well. It's just hard to fight nature. You can only do so much.

It's also for this reason that if you want to build a very, very high retention, high frequency app, you have to probably build within some of the categories that people are already identifying as core daily products. This means that most likely if your app is successful, it takes away from some other daily product. It's no wonder that my constant use of ChatGPT has dramatically diminished the number of Google searches that I do. Or that when I began to use Substack for reading and writing blogs that I stopped using many other kinds of social news software.

Retention gets worse as users expand and grow. If you are lucky enough to build a product that experiences great retention, one of the natural things is simply to extrapolate all the behavior, monetization, and usage to a much broader market and assume that naturally you end up with a very, very big good number because you're multiplying a bunch of small good numbers together with a big one. The reality is that as you start to scale your user base, bad things start to happen. Let’s say you begin adding Android and international users and you acquire more customers using paid marketing and other channels — you'll quickly find that all of your metrics get worse.

The reason is that the best users show up early. The ones that are the most highly monetizable, that have the highest intent, that are the most digital and plugged in, well, these guys tend to find your product early and start using it due to a recommendation from a friend. Later on, as you bring in new users from other sources, it's likely that your product just isn't as good for them. It could be as simple as building an iPhone app for college students in Western countries and simply getting worse metrics as you bring in Android users from emerging countries where the feature sets just don't quite work. Of course you can work to improve this over time, but I assure you it will never be the same.

Instead, the question is: As you grow your users and they get worse and worse, are they still valuable and can you still operate your product profitably? And more importantly, are you able to hold on to that core highly valuable user base that came in early?

No wonder these early users are often called The Golden Cohort.

Churn is asymmetric. It's incredibly easy to churn users. In fact, most products churn 90% or more in the first 30 days. Simultaneously, it's incredibly hard to win back a user that's already quit. This is the core asymmetry around churn. In fact, it's so bad that it's often easier to simply try to acquire a new user rather than to try to get someone back.

It's for this reason that life cycle marketing that involves trying to resurrect dormant users by sending them discounts or offers tends to be extremely painful and expensive. The version of this that often works is to get existing engaged users to resurrect somebody through the natural usage of the product. For example, if somebody at work tries a new project management tool and it doesn't stick, then you probably won't get them back by bombarding their emails with reminders of features. Instead, you try to get one of their coworkers to invite them back into the tool to work on a new project. That's what works. But again, insanely difficult and complex and is really only available to products that have network effects (i.e., sharing and collaboration).

Retention is weirdly hard to measure. When people talk about retention, they tend to try to measure what happens in the first day, first week, and first month. But they'll rarely talk about what happens two years ahead. The reason is that when you're working on a product, you need a short enough time frame and a thing that's easy enough to measure that teams can make decisions about what is happening. As a result, although annual churn or long-term monetization is incredibly important, you tend not to measure it, instead focusing on what's right in front of you and what's easy. However, this approach has many problems.

Unfortunately, many categories of products experience huge amounts of seasonality. Anything involving commerce, travel, wellness, or online dating are obvious examples. But there are cycles even to the way that companies use business software as well. Seasonality throws things off because you might be down month over month or quarter over quarter, but is that because of features that you launched? Or is it because user behavior is simply different in this quarter? It's just hard to measure retention when it's super laggy.

Same with bugs or new tests that you're running or new market launches. These are all things that muck up the data, and you end up finding yourself reviewing reports where retention curves went up or down. But there's an asterisk on every number because they're trying to validate that the new Android launch didn't create an apples-to-oranges comparison.

Crazy viral growth with shitty retention fails. Many folks working on new products find themselves very focused on signing up new users and not on retention at all. After all, if you just want to see a graph that goes up and to the right, why not simply ramp your top of funnel and show that you're growing quickly, raise a bunch of venture capital money, and then you can figure out the retention issue later.

We're seeing this all the time right now in the industry when products have a crazy TikTok ramp because a creator pushed their app to millions of followers or because a launch video caused a bunch of revenue growth. Even though the usage and churn are not in a good place.

The tech industry has already run this experiment many, many times. And the conclusion is the same: Highly viral products with shitty retention do not last because it's so hard to fix retention. Eventually, the user acquisition fades as the novelty factor goes away, and eventually you're left with shitty user acquisition and shitty retention, and what goes up must come down.

We've seen this across many contexts. During the early social network phase, there were many products that used email address books to spam their way to growth, but drove users to bad products. Sometimes, if you could get them to sign up to some shitty ringtone annual subscription, you could try to monetize them and make some money along the way. It wasn't until Facebook, of course, with their UX innovations like The Feed and Real Names, that eventually created a product that was both highly viral and had very high retention. The same thing has happened in mobile apps as well, where you see big hits pop up sometimes caused by forced invitations via SMS, but again, if the products aren't sticky, the whole thing collapses quickly.

Great retention is magic. You might read this whole essay and feel a little bit depressed. I know that sometimes it's hard to get things going. However, it's amazing when something really works. When you see a product out in the wild with a 50% D30 (I do once every couple years), it's just amazing. I've come to believe that these lightning-in-a-bottle products happen not because the builders had some incredibly systematic way to A/B test their way to great metrics or that they employed some kind of high-velocity iteration process that got them there, but simply that there's a little bit of magic that's required. This magic comes from a fresh insight about the market or customer needs, and while it might seem obvious in retrospect, it drives super high retention because this product is the first to figure it out. We can say this now about video conferencing software or disappearing photos or a magic AI that replies back to you about any topic. There's just a magic here that no amount of iteration and metrics-driven testing can get you to.

The Big Question
You might read all of this and still have a big question: So wait, how do you get to great retention? (If I knew the answer in a deterministic way, my job as a startup investor would be so much easier, wouldn’t it?)

But let’s try our best. In my points above, there’s a few clues:

• The idea really matters.

• If you want a high retention product, you need to pick a category that is high retention already.

• You need to pick a product category where you already use an existing product every day.

• You're going to build something that directly competes against that.

• If you win, then you'll stop using that other product and use your product instead.

That's a high bar, but I think it's a good start.

Of course, if you build something that's quite head-to-head with something that already exists, you might suitably object: "that's going to be really hard to switch somebody over." It is. So then this is where you need to decide to take enough market risk, but just the appropriate amount, where you do something new and different that reinvents that core interaction. But you're probably talking more about a 20% remix rather than an 80%. Ideally, you need to be able to describe this to your users in a way that they can understand quickly and viscerally within the first 60 seconds of usage.

This is where the dreaded investor question "why now?" starts to matter quite a bit. Because what you're saying here is that ideally there's some kind of new development in the industry, whether that's a general-purpose technology like LLMs or a societal difference like the oversaturation of social media, that allows you to make this twist happen at exactly the right moment.

This gets you into an existing market quickly, and you're more likely to have great retention numbers early on. Timing matters a lot. If you get the timing off, and it's a low-interest category, and your differentiation isn't different enough, then what you'll find is you've traded a retention problem for a user acquisition problem. Here's the difficulty with building a new kind of internet browser: if you win, it's incredibly sticky. But people are so happy with their existing browsers that it's very expensive and complex to get them to try yours in the first place.

This is why I don’t blame folks who have a “Cursor for X” idea, or “Figma for X,” just like the “Uber for X” ideas of the past generation. They’re trying to piggy off some existing markets and behavior so that they don’t have to take crazy market risk.

And if you get the differentiation right, the timing right, and there’s a ton of user demand, and you’ve nailed the right base product category, then I think it can really work.

But what about new markets?
The natural counterpoint is that new markets are often more exciting than existing ones. Isn't tech about building brand new things rather than innovating 20% on old stuff? Of course this is true, but I think this is the tiny tiny minority of products.

My counterpoint to this counterpoint is that most products actually have some kind of prior lineage, even if those prior products are quickly forgotten.

Before Instagram there was Hipstamatic, which had become the #1 paid photo app in the early App Store. It demonstrated the success of photo filters. Of course Google was not the first search engine, it was actually #10 or whatever, after Lycos, Excite, Infoseek, etc., which demonstrated consumers wanted search but that it was impossible to monetize. Tesla was not the first electric car, nor iPhone the first smartphone. Sometimes it’s the 10th iteration that matters. Some call this “last mover advantage” rather than first mover. I think an important point.

Yet sometimes new things do happen. Uber was created to turn an existing offline action — calling a cab — into an app, not because there was already a hugely successful ridehailing app. (And no, not Lyft — it was a weird bus booking thing at the time). Of course a lot of ChatGPT, with OpenAI’s 5 year journey between inception and v3 which really took off, and without any real blueprints for what it might replace. These types of journeys are remarkable, and the tech industry is better off for it, because they involve real risk as part of new category creation.

Hello!
I’m back in sunny LA until the end of the year. Just announced that applications are now open for a16z speedrun 6 — where we invest up to $1M in brand new startups. I’ll also be hosting a ton of events at SF/LA Tech Week in October. Register here for more info. And if you want to host too, here’s more info.

The Home Screen Test
It's the golden age of AI, but it's still incredibly early. The "Home Screen Test" to make my point:

How many apps on your phone's home screen are AI-native? How many of those apps were created using AI coding tools?

Weirdly, the answer for most of us might be a lot closer to zero than you’d think, particularly beyond the obvious LLM apps — ChatGPT, Grok, etc. Of course people are working on this, but for now, the Home Screen Test says that we have 4x7 grid of apps on our screen, and very few are AI-native. One day shouldn’t it be all 28? Where’s the AI-native Calendar app, or the AI-native Social network, and so on? I remember in the web-to-mobile shift how quickly my engagement went completely to the mobile app versions of messaging, social, email, etc. What’s going on here?

This tells you that there's an enormous opportunity because we haven't yet started to really change the way we work by using AI. So far, I just Google a lot less and I prompt a lot more. But, certainly, there's a lot more to it than that. I’ve written more about vibe coding here and network effects in the AI world here.

Other big questions on AI and how it’ll change the way we build startups
And in addition to it being early on products, I think we’re also very early on the questions related to how we’ll build those products. That is, AI will change the way we build the startups that build the products, but it’s not yet clear how that’ll happen.

A few examples of the open questions that you could have very reasonable differences in opinion on:

• Will the startups of the future need fewer (or more) employees? The argument is that AI might provide 1000x more leverage. As a result, what might have required a whole company to achieve can now be done by an individual. Naturally, this might mean that you have 1 person who supervises 1000 agents who code all day, and this becomes a billion dollar company. This would be amazing of course, but the counterpoint is this: if an AI-native startup can scale very fast, but some percentage of its capabilities still have to be done by humans because AI can't handle them yet, then you'd expect that they'd still hire a ton of people. Maybe the bottleneck ends up being “taste” (as some folks claim) and you end up needing a ton of designers. Or maybe you can prototype a lot of popular products well, but you

• How will defensibility work with fast-moving AI competitors? If it becomes incredibly easy for a product to be copied instantly, how do you define a moat if AI commoditizes all capabilities? One easy answer is that the last decade of consumer apps has taught us that, in a low technical differentiation ecosystem, user growth and network effects are all that matter. On the other hand, maybe the pace of AI software creation will speed up so fast that the ability to constantly iterate, launch new features, and create new products becomes the differentiation in itself. One thought I have is that the only products worth building will be those with multiyear horizons and heavy CapEx requirements. Think space tech or B2B hardware—areas where you need deep intuition about today's markets to make the initial bet. Any software product that AI can build within a few years will be arbitraged away. The advantages there will be more like those in direct-to-consumer, where brand and temporal distribution insights matter most. But it's much harder to build a great business that way.

• Will AI make startups cheaper or more expensive to build? Obviously, some infrastructure layers, like foundation models, are very CapEx heavy. And theoretically, some apps *should* be very easy to build. But growth and distribution cost a lot of money to get into customers' hands. One thing we've learned over the past decade is that even though it's relatively cheap to build a web app, acquiring users can still cost millions. The failure rate remains high because there are just so many products competing for people's attention, so maybe that’s the limiting factor.

• How will we organize the team of the future? Does it still make sense to have separate functions like engineering, product, or design? Or, as building a product becomes fully multimodal with AI tech that can build software using a PRD/wireframe/whatever, will all of these disciplines collapse into one? As many amateur historians know, the structure of work has evolved significantly over recent centuries. We moved from cottage industries, where families made things in an artisanal fashion, to factories and corporations thanks to industrialization. More recently, software platforms have enabled the emergence of massive numbers of 1099s and part-time workers, like delivery drivers. It's unclear how AI may affect the next generation.

• Does it still make sense for the San Francisco Bay Area to be the central tech hub? It used to be that Silicon Valley was the center for startups because of the network effects around talent, venture capital, and knowledge. I'd argue that while SF’s advantage has stayed strong, but has definitely weakened as a lot of that talent migrates to other hubs like New York, and as that knowledge gets shared more widely via podcasts and Substack. The argument for continued decentralization is that if building products becomes trivially easy, it's almost like a form of content creation. And as we've seen with content creators, there are local and global creators who can be based anywhere. If you don't need to scale with employees or venture capital, maybe the future entrepreneurs will be based everywhere, too.

• How will venture capital work in a world of startup fragmentation? For a long time, venture capital worked simply by investing in the best spinouts out of Stanford and the adjoining towns full of tech workers. What happens when it becomes super easy to build a brand new product and test it in the market? If people are willing to pay for it, maybe these products will be super profitable from day one, especially if they can be built by just one or two people, keeping costs low. What happens if these kinds of products are being built all over the world? Perhaps venture capital itself gets more distributed, as it can be used as growth capital rather than risk capital. This makes it more obvious and accessible to a wider number of investors, even if those investors are located globally rather than just in the Bay Area.

• Related, does preseed versus seed versus Series A/B/C still exist? We have these distinctions to group different phases of growth. But maybe more products will just jump from zero to Series A. And does it make sense to fund people to wander around experimenting, or will this just all happen as side projects.

What's great about these questions is that the contemporary tech industry is decades, not centuries, old. A lot of the modern constructs, like venture capital, mobile/internet, or building startups in SF (not the peninsula) have happened relatively recently. Because things have happened so fast you can easily imagine it changing over the next few years as well.

The historical precedent
The aforementioned questions seem like obvious avenues once you start to think about how business structures have evolved alongside technology over time. Think back to the artisanal cottage industry of pre-industrialized Britain. In the 1700s, a blacksmith might work in a workshop attached to their home, with the entire family producing, and the end products would be made one-by-one and sold at the market in small batches. Fast forward into the 1800s, industrialization meant massive pools of labor working inside factories. At that point, you need corporations to organize the efforts of operating the businesses, shareholders to finance the expenditures, and layers of specialized professional management. (Check out Burnham’s Managerial Revolution as a fun discussion of these dynamics).

Or you might look at the 1600s and the development of long-distance naval power in the form of merchant fleets and intercontinental trading networks, and the simultaneous invention of the limited liability corporation to weave it all together. The East India Company required both a technological and business innovation to build the largest corporation in the world, with a standing army of 260,000 soldiers.

That's why, when you examine the potential impact of AI, it seems inevitable that the business structures we've created today will ultimately be insufficient to organize all the potential production that comes out of the technology. If we moved from “the family unit organizes labor” to “the factory organizes labor” then what will that mean in a world of agents, compute, and models?

The most optimistic view
In the most positive version of the above, I think we want to see a world where AI enables fewer people to produce more. As such, it would be wonderful if AI-native startups need far fewer employees to build. Defensibility will be achieved through killer features and technologies, not by the simple strength of distribution and monopoly. (We want more new things, not the same old big guys to rule the next geneeration). We’d like for all these new technologies to make startups even cheaper to build. Ideally, the Bay Area will remain the central tech hub. Even if people can create companies anywhere, they'll still move to San Francisco to tap into the expertise and capital. And of course venture capital will evolve to figure out a way to make money out of the whole thing :) I hope that's right.

But you could also make plausible arguments in the opposite direction for each of these. The biggest AI winners might be those with massive data centers, access to tons of data, and lots of compute. This implies centralization, where the big get bigger, creating an anti-startup world. Alternatively, AI could be an amazing feature set that doesn't actually help much with marketing and distribution. Incumbents could slowly convert their pre-AI products into AI-native feature sets, eventually outcompeting startups.

As smart folks I know have said, the question is: “Will incumbents get innovation first? Or startups get distribution first?” Incumbents might win.

The next few years in startupland are going to set the stage for a lot of what will happen in the future. The last few years have been eventful. We've seen a wave of high-end AI research teams building foundation models. But now that these models have absorbed all the data in the world, they're asymptoting in their effectiveness. The next few years will be about the folks who build the business logic that sits on top of these models. They won't do AI research or train their own foundation models. Instead, they'll be model-agnostic, creating a compelling UI that sits on top. We're seeing these kinds of products absorbed into every different kind of vertical industry, with approaches spanning from selling tools all the way to rolling up industries and implementing technology that way.

It's gonna be a wild couple of years.

Micromort: A 1 in a million chance of death.

You can use this measurement to compare/rank the risk of any activity, and in fact your favorite LLM is quite good at generating tables of activities versus micromorts. You can even qualify certain activities, like “tell me the micromorts of road cycling, but only during the daytime in suburban/rural areas.”

Here’s a handy graphic that shows some micromorts for a wide array of activities. (tldr; don’t mountaineer, BASE jump, or get old):

This is an endlessly interesting measurement and you might ask yourself, isn’t this about exposure? If I cycle for 10 hours a week versus sky dive once in my life, how does that compare? Of course you need to estimate micromorts per hour — and an LLM is great at this — and normalize the exposure.

Another fun prompt is to ask, “based on what you know about me, what do you think are my highest micromort activities?” Endless fun. I’ll leave it to you as a reader.

Other ones I like:

• Cost per hour of pleasure (CPHP): Going to a concert or basketball game is high CPHP. Buying a super nice treadmill that you use 3x/week is low. Basically anything that you use all the time tends to be low, unless you’re fooling yourself that it’s an everyday forever product and it’s actually going to collect dust in your closet (yes, I’m very guilty of this).

• Complaints per Hour: This is as much for you as when you talk to your dear friends that are just going on and on. High CPH means high negativity. Maybe mostly a metric to lower? However, if it’s very enjoyable to complain (and sometimes it is) then it’s free, so it’s also effectively infinite zero CPHP, so maybe everything cancels out.

• Phone Pickups per Hour: If you are with boring people, or hanging out somewhere boring, the number of phone pickups is very high. As many of you know, when I wrote a book a few years back, the process is so intense (and sometimes very boring) that I had many many phone pickups. Eventually I just locked the phone away in a timed safe. Seriously!

• % Conversational Autopilot: We all hate boring conversations, but we can’t help ourselves. But we all end up stuck at group dinners or sometimes you take a random 1:1 meeting that just turns boring. You introduce yourself, you talk about your job, then maybe you talk about places you’ve traveled. It’s the lowest common denominator conversation, and you just put the conversation on autopilot because you’ve heard and said all the same things already. So what % of the convo is in this mode? Personally I find that great chats often hover around 20-30% — enough commonality to be comfortable, but then you tread new ground that will be memorable later. But let’s admit sometimes they’re like 75%+ and that’s when you want to dine and dash.

As you read through these, I think you realize (rightly so) that these ratios are mostly a nerdy way to complain about people and situations. It’s true. I think of my writing as usually low Complaints Per Hour but now you are getting a better sense of my inner self.

OK but what about business metrics (not micromorts)

The title of this essay promised you something more than micromorts and phone pickups though, it promised some new business metrics! And I have some for you. These are real metrics I have heard used or use myself:

• Lies per Second: In startupland, there’s a fine line between pitching and lying. Hopefully you’re just telling the truth attractively, but sometimes people are just making up numbers, saying they have customers when it’s just pipeline, removing labels from their graphs, and so on. In that case, the “lies per second” might start high and go even higher as the presentation progresses. When you find yourself in a high LPS meeting, it might just be time to end it early.

• Meetings per Decision Ratio: There are big decisions and little decisions. How many meetings did it take to make that decision? Take something simple like deciding to hire someone — some places have an organized, streamlined interview loop of 3-5, a presentation, and then boom, you decide. Other places it’s open-ended and you might have 3 initially, then another 5 then 1 or 2 more, then another 3, and so on. Why does it take so many meetings to make a decision? Sometimes there’s no owner to the decision so it’s unclear who has the authority to even give the greenlight. Sometimes it’s an ill-defined goal or process. Sometimes it’s constantly changing. Either way, a high MPDR tells you that something is broken, and the situation needs to be escalated to a decision maker to simplify.

• Time to First Excuse: You have a bad month. Why? “Seasonality.” Your customers hate your new product. Why? “Our marketing was poor.” You often sit in a review meeting and you can count the minutes before a poor result is presented, along with an excuse for why it happened. (It’s never the team’s fault, by the way). A low TFE often leads to a low TNE — that is, Time to Next Excuse. As the TFE approaches zero, this is when you know the team probably needs to get replaced/rebooted.

• Numbers vs Text Ratio: In startupland, one of my favorite stages is the very very beginning when the idea is being formed. That’s when it’s all story, and all text/narrative. There’s no numbers because there’s no revenue, no customers, and no product. Fast forward a few years though, and during a Series B, you should start to see numbers everywhere. You want a presentation that’s just chock full of retention curves, financial metrics, etc, etc. That’s why the “numbers vs text ratio” should start getting very high. Sadly, often that is not true. Sometimes you see startups that have raised $50M show up and just present a deck that’s all numbers. If the Numbers vs Text Ratio is too low, then the Meetings Per Decision metric is going to skyrocket. Or it’ll be low, because no one will want to invest.

• PowerPoints per Launch: Every product team in Silicon Valley has their own culture. Many startups learn by shipping — they launch features, see how people react, then launch more features. That’s zero powerpoints per launch. Somehow every startup turns into a big company over time, and there’s more coordination costs, and you find yourself unable to launch new initiatives due to “death by PowerPoint.” You want to pitch a new idea? OK make a powerpoint. You gotta go roadshow it though. Make more powerpoints. How about the market research that tells you this is a good idea? How about what our competitors are doing? What should be the KPIs? Make more powerpoints. The only way to reduce PPPL is to involve senior executives who want to push it forward, otherwise you’re in consensus-building hell.

• Dollar per IQ Point: If you hire a top 1% stanford CS student, believe it or not, you’re probably spending moderate dollars per IQ point. You might get an even cheaper Dollar per IQ Point if you hire a top 1% student from a state school — the IQ points may not be as high, but they won’t demand the king’s ransom. How do you hire the most expensive Dollar per IQ Points in the world? Probably by hiring a bottom 10% Stanford student :)

• Decision to Rumination Ratio: There are big decisions and small decisions. Sometimes you take a lot of time — weeks — to think through something, and sometimes you make the decision in a snap. Weirdly, they are often not correlated at all. It might take you months to pick out a new car — researching, test driving, etc., but then you might jump on a new opportunistic job with only a week of thought. Ideally, you’d like to think that you normalize this ratio so that it’s all relatively the same — so that you take a lot of time for big decisions, and only a little time for small decisions. But we all suck at this. One important prompt: What are the most important decisions in your life? Probably who you marry, what city you live in, and your career. How many of us are willing to take months or years full-time to actually figure these out? And to talk to experts? And to get expert help? Instead, we just kinda do what feels right and sometimes it works and sometimes it doesn’t. Thus — shotgun weddings, moving to a new city on a whim, ragequitting a job, impulsive breakups, getting a tattoo while on vacation, impulse buying a pet, etc., etc.

These are obviously fun/conceptual ideas, but one day I’m sure we will be able to measure them once we’re tracking all the meetings we have via our AI note taking apps! And maybe we can even have alerts automatically fire when the Lies Per Second metric exceeds a certain threshold, since we’ll be able to automatically fact check every sentence on a video conference.

I’ll stop here, but if you have any examples to add, please reply in the comments! Always looking to incorporate more of these into my dinner conversations along with micromorts and so on.

We all use social media these days so we know what it means to sound real. By now, we know this intuitively, and get instinctive cringe when things are inverted. So why do so many marketing teams still communicate in corpospeak? “We are delighted to announce a new product.” “We value your patience.” “We believe in breakthrough, innovative approaches.” And all of this said from anonymous, blank, faceless brands, that can’t and won’t ever reply to their customers. It’s wooden. We all know it’s bad, but we can’t help ourselves.

I know many of you work in these companies! And I know you’re working on it, hiring social media interns or someone presented a short form video strategy, or whatever. But IMHO that is not addressing the root cause.

There’s a lot more working on as a broader transformation of marketing:

• This is a massive reordering of trad marketing. Before the invention of mass communication, there actually weren’t national/global brands. If you were thirsty, you’d go to your local carbonated soda shop and just have whatever they had — not ask for a Coke. Using radio, newspapers, and TV to reach billions of consumers spurred up entire industries (just watch Mad Men, or read Claude Hopkins), mass industrialization, and mega brands so that everyone will buy and use the same things. So what happens when the big centralized mass communication channels get turned into million of microniche channels? What happens to all of our cherished brands, and the marketing playbooks that have been written over decades? It might be that the notion of a brand, and a brand voice, and strategies we’ve used to consistently reinforce a brand — well maybe they away.

• Marketing channel reboots are ugly. Have you ever seen a company figure out that a certain marketing channel works — say, buying paid ads, or heavy SEO — and they double down and triple down… and then suddenly it stops working? (because of saturation, channel competition, platform enforcement or otherwise?). It gets really ugly. If you look at some of the biggest blowups in ecommerce startups, it’s usually because of this — when CACs go up and LTVs start to diminish in more marginal audiences, and the paid marketing cycle is inverted. (I’ve written about these dynamics here). This usually causes an existential moment as marketing needs to get rebooted. This is what’s happening to trad marketing. The gradual then sudden implosion of corpospeak is this played out at an economy-wide level, because the centralized gatekeepers (press/“experts”/legal/etc) are in severe decline, but companies are so used to interfacing with these middlemen they don’t know what to do next. They need to reboot.

• It’s all about vibes now. Whether you like him or not, Elon talks to you like he’s a real person, and with his 24x7 posting, you get a good sense of the thinking behind all his companies. In the political realm we have seen the rise of Mamdani (sorry), or Trump (sorry to the other half!), and we could make a list of dozens more examples that fall into this bucket (AOC, Kim Kardashian, Warren Buffett, etc etc). I remember being shocked when I heard Steve Jobs would actually answer his email when random people emailed him. Being real matters a whole lot to people, and they will even overlook many flaws if they connect with you. Vibes.

But look — I don’t need to tell you all this. You already intuitively know all this. Again, because we all use social media these days. As a result, we’ve all developed the ability to discern what people vibe well (and what doesn’t) in this new order. We have the taste, but may not have the ability to actually create the content we want.

By default, marketing professionals create corpospeak. It jumps out particularly in email marketing, or on the about page of a website, or with a lot of conference content. Ironically, the more effort and the more people are involved in something, the worse it is. How do we get rid of corpospeak in our companies? How do we talk to people the way they want to be talked to? And why is this so hard?

The slow long death of corpospeak

The funny thing is, we actually all know the root cause — we’ve spent about a hundred years creating an intricate ecosystem of marketers, lots of jargon like “positioning” “brand” and so on, plus gatekeeper media corporations with trad channels, lots of experts, hundreds of books and HBS case studies, etc., etc. — and the mega fracturing of channels in the age of the internet is about to ruin it all.

We created this system, and here’s how it works:

• Voice of the marketer versus the builder. Corpospeak happens when a professional marketer writes the communication (which is then subsequently approved by lawyers/etc), rather than when it from the voice of the founder/builder/creator. It’s when the speech is formal, trite, and uses business idioms that you’d never use in everyday conversation. Further there’s an entire ecosystem of “experts” (sometimes I call this the PR/brand industrial complex) of agencies/PR staffers/etc will tell you it’s important to be elevated and polished. It’s meant to be short and sound bitey, and not nerd out too hard. However, it’s exactly the tone that pushes customers away in a world when they’d rather be real.

• Created as marketing, dies as marketing. People can always sense when something is created as marketing. Your customers are smarter than you think, and they immediately know when they are being sold to. And when you only communicate with them via gatekeepers — traditional marketing channels like journalists, or advertisements, or customer support call centers, then you’re like the “friend” that only calls you when they need to borrow money. Particularly when the marketing is all just one way. Often, brands are confused when they post to social media channels and no one interacts with them, no matter how many followers they buy. People know when they are being marketed to.

• Culture of scarcity. Another funny thing is that corpospeak tends to be published only at very specific, calendared intervals — after all, you have to worry about “brand dilution” and “making the biggest impact” in a world where print media is physically limited in size, and you can’t launch and re-launch a product. (what would the journalists think?) Contrast that to Elon’s hourly/daily discussions about Tesla — is that marketing? Or just a guy talking about his great passion for his cars? If you’re asking this question, you’re going to the right place.

• The misguided desire to please. The temptation is always there to seek recognition from the trads. Yes, 20-somethings all want to be Forbes 30 Under 30 even with the occasional criminality of their alumni, the complete decline of the Forbes brand, and the fact that there’s like 500 people on that list, not 30. But I see this in startups too — even though they ought to know that Techcrunch doesn’t matter for growth, or fundraising, or hiring, they still want a launch article. If it doesn’t work in any practical way, why do it? (we all know why). The same happens in the agency industrial complex where there’s numerous awards, conferences, and boondoggles for advertising professionals. These insular networks perpetuate a certain style/culture/tonality that furthers trad marketing.

• Fear of the public. This all happens in companies when people fear interactions with the public. If someone complains about a bug, it’s not the job of the actual app developer to reply and figure out the bug — no, you give that to customer support. What if you say something wrong? Aren’t you stepping on toes? From a customer’s POV, of course, interacting with a customer support agent is like the DMV experience of the corporate world. You have to wait, you are talking to someone who’s probably either a bot or based overseas, and you have to constantly explain your situation over and over. Imagine just interacting with someone at the company who can solve your needs — we’ve all seen the stories, and when it happens, it’s magical.

Much of the root cause is the functional distinction between marketing and R&D.

This will all seem familiar to many of you because for those of you who are builders — in the EPD teams of tech companies — you naturally turn all of your communication over to the marketing teams. They are the experts of course. If they need you to stand on a stage, or be in a webinar, they will help with that. They will help script your speech so that you say all the right things. The separation of the builders from the marketers is fundamental to the dynamics at play here.

In same ways this all mirrors what I saw first-hand with marketing as separate from product management when I first got to Silicon Valley. Back then, the marketers knew all the customer metrics and PMs often didn’t understand basic concepts like CAC, LTV, retention, or otherwise. A decade later, if a PM didn’t understand growth marketing concepts, they would be considered ineffective. In many ways I wonder if the evolution of corpospeak will require decentralized participation from an even larger number of individuals within a company.

The weird thing is, no one will defend corpospeak and its associated strategies. Because we all use social media, we see where it’s all going. No one wants to be trad. No one wants to be stiff.

Why corpospeak can’t just be delegated away

There is always the go-to fix for all this. “Can’t we just come up with a social media strategy” you ask? And then try to work with the brand/PR agency industrial complex to implement said strategy. Like any project, the company leadership wants to just delegate it away to the experts. Sometimes these experts then further delegate it to younger, junior staff. But this generates a lot of problems:

• People don’t want to hear from your 22 year old social media intern. They want to hear from the people who are building the product, and often they want to have substantive conversations. The more customers care about the product, the more they want to have real discussion and not just funny witty social media memes. This makes this entire task hard to delegate.

• It’s hard to be an expert in one thing, get disrupted, and then be an expert in the new thing. If a marketing team is an expert in advertising, conferences, traditional media, etc., they will simply think of social as “just another channel” rather than a fundamentally different approach. The natural approach will be to create an integrated top-down strategy with social as one of many things to consider, and staff it with some junior people and collect the invoices

• The actual builders inside a company are busy. Or they don’t want to be in the limelight. Maybe they don’t like making videos or hanging out on X. This might be, frankly, a generational thing between boomer pre-social media professional expectations, versus what’s coming with the digital native workforce. I’m sure the GenZ founders who are building the next generation of companies will be far more comfortable. So again, it gets asked — can’t this be delegated? No, sorry.

Any by the way, social media “experts” are often not actually good at social media. I’ve now interviewed a ton of them over the years, and one of the wild things is when they don’t actually use social media themselves. Or they don’t actually create any content on any platform. I’ll ask, “what’s your favorite platform” then look up their profile, which will have 200 followers. That’s an auto-reject for me. Why is this so common? Oftentimes social media jobs are more about perpetuating something than building. They might know some of the analytics packages, they might be able to edit videos so that they work on social, but they haven’t actually built up an account successfully from zero. Instead, many of these experts have just worked on big brand accounts that bought all their followers, and posting daily inconsequential slop. From that expertise, they are hired to do the same somewhere else.

These are common pitfalls. They may not affect your company, but this chain of events starts with the desire to outsource the responsibility of talking with customers, partners, stakeholders to marketing experts.

So instead, I argue the opposite — you can’t outsource.

The new normal is that the builders of products — the founders, the engineers, the designers — will have a decentralized responsibility to actually talk to people. They will have to learn the skill of asking questions, replying, having interesting things to say — all without causing sharp changes in their company’s stock price due to unfortunate outbursts. They’ll need to figure out lightweight frameworks to work with legal, or compliance, or brand marketing, so that they can actually make this happen. And maybe those teams will be set up to enable these interactions, rather than protecting the downside.

A few small tactics in a new playbook

I wanted to share a few things I’ve seen work out in the wild — here’s how many new startups I’ve been working with act and sound real:

• Make it everyone’s responsibility to talk to customers online. For the camera shy, this means yes, you’ll need to get used to creating content — written or increasingky in video form. And yes, it’s OK to reply to people on social media. (And if someone on your team says something off, well, tell them why, add it to the playbook, and keep going)

• People want to hear from the founder/CEO. It’s especially the founder/CEOs responsibility to talk to customers and represent themselves online — not their marketing team. Maybe the marketers can help flag important threads, ghost write, etc., but you need the real main person. That’s who people want to hear from, and who they will build a relationship with over itme

• It’s OK to talk about your products in relation to your own motivations. Say “I thought” and “Why I did X” about your why you did something, rather than referring back to the company’s sales pitch. Post from your personal accounts, not the companies. Use photos of yourself, not exclusively glossy professional content

• Don’t be afraid to nerd out. If you’re passionate about something and could talk about a thing for ages, do it. Your most hardcore customers, potential employees, or partners will want to talk about the same things. It’ll attract them

• Don’t delegate the doing. Maybe you can learn from folks who are native to a channel on what works, but then you actually have to do it yourself

• Quantity beats quality. Brand ubiquity beats brand dilution

• Imagine that you are signing off every email as yourself, not as the company. When you say, “hello all” then write a message, and sign off as yourself, there’s something special and important that is being communicated there. Write your message that way, then remove the header and footer. It’ll sound much more as you intend

• If you decide you need a social media strategy, look at the profiles of the people pitching you the strategy. Make sure they’re actually social-native and not just making a great pitch deck

Finally, thank you for your time. We hope you enjoyed this content, and that it reflects our innovative and thought provoking values. You are sincerely appreciated in this regard. :)

Hello all!

Wow — lots of you read my recent post Every marketing channel sucks right now. It’s now one of the top 5 most widely-read essays. Thank you for that, and wanted to refer to a few other essays** I’ve written previously on related topics.

Couple other updates:

• a16z speedrun deadline this Sunday. As mentioned prev, I’m organizing 12-week program here at a16z focused on very early startups, where we invest up to a million! I’ll personally spend time 1:1/small groups with you. Applications are due this Saturday (May 10), so consider this your deadline and final reminder! Apply here — really only takes 5 minutes: http://sr.a16z.com. More details here.

• Also, I would love to see y’all in NYC when I’m there next month. I plan to be in town during June 2-6 for New York Tech Week, organized by the a16z team. There’s a bunch of panels/fireside chats where I’ll be speaking, including the Official a16z Kickoff Event, “AI Meets Culture: What’s Next?” hosted by Fenwick, Titans of New Media co-hosted with Substack, bunch of a16z speedrun events, etc. Check the calendar for more. (And friends/ex-colleagues, please DM me! Would love to say hello)

• I joined a16z 7 years ago, in May. Holy shit what a ride. Obviously it seems like so much has changed, and there’s been multiple distinct phases. A 2018-2020 set of boom years, pre-covid, where we were all investing in the post-Uber consumer age. Then lockdowns where I worked from an RV and lots of product categories thrived — particularly gaming. And I shipped my book. Then the past few years when I’ve been focused on tech, entertainment, gaming, AI, etc. A very fun, unpredictable, enjoyable journey so far — for which I’m grateful.

• Kicking off an LA speaker series program. I’m still living in LA! Even though the AI center of gravity remains SF, so I end up visiting every month. But as a side project, I’m now working with a bunch of folks in the LA ecosystem to put together a weekly program to hear from executives/founders from the top SoCal companies. Am hoping to get the big players like SpaceX/Tesla/Snap/Riot/etc, lots of startups, and VCs too. The idea will be that we can lend our Santa Monica offices to host events but will let outside teams host cool get-togethers. So watch this space! And reach out if interested to partner.

• Happy birthday to my wife! This weekend was her birthday and we went out for a wonderful (casual) meal. Pretty chill. I’m not sure she will read this far down this email, but if you do, you’re great and love ya.

Thanks to all of you for reading, and more to come soon!

Andrew
written from a hotel room on a 24 hour visit to SF :)

**related essays:

• Data shows losing 80% of mobile users is normal, and why the best apps do better

• How startups die from their addiction to paid marketing

• The Next Feature Fallacy: The fallacy that the next new feature will suddenly make people use your product

• The red flags and magic numbers that investors look for in your startup’s metrics – 80 slide deck included!

• Why consumer product metrics are all terrible

These days I’m spending a lot of my time with very early stage startups (yes, as part of the program at a16z to invest up to $1M into each, called a16z speedrun) and as part of this, I spend a lot of time talking about launch and marketing strategy for new products.

The options for marketing are pretty grim right now. Here are my complaints:

• SEO: Takes too long, you’re competing against listicles (and Reddit threads), Google might screw you over at any moment, maybe AI one boxing will kill your traffic next year anyway :(

• Influencer marketing: Get a big spike of traffic but none of it converts, the spike goes away after a few days, big creators are too expensive and a slew of small creators need to be cobbled together, lots of babysitting :(

• PR/comms: Doesn’t actually generate signups, doesn’t scale and not repeatable, expensive retainers for PR experts to grab coffee with journalists, your competitors will get the same article next month, and press is as likely to attack you as to cover you :(

• Email marketing: hope you like spam folders, building a good list takes forever, open rate rates are <30% and CTRs are <5% so hope you weren’t expecting a lot of clicks

• Viral loops: doh you need an actually great product, all the contacts spamming techniques no longer work (neither email nor SMS!), your UX will be ruined by aggressive popups and onboarding schemes, and it’s nearly impossible to get viral factor >1

• Ads: way too expensive, always getting more expensive, your competitors will just copy your copy, lots of fake/low-conversion clicks, your investors hate it, and it’ll kill you if you don’t have a strong business model already

• Referral/affiliate/etc: be ready for a crazy amount of fraud, it’s just as expensive as paid marketing (though people fool themselves into thinking it’s not paid), and surprisingly most people don’t care and will get fatigued quickly! :(

• Big launch on social: it can only happen once, you’ll have to spam all your friends to share and over time they might come to resent you, the algo is always working against you, and it only lasts for a few hours :(

I could keep going… but you get the picture!

Unfortunately this is the state of growth marketing. A lot of channels are not working, or are slow, expensive, or one-time only. This is the natural end state for things, and maybe we’re in a bit of a lull due to the technology super cycle as we’re 15+ years into the mobile wave, we’ve had various kinds of paid ads for 20+ years, and so on. All of these aforementioned marketing channels are now fully mature.

The Law of Shitty Clickthroughs, Redux
This is the logical ecosystem-wide conclusion for the concept I wrote about many years back, The Law of Shitty Clickthoughs which inspired by the efficient market hypothesis. It said the following:

Over time, all marketing strategies result in shitty clickthrough rates.

I encourage you to read the full essay if you haven’t, but the tldr; is that when marketing channels work, everyone jumps in on them, and they start to decay like crazy.

Why do they decay? Because customers stop responding over time — they start ignoring things as the novelty wears off. The ROI also goes down, as all the intermediaries that charge you for access to their audiences jack up prices. The super slick self-serve, automated auction model for paid marketing offers many pros and cons, first that it lets you aggregate much larger audiences with simple tools, but by making it easier, of course you end up with a ton of competition that then drives up CAC. In other words, as a marketing channel scales and ages over time — and yes, we’ve had SEO as a concept for 25+ years now — consumer engagement goes down, costs go up, competition goes up, ROI drops like crazy.

Big channels versus Little channels
All the aforementioned marketing channels are what I describe as “Big Channels.” They have scale, can be moved using $ (or labor, which costs $). And if you’re a big successful company, you might be fine because you have a (hopefully) strong brand, you get a bunch of organic traffic from word of mouth, you (hopefully) have a successful product that can cross-sell into new products, etc. You can have badly performing Big Channels because it’s all blended into a longer LTV recovery period, more organic, and all the other advantages that successful companies have.

But let’s talk startups, who have none of these positive dynamics? I want to lay out a few thoughts:

• Don’t focus on Big Channels, focus on Little Channels. first, you should just know that all the aforementioned big, mature channels will suck for you. These channels have mostly all been bid up by bigger companies, and they are mostly stale to consumers, and you don’t have the same LTV and financial strength as an established product. Instead, a new startup has to be asymmetrical — what you can do that they can’t? The natural solution points towards Little Channels, which are all the smaller marketing strategies that are tried in the early days and abandoned over time.

• Don’t worry about scaling. Let’s say you have a new product with only 100 active users. If your marketing campaign gets you +500 actives, you’re ecstatic. Gain a few hundred users inside an established company, and you should clean up your resume. Thus, little channels can work: Running mini events with cool speakers, organizing a Facebook group, going after a single college/company/town, emailing your ex-colleagues/friends to try a new thing — these can all help in the early days. You won’t have the competition, the response rates will be higher since you’re doing it all by hand, and you can always scale over time by moving onto the next channel.

• Novelty is in your favor. There are only so many ways to sell a big established product to customers. The “wow” moments have all been used, the value prop is already understood. The response rate on marketing declines as a result. But if you’re a new, bright-eyed and bushy-tailed product on the market, and you have a “wow” product that presents well in a hype video, then you can get a nice big spike when you launch. And maybe even a few more smaller ones over time.

• One-time is fine, repeatability is what you do later. A lot of marketing tactics that work in the early days only work once — like a social media launch — but that’s OK. If you can prove that a few unscalable tactics work, and it helps you gain momentum via funding/hiring/otherwise, then over time you’ll have more time to try additional strategies.

• Your product is brand new, so your marketing can be brand new as well. When you build a new product, it’s always smart to take advantage of the new tech wave, so that you can create something different than what’s existed before. You want to be building a mobile app in 2010, not a website, and you want to be building in AI now. But in the same way, you can use these new technologies on the marketing side too. What does AI allow you to do in marketing that previously couldn’t happen? Whether it’s rapidly creating personalized creative, or generating concepts faster, or creating an interactive bot — what can you do different than no one is doing, using the new tech?

• Take risks with your brand. You can attract people with your brand by being polarizing. Say “this product is not for you, it’s for these other cooler people.” Or hit competitors directly in the face, in your marketing. Be aggressive. These are things that long-term employees of the world’s trustworthy brands can’t do, because they are being protective, and managing the downside. You need to do everything you can to stand out, because being ignored is the worst outcome for any marketing tactic.

The point is — yes, today’s marketing channels sucks, but that has everything to do with the level of competition and the rate of customer fatigue. So go innovate! Try reaching people in new ways, say novel things to people they haven’t yet heard, and if you have to trade off anything, just trade off scalability. (You can deal with that later once your product is successful even at a small scale)

Product is (unfortunately) King
So I have bad news: Your product actually has to be very good. I wish I lived in a world where you could have amazing marketing and growth strategies, have a shitty product, and you would win. Then marketers would run tech, and they do not. It’s the people visionaries that create the products that run tech, and that’s a good thing!

The reason is that even if you do a ton of work to acquire a bunch of users, it won’t matter if they leak out of the DAU number. I’ve come to think of great marketing strategy as a multiplier effect on your inherent product quality. If you have a great product, you will multiply that into greatness. If you have a shitty product, you will multiply that into… well, you get it.

I think we’re seeing this in the AI wave at the moment. You have great, highly novel products and when one startup makes a hype video, it just goes viral. They don’t know anything about funnels, A/B testing, CAC, etc. The product is just killer, and the output is highly marketable, and as a result social media really works well for them.

A year from now, I think the novelty value will have faded a bit — we aren’t as impressed by the output of image generation models compared to when they first came out. If the novelty fades, as I predict, then in a few years we’ll need to really figure out how to market these products. (Sorry AI researchers, you’re going to have to learn marketing and sales!)

Every marketing channel sucks, but it’s going to be OK
This essay was not meant to be depressing, but rather to just call out the idea what we can all see — that most of the channels we work with are decades old, that the performance is teetering on an edge. Just as products are innovating by adopting new tech — AI, XR, web3, and so on — we in the marketing field need to innovate as well, by asking ourselves, where can I do something new with XYZ new tech?

And it’s very clear that we should not copy incumbent products. They have too many features — try something smaller and more targeted. But the same is true in marketing. Don’t copy the established products and try to execute in Big Channels — instead, think asymmetrically. What can you do that they can’t, and going with small channels is always a great place to start.

dear readers,

Alright, I’m back in (currently cloudy) LA after spending a couple months in San Francisco. It was super fun to be back in the bubble. It seemed, right after COVID, that so many other geos were ascendant — you saw a big explosion in LA and NY, and a lot of VCs (including us) starting offices in London. But very clearly the AI revolution has brought a lot of the action back to SF, even if folks are doing it remotely and SOMA is a bit dead. It was great to catch up with many of you over dinners, events, mixers etc. Bumped into a lot of friends during urban hikes in the city also!

I was in SF working on my big project at Andreessen Horowitz right now, which is building a16z speedrun (more details here). This is a new program where we invest up to $1M into very early stage startups and then work with them closely over 90 days, then launch them into the world. I am very involved, and it’s been fun to get my hands dirty working 1:1 and in small groups with founders early on their journey. There’s a new cycle kicking off, so would love if you could:

• recommend anyone I should be working with!

• fine to share the website below, or reach out via neutrals

• details + dates are: up to $1M investment from a16z speedrun plus $5M credits from our main partners (AWS, GCP, OpenAI, Microsoft, NVIDIA, Stripe, Deel, etc)

• the next programs kicks off July 28 to October 10, 2025 this year.

• speakers include the founders of Figma, Zynga, Zillow, DoorDash, Twilio, etc etc

• application deadline is May 11, so please get ahold of us by then

Here’s how to apply to the program.

And more details, and an FAQ.

OK — so a bit more details on how this all happened:

I started working on Speedrun at a16z because there are a ton of very early stage startups out there that are building in and around our focus of tech, entertainment and AI. Many of these early wouldn’t know what to do with a16z’s usual check size of $10 or $20 million, but they would benefit from the a16z team (now numbering nearly 600 strong!) that provide access to everything from hiring, partnerships, financing, marketing, and everything else. I love working with founders at this stage.

Also it does not slip our mind that many of the best startups on the planet were started by first time founders and also repeat founders that also have a chip on their shoulder. We feel like we’re not able to help as many entrepreneurs if we just focus on ones that have a ton of revenue and are growing quickly (though we want to work with those ones too!). So starting about 18 months ago, we kicked off a16z Speedrun which is now heading into its 5th cycle. It’s been awesome and very rewarding!

If folks have any questions, I’ll try to address stuff in the comments here on Substack. But best thing to do if you have a specific startup in mind is visit the website.

Thank you!

Andrew

The DoorDash product, version 1.0
Many of you may know I’ve been busy organizing a16z speedrun, our 12-week program based in SF/LA where we invest up to $1M in seed/pre-seed startups. As part of this, I enlisted a number of great founders to come speak to the startups, including the founders of Figma, Supercell, Zynga, Carta, Twilio, and many others, and had the opportunity to do a Q&A with Tony Xu about the early days of DoorDash. And there was a funny and informative story about how it started.

Here was the “version 1.0” of DoorDash:

• static HTML page

• 8 restaurant menus in PDF

• a google voice number that would call one of the founders

• originally called "Palo Alto Deliveries"

• took all of 45 min to build

Incredible that something so simple could eventually blossom into a product that millions of people use every day.

This essay could just be a parable about launching a new product, and the benefits of the Minimum Viable Product. But I’m not here to bore you, so we’re not here to talk about all the benefits of MVPs, since the idea has become part of the startup vernacular after Eric Ries popularized it in his book, The Lean Startup.

We don’t need to rehash things — instead, I want to talk through the lessons learned in the past decade from actually trying to put MVPs into action, and further, what it means in the era of AI. After all, isn’t it confusing that many of the AI products that are launching today had multiple years of heads-down academic research and weren’t shipped as MVPs? Is the idea of an MVP still applicable?

The problems that plague MVPs
These are important questions, alongside a number of common problems that constantly trip up as MVPs are employed during a zero-to-one period of product development. Product teams often encounter issues like:

• constant testing with inconclusive results (and thus, unclear direction)

• false negatives because of incomplete products

• works for “single player” products but hard to test within communities/networks

• hard to justify breakthroughs/research that inherently require a long build cycle

• overreliance on data versus true customer insight

• inability to compete against pre-existing products in an existing market

• local optimization into a mediocre product that is never great

… and much more.

By now many of these probably seem familiar. So now you see the real topic of this essay: we are here to talk about all the endless problems and debates that happen when you try to implement the MVP as part of your product development strategy so that you see it’s not a silver bullet.

Let’s start with a few observations on testing and interpreting results:

• you should expect your tests to be mostly inconclusive. The zero-to-one process of building new products is brutal - it's mostly just repeated failure, particularly if you are in a new category. One thing you’ll notice about v1s of products is they tend to end in ambiguous (and negative skewing) results. Did your product fail because you didn’t build enough features? Or maybe the branding was off? Or maybe you just needed better onboarding? Expect your product team to spin their wheels each time there’s a failed experiment. To make these fails useful, you need to actually run a clean test that leads to a conclusion. If the most likely outcome is failure then you need to at least be able to cross off a few things and maybe get a glimmer of success. This naturally leads to strategies like testing one thing at a time, and making the One Main Feature the core of the product experience. If it’s buried underneath a bunch of other things, then the results will be inconclusive.

• false negatives cause false restarts. The most dangerous outcome in product testing is getting false information, which is most likely to arrive in the form of false negatives due to the prior discussion about new products being about repeated failure. MVPs that are too minimal often fail because they look weak compared to existing products - they're bare-bones in features, branding, and UX. The resulting engagement metrics will likely hit rock bottom compared to what you'd see after proper iteration and refinement. And here's the real problem: you might wrongly conclude that the entire product direction is worthless. This is doubly damaging - not only might you abandon a potentially promising path, but after accumulating several of these false negatives, you'll likely get discouraged and give up. This triggers a complete product reset where you pivot to an entirely new direction and start over. It's a cycle of experimentation that generates very little actual market insight.

• expect to launch, then re-launch, then re-launch again. What’s your strategy? Of course, every test of an MVP actually requires you to follow up with another followup test. That way you can double check an insight by building on it, refining it, and seeing if there is indeed a casual link as you surmised. But running a followup test is hard, because your existing users are now tainted by the previous test — so what do you do? Perhaps they’ll reject the v2 of the product because they already tried v1. Thus, you’ll need some kind of experimentation process in which you can pull new users off of a waitlist and put them into a new experience. Or if you have a social product, you’ll want to onboard various self-contained teams onto your collaboration tool, or different highschools onto your communications app. If you don’t have a clear strategy that allows you to test across dozens of cohorts then you will be limited in the number of tests you can run, and how quickly you can learn.

While testing helps startup teams navigate the Idea Maze from MVP to market-winning product, this view overlooks something crucial: you can learn immensely from studying the successes and failures already in your market, rather than trying to recreate all that knowledge from scratch. This is why building encyclopedic market knowledge is so valuable when entering an existing market versus creating a new category. In an established market, you start with clear signals about customer needs and how different products position themselves. Yes, you'll need to build more functionality to compete with existing players, but you benefit from the accumulated domain expertise in the space.

I strongly prefer this approach over pioneering entirely new product categories. With a new category, you have no idea if there's actually a "there there." Even if you iterate to a seemingly viable product, you can't be certain it will have the business characteristics you want. Take travel products, for instance - their inherently low usage frequency and high customer acquisition costs are fundamental to the category. You could learn this the hard way through iteration, or you could recognize these characteristics upfront by understanding the market dynamics.

MVPs may or may not apply to your product category and stage
As I mentioned earlier in the essay, it’s a weird time for AI products and the MVP concept. What does it mean to MVP a foundation model product, when the initial steps of creating it might entail many hundreds of millions of dollars of training costs and R&D? Maybe it’s just to say that the “M” in MVP in this case is rather large, but I think this idea crops up in a number of places because the MVP is market-dependent.

Here’s what I mean by that:

• MVPs are often okay but not great products, and okay product lose in mature markets. Building a mobile app now is different than in 2010 when apps were brand new — today, your customers expect strong design and polish, and you are likely competing against a large number of pre-existing apps. Contrast this to the earliest days of mobile apps, when competition was low and even scrappy small apps could get big (ex: all the flashlight and fart apps). MVPs often have the problem where they can compete well in the early phase of an S-curve when the “it works” feature is all it takes, where minimal can actually mean minimal. But I am not surprised that a product like Figma or Notion took 4+ years to build the v1, because expectations are much higher in their respective categories. Further, great products often require that extra polish that is almost feels low ROI — but loyal, repeat users can often see and appreciate all of that polish, and the little bits of ROI accumulate over a long time. This has led to the observations that many of the products following the MVP theory end up with shitty UXes, iteratively bolted on, unless they take a real beat to polish everything out as they converge.

• some categories just require big upfront investment. These might sound like extreme examples, but consider startups aiming to build new nuclear reactors, supersonic aircraft, cancer treatments, or humanoid robots. While there would clearly be customers if you succeeded in building these products, it would take hundreds of millions or even billions of dollars just to create the first working version that someone would actually buy. This isn't to say startups in these categories can't become extremely valuable - they absolutely can. But the reality is they require massive upfront capital, R&D resources, and time. You simply can't create a truly minimal V1 in these spaces. And that's okay. These are obviously extreme examples, but we're seeing similar patterns emerge in other categories. The AI foundation model startups we discussed earlier share some of these characteristics. This is just to say, an MVP might not be possible and may not be a great tool in these cases. Yet these new product ideas might still be amazing opportunities.

• it’s easy to create an overreliance on data versus true customer insight. Over the past decade in tech, we've seen metrics come to dominate product strategy over qualitative insights. This is natural given our access to A/B testing, analytics, and systems like OKRs that drive rigorous execution across product organizations. The problem is that the data that tends to dominate is what's easily measurable - not necessarily what drives meaningful product outcomes. While these incremental metrics can help scale an already-successful product, they're simply not enough for zero-to-one products that need to multiply by orders of magnitude to become relevant. Product leaders need to understand the true market opportunity, make the hard calls about direction, and then leverage these quantitative tools to optimize once that bigger strategic goal has been set.

A corollary to all this — in today's tech landscape, where the product culture has turned so metrics-driven, the biggest opportunities might actually lie in areas that require intuition to discover. Taking an intuitive, qualitative approach can be faster than relying on A/B testing - especially for new products where low user numbers make data collection painfully slow.

After all, your initial product direction requires exceptional judgment. Pick the right starting point, and you'll be miles ahead of someone who chose poorly and tried to iterate frantically to success. The ubiquity of metrics-oriented thinking means that breakthrough opportunities often exist precisely where data-driven product leaders won't look. This typically means entering mature markets or categories requiring significant upfront investment. After all, if it were simple and immediately measurable, big tech companies would have already pursued it.

What does this mean about MVPs in today’s age?
Let me be clear: my concerns about MVPs shouldn't be interpreted as a wholesale rejection of the concept. I'm often the first person telling companies who've been building products in isolation for 12+ months to "just ship something already" or to choose product areas where shipping a V1 is actually feasible.

However, I've grown increasingly skeptical of certain product validation approaches I previously championed: Landing pages with email capture, social media traction metrics, or viral preview videos - while seemingly indicative of market interest - rarely translate to actual product stickiness. The key is solving existing customer problems, ones that likely have precedent and current solutions in the market, rather than attempting to create entirely new categories from scratch. The romantic notion of ideating cleverly, shipping an MVP, zooming in on a promising feature, shipping another MVP, and repeating ad infinitum often leads teams to iterate endlessly without direction. Instead, products need a strategic starting point in an attractive market - typically validated by the presence of other players in the space.

Further, the startup journey isn't really about shipping a single MVP. Instead, we need to recognize it for what it actually is: shipping MVP after MVP in a long series of potential failures, punctuated by occasional glimmers of customer interest. You'll need to repeat this cycle many times - drawing conclusions, raising money, keeping your team aligned, and maintaining customer relationships throughout. This constant iteration is what makes the startup experience so challenging. When you view it as a long-term hill-climbing exercise, you quickly realize that reducing your product vision to a single minimal version only gets you through the first few steps of what is ultimately a very long journey.

I started this essay by highlighting DoorDash's elegant MVP experiment, and while my short 40-second post resonated with many readers, it only scratches the surface. The full interview with Tony ran nearly an hour, revealing the countless iterations and tests they ran before finding product-market fit. It took several years before DoorDash's trajectory became clear - a reminder that the startup journey is a long, challenging road. I'm grateful to Tony for sharing these insights during our a16z speedrun session, offering a rare glimpse into the methodical process behind what's now a household name. The longer version of the DoorDash interview here.

(above: a vibe coded flight sim from NicolasZu, created from several thousand prompts. Included his notes on his method at the very bottom)

Vibe coding is happening, you guys
We’ve all been surprised by LLMs being good at writing/brainstorming/generating text, but along the way, we also discovered it was surprisingly good at writing code. This was first harnessed by coding co-pilot features in IDEs like Cursor, but as many of you have followed, “vibe coding” is the new thing, coined by the great Andrej Karpathy:

There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It's possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper so I barely even touch the keyboard. I ask for the dumbest things like "decrease the padding on the sidebar by half" because I'm too lazy to find it. I "Accept All" always, I don't read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension, I'd have to really read through it for a while. Sometimes the LLMs can't fix a bug so I just work around it or ask for random changes until it goes away. It's not too bad for throwaway weekend projects, but still quite amusing. I'm building a project or webapp, but it's not really coding - I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.

And of course, since this tweet from Feb 2025, we’ve seen an explosion of people playing around with all the tools to build a lot of cool things, particularly a bunch of fun little games like flight sims, tank battle games, and even first-person shooters.

So what happens next? I have a bunch of fun/random thoughts and wanted to share them:

• most code will be written (generated?) by the time rich. Thus, most code will be written by kids/students rather than software engineers. This is the same trend as video, photos, and other social media. Of course this seems surprising because today most software is being written by highly trained adults, who are generally time poor but money rich. This will change, and it means that over time, software will become dominated by youth culture the same way that social media is. Are you ready for software memes? Or memes in the form of software? It’s all coming.

• we are in the command line interface days of vibe coding. For the majority of creators, vibe coding will eventually fade, and vibe designing (with a visual paradigm) will come to dominate. People ultimately think better in a GUI-like format than a CLI-like format. Thus, in vibe designing you will show the AI the design outcomes you want, and then everything else is done for you. Yes, you may end up with tools to tweak the design details for extra controllability, and provide additional mockups that then get filled in underneath with code. But maybe folks will build software without seeing or learning a programming language.

• vibe coding could reduce the need for open source libraries as more code will be generated from scratch by AI. Code will be more of a disposable commodity, with less reuse, and instead generated on the fly for personalized use. It's interesting to see right now that creating a new project is easier than editing a project, because the latter requires a lot more context/complexity. Interesting dynamics if something like this continues

• "trad UX" and design standards give way to post-modern/fragmented software, as millions of new vibe coders create experiences with no prior know how and new perspectives. New patterns will emerge, as TikTok/YouTube has done to film making and trad entertainment. The world will go beyond buttons and modals and scrollbars and other things. Software may become unrecognizable before it coalesces again

• new bottlenecks for software. if vibe coding makes software trivial to build, then the bottlenecks shift to other places: 1) consistent creativity that stays ahead of everyone else. Anyone can write a tweet, but the best creators are the ones who consistently come up with new ideas. 2) distribution and network effects, where the first vibe coded product doesn't win, but rather the first vibe coded product that hits scale that wins

• adaptive software, specified by outcomes not code. imagine products that automatically adapt based on user behavior, rather than based on the actions of the vibe coder. For example, if the vibe coder has specified that the signup funnel should easy, then after seeing users struggle with it, the software can automatically vibe code itself to improve the flow by dropping steps or adding explanatory text. Right now we are in a paradigm where PMs specify behavior that software engineers specify in code. Imagine if PMs can specify outcomes, and the software is configured to automatically adapt to hit those outcomes

• this will accelerate software eating random long tail industries. Sort of like the advent of spreadsheets (aka programming for non-technical business people). Previously, you couldn't have gotten high-end software engineers excited about small/boring/slow-growth industries but now people within those industries will vibe code their way to greatness. Thus, those industries will get absorbed into the greater tech ecosystem. An accelerant to software eating the world

• software dev teams will change. today, there's a ratio in a typical software co of 5:1:1 of engineers to designers to PMs. What will that ratio look like in the future? Are we still going to have the currently dominant "EPD" paradigm or will the job titles be changed significantly. There's two arguments here: One version is that the ratio of engineers will go down, because things will be lower cost to build, and more of the emphasis will be in figuring out what the hell to build. But the counterargument might be a form of Jevon's Paradox, where the cheaper it is to build software, the more engineers you might want so you can build more, faster?

• marketing and sales. what does it mean to "vibe market?" or "vibe sell?" Perhaps in the same way that you can casually specify an app and an AI will vibe code it up, perhaps the same thing can happen from a marketing context. Just tell an agentic interface that you want to market your new app to kids via short form video, and it'll then go figure everything out. Hit the "execute" button and it'll go contact influencers, buy sponsored ads, and start creating customers. Same thing with a virtual sales team, but your vibe selling comes from creating outbound lists, scripts for selling, and then an interactive AI that can try to sell for you. What business functions won't be affected here?

• bugs, hacks, and other bad stuff. One of the weirdest things about building software is that, yes, the coding is hard, but so is just figuring out all the complicated logic. When you work on any product of sufficient complexity, you realize there are just lots of business decisions and edge cases and weird complex situations that need to be addressed, even when it’s rare. Can a customer redeem two coupons at the same time? If a customer wants a refund, but the delivery driver has already been paid, do you refund the whole thing or just the product? There’s a million of these things. And these issues bridge into problems of security, privacy, etc too. If two people share a folder with each other, and one person deletes the folder, does the other person also get the folder deleted? Or are they just logged off? When you make the wrong decisions here, people can get very upset. Vibe coding is fun, but addressing the infinite number of edge cases in software is not. Very curious to see how products will address this in the future.

These are just a few thoughts from a weekend of research and playing with the tools, and if you have further ideas, I hope you comment and add some more ideas.

Wait, but aren’t vibe coded apps bad?
One thing you’ll notice when play around with all the vibe coded software people are building — they are kinda bad. That is, if you didn’t know it was built in a different way, you may not be that motivated to interact with it. After all, the hyper-blocky/pixelated tank battle games are not nearly as cool as something like World of Tanks, which has been under development for 10+ years and has millions of players.

Two points on this:

First, remember what happened with the progression of images from 2D diffusion models — it’s gotten very good very fast:

And of course, the same in LLMs, video, etc. My assumption is, we are on a fast climb towards much more sophistication and complexity with AI code gen. While the current output may feel almost trivial, imagine what all this tech will look like in 5 years? In the positive case, it’ll feel like everyone has a team of 100+ engineers in their pocket, ready to build any piece of software they want.

And second, maybe the triviality of the output doesn’t matter. If you compare to prior content creation trends, you could argue that the photos that people post online are a lot worse than what professional photographers post. Same with YouTube videos and what a filmmaker would create. But it doesn’t matter, simply because the sheer quantity of content — that is comprehensively able to fill every niche and nuance of demand, and created by real people as the creators (not corporations), make social media dominant. The same could be said of software.

In the past, maybe you would only mobilize a team of 10 developers to build a product that could generate $1M in revenue and scale to a multi-million dollar company. Anything less would be sub-scale. But maybe now you’ll build a new software product the same as what you’d do with a capturing a iPhone video. You’ll do it for fun, and it’ll be ephemeral, and you might share it with a few people. But then you’ll move on and build the next thing.

How vibe coding works today
For those who are ready to explore, it’s not hard. Here’s a quick video on the topic, where a simple strategy game is built quickly so you can get a feel for it.

The tools are actually pretty easy to play around with and set up. I recommend Replit Agent if you want to get going quickly with a demo, but most people seem to be using Cursor.

Finally, I also wanted to link to NicolasZu’s recent post on how he built his flight sim game, to give you a sense for how to build something with thousands of prompts (but without coding).

I don't think the secret sauce is public yet. I can give you the fundamentals

- Grok 3 to transform a game idea in a game design document

- Grok 3 to define the best stack: front & backend for your game and ambition (multiplayer? etc)

- Grok 3 to transform the game design document into an implementation plan with super easy steps

IMPORTANT: start small. You will do this kind of document for every major upgrade (like multiplayer, or mobile, or server optimization, ...)

- Cursor setup with .cursorrules to clarify your stack (based on Grok), e.g. "optimise for mobile"

- Cursor with Sonnet 3.7 to learn the GDG and the impl plan.

- to Cursor "is it clear? what are your questions" => usually ask 9-10 questions. You answer and ask him to add to the impl file

- Sonnet 3.7 to start with phase 1. Then YOU test phase 1. You are the QA here. If phase one works, Sonnet to document what he did.

- Important: New chat between each phase

- Sonnet to go with Phase 2 ("please read the GDG and what was done in step 1). Then you test, etc.

- What if the test doesn't work? prompt away to get it done.

- Stuck? Grok 3 => your github repo => generate prompt for Claude

- Super stuck? go back 1 or several commits / prompts and try different ones.

- Hyper Stuck? Think what files could be the location of the error and drag them in the chat to investigate with Sonnet Thinking

- Project bigger than 20k LOC? Ask Sonnet to create an http://architecture.md. this file will help all future prompts.

Anyway, I am at 3000+ prompts in Cursor. I really feel like I am a whisperer 😂 I jump from Sonnet 3.5 to 4o and now to 4.5 depending on nuances I know those models have. For instance, "improve my messaging" => 4.5

If you’re reading this, the following will probably ring true:

• You work in tech, and you’re a busy person, with endless appointments on your calendar. 6-8 hours of meetings followed by 2-3 hours of email every day seems to be the usual thing

• You are constantly being asked to interview new people for your startup/company

• You’re told, and you believe, that having great talent at your company is incredibly important. And if you make a bad hire, you will be cursed for months and maybe years until that person is managed out (which sometimes, unfortunately, never happens)

• If it’s on your direct team, you might have the luxury of a long interview. This might be an hour+ where you take them to lunch, have them do a presentation, and do 1:1s. Particularly if they are an exec hire

• However, even more likely, you are just part of a longer interview loop and you’re booked 30-45 minutes to assess and ultimately offer an opinion. If you are the CEO or a VC, this is often your situation — you can stop the train, but otherwise, they will probably hire this person

So the question is: How do you form an accurate assessment of a person within 30 minutes?

In a way, this ought to be easy. After all, you might ask, how many SECONDS do you need to talk to a person (in a work or social setting) before you decide if they are smart or not? This is a favorite dinner party question, because the answer is inevitably kind of rude and inappropriate — it’s probably 15 seconds? 30 seconds? Not a long time before we all make up our minds. Interviewing is like this, but of course because you are assessing for more than raw horsepower, but actual skills, you need a bit more time. But I think 30 minutes is more than enough.

Focus on high “information per minute” questions
I’ve spent over a decade interviewing very senior people (and junior people) in my role in venture capital, and at Uber. Sometimes I interview people where I know the discipline well, like marketing/growth, and sometimes I’m interviewing people for roles where I have no clue.

The biggest waste of time is to ask questions where you can’t say “hire” or “no hire” based on the response. This is why long windup stories detailing their every career move is low information per minute, and not useful. This is why I’m mostly negative on using culture fit as a strong indicator — everyone on the interview loop will have slightly different POVs, and ultimately you won’t ever hire someone on culture fit alone. Same with the obligatory, “why are you interested in this role?” (Yes sometimes someone will say something that instantly filters them out, but let the recruiter ask this one — you should focus on the deep skills assessment)

The best interview questions filter people out very efficiently. This is what you want. So what are the kinds of questions that filter?

Here are some of the “super questions” I always try to fit in:

• The infinite depth question. For every job/industry, there ought to be a class of questions that are so open-ended, and so deep, that the candidate might take days or weeks to give a complete answer if they are actually an expert. For marketing, this might be — “tell me how you would grow Discord (or pick your app) by 3x over the next year?” The right answer might break down acquisition versus retention, their current go-to-market, what channels might make sense, etc. And then how to execute the strategy, timelines, and so on. For a non-technical person talking to an AI person, I might ask, “I type something into ChatGPT and hit enter. Tell me what happens next?” The infinite depth question reveals a lot about the comprehensiveness of knowledge they have in their industry — if you keep asking why/why/why and they can’t answer, eventually you understand the limits of what they know

Read more

Look at this graph, illustrating the rise and fall of digital cameras, and their collapse triggered by the ubiquity of smartphones. This graph is incredible of course, and it captures the simple story of technological disruption in one visual. It’s so easy! Just don’t get disrupted.

Of course it’s apt to start an essay about the coming AI disruption with a chart like this, but honestly it’s such a cliche. Duh. Boring. We all got it. After all, everyone has read the same business books, listens to the same podcasts, and knows all the same jargon and the same predictions. In fact, all the these legacy incumbent industries these days they are staffed with a slew of tech-forward millennials and MBAs who have all read the relevant Harvard Business case studies.

We know AI disruption is coming for many industries. The AI Horde, consisting of thousands of VC-backed AI startups, are starting to go after every industry as a general computing technology that can replace and augment labor. So what happens next, as they compete with incumbents?

“AI will disrupt Hollywood!!!” — really? will it?
To take an example of a single industry, scroll enough on social media and you’ll see a prediction like this: “AI will disrupt Hollywood!!!” This kind of assertion is usually accompanied with an amazing 30 second video that gives a taste of the latest in AI — the output of a bunch of video generation models, speech synthesis, and AI music, all cut in the way that is highly entertaining for 30 seconds. And the prediction mostly kinda feels right. And it is genuinely amazing how fast gen AI video models are progressing — today the output is still a little janky, but the rate of improvement is incredible.

We’re going to dig into this specific example of AI video and Hollywood/YouTube/etc, but this is being said for many industries that are thinking about AI. And I want to give the AI startup’s view, not the view of the incumbent. This is a different approach because the disruption theory, and the viewpoint presented in the Innovator’s Dilemma is focused on how an incumbent should best react and protect a well-defined product category. The POV makes sense because most management theory exists to perpetuate the incumbents, and the MBAs are hired to middle manage big, enduring conglomerates.

Contrast that to brand new startup, which has a radically different POV. The startup is just trying to survive. They are trying to find/scale product market fit, and navigate a complex system of niches and markets trying trying to find a foothold in at least one, so that they can continue to get more funding and continue their journey.

The incumbent thinks defense, and is trying to defend their particular castle. They ask, how do I protect my specific position in my industry while technologically disruptive AI startups enter it?

The AI startups think offense and acts like a barbarian AI Horde, exploring and probing and retreating, trying to figure out what will work among many possible targets. They’ll attack any castle that feels weak, and switch to another if it doesn’t look easy. They ask, if I can build products with this new tech, which market should I go after where competitors are weak? Should I double down where I am, or pivot to attack another? Should I attack the incumbents directly, or try to sell tools to them? Do I go premium and white glove, or build something simpler and low-end?

This is why the AI versus Hollywood situation is so interesting.

The big choices for AI video
The AI Horde’s upcoming decisions are complex. Let’s assert that the AI video gen tech is just going to get there eventually, and we’ll soon be appreciating video in many formats. Ahead of this inevitable outcome, let’s say you are a skilled new startup team trying to best take advantage of this upcoming trend.

The multitude of decisions might start like this:

• Should we work within the existing ecosystem?

  • Sell tools to existing entertainment companies

  • Create film/TV and sell the content (“Pixar for AI”)

  • Buy an existing film/TV studio and vertically integrate AI

• Or, should we target the social media ecosystem?

  • Sell tools to creators

  • Become a new kind of creator that uses AI to create a lot of content

  • Play some other role in the existing ecosystem (ex: build AI video etc)

• Or, build something new that exists independently?

  • Create a new app that uses AI to disrupt an existing category (like AI ReelShort)

  • Innovate on a new AI-first format (interactive? companions? a new kind of game-like experience?)

  • … and much more

These are a few of the big choices, but really there are an infinite number of decisions and sub-decisions. We could add even more detail by sitting on beanbags and brainstorming more, but this is a good place to start.

Imagine yourself as a new AI startup trying to figure out where to attack. There’s castles defending every bulleted line on the list, which is on one hand very bad, since now everyone has read all the same books and folks are rapidly writing long memos about the threat of AI. But on the other hand, the entire advantage of being part of the AI Horde is that you’re small and fast and can attack wherever is weak. And barring that, many of these bullets are B2B and you can just arm the big guys with weapons to fend off the AI horde.

Work with the incumbents? Weak-form vs strong-form?
Depending on your interests and prior background, you might have an attraction or an aversion to interacting with players in the incumbent industry — for the sake of this example, we’re talking about Hollywood specifically. (I find myself mostly disliking startup ideas that depend on these incumbents to adopt technology, since they have a long history and culture of avoiding it whenever they can, which I’ve discussed here: Why Hollywood and gaming struggle with AI). But it does come with one major advantage, which is that they have money and motivation. They also have well-known and loved brands, characters, and franchises. Theoretically a new AI startup could cut some amazing deals and have a jumpstart on their journey.

So here are some ways to approach working with Hollywood:

• Pixar for AI aka a new AI Hollywood studio: If an AI startup can hybridize with a studio, then theoretically one could create the next major film idea, sell it to a major studio (Sony, Paramount, Warner Bros, and the like) and receive financing to make the film without raising additional venture capital. This would generate revenue, and the AI studio startup could continue with a series of these, and if one is a hit, it might become a very cost efficient way to get going. This is, plus or minus, the story of Pixar, Toy Story, and Disney of course. The biggest challenge here is that you’re required from a tech perspective to build something at the highest quality level for the silver screen right off the bat, and I don’t think Clayton Christensen would approve nor would it be the easiest thing to do. But this option is there. My colleague Jon Lai wrote an essay on this whole approach (and more) and it’s worth reading. One interesting side question here is whether you’d rather back the luminary Hollywood team who then goes and adopts tech, or would you rather fund a tech founder who then sells into the industry? Both are hard, but again, depends on your tastes. Many tech VCs would rather back the tech founder, but the dream of course is Ed Catmull with John Lasseter.

• AI-enabled/tech-enabled studio: There’s another interesting approach here which is to actually buy a production studio that already has a business, team, and process, and create an AI team working alongside them to aggressive build/buy/partner on all the tech needed. We’re seeing this attempted in entertainment already, but also fields like legal, customer support, accounting, and so on. This might actually be easier to do than to try to sell software.

• Sell tools: There are a ton of workflows within the film industry that could use AI tools today. Some are obvious, like dubbing, internationalization, etc. These are nice niches that might be good starting points. Or you could look at other issues, like digital asset management or video search. Or editing, including the ability to fill in gaps of video with AI-generated content, rather than reshooting. The failure case for all of this is that you’re selling tech into companies that have tight budgets and historically have not been great customers. And maybe your customers also hate AI. But on the other hand, if you can make people more productive and/or your new AI app works on an outcomes basis (like the total project fee to dub a film, rather than as a $/seat tool) then that might work too. But also might be hard.

All of this might work, but there’s a major problem: There are major constituents within Hollywood that absolutely hate AI. Given that there’s so much pushback from unions and creatives against AI, will this approach even work? Maybe major studios will object to all uses of AI tools placate their creatives, and that means they won’t use tooling, won’t engage with companies that use AI, and won’t buy content that was created with AI. Even if it’s good for them financially, one of the big differences between the entertainment industry and the software industry is the former has very strong union control, and strikes can cripple an entire year of films. If you’re a pessimist, this is your take, and that’s why Hollywood might take a long time (like way beyond what’s reasonable) to adopt AI, and instead all the AI storytelling will happen on social media and other freer platforms.

The optimist’s POV on this is to observe that actually there is a battle between the business people and creatives in the entertainment industry — the former actually know their whole business model is broken, and are excited about AI vastly increasing quantity, promoting more creativity, and reducing costs. So they will find a way. And also that Hollywood itself is very fragmented with a zillion smaller production studios and companies, many of which are excited about AI. (We’ve been pitched by many!) These techno optimist studios are bound to do amazing artistic work, maybe initially with a lot of human-in-the-loop, and as a result will open doors. Some of the major distributors like Netflix and Amazon might be less affected by union issues, so perhaps they could buy a new AI film and open the floodgates. Or perhaps smaller distributors will prove things out in other geographies, show success there, then bring it to the US as the Overton window shifts.

Building for the digital-natives
If you decide not to build for Hollywood, then you are building for the Internet, which might mean social media or also potentially just building your own app. This could manifest in a bunch of different ways:

• AI Cocomelon: One version is to create a next-gen media network like Cocomelon (if you haven’t heard, a very valuable kids network sold for $3B recently) or Mr Beast or something else similar, but with AI-generated content. These advantage here is you don’t need anyone to green-light the content the way Hollywood works — if people watch it, you will make money. However, the business model is a residual stream of advertising against the videos over a long period of time, rather than an upfront financing that helps defray the cost of building the content. Thus, initially content in this format will look cheap/low-quality, and they need to build a huge portfolio to generate the stream of capital needed to be successful. This is also where AI video gen startups creating anime or animation (kids/adult) will be great because this tech is working well enough right now — you just need the creative talent in there to create new stories.

• Tools for social media creators: We are already seeing plenty of these, and it’s sort of a “sell picks and shovels” — helping creators with video tools that allow them to use AI to create new content. This is potentially a very large market since billions of people now use social media and most of them create content of one form or another. However, obviously very competitive and many utilitarian products lack defensibility. (Btw, I personally love Captions (an a16z startup) which has really pioneered this space!) But perhaps these tools will undoubtably become popular and am sure there will be multiple billion dollar companies here

• New app that’s an AI ReelShort/TikTok/Twitch or other video type: You could build a new app, but have it centered around the output of various AI video models to target one of the many video formats that are already out there, for example, an AI TikTok app, where the model generates engaging 30 second videos, or AI ReelShort, which is a micro-transactions driven narrative content app. Or an AI Twitch, AI OnlyFans, or something else entirely. These are all new standalone apps that are fast-following existing video formats. Perhaps the user doesn’t even know they are AI generated.

• Real-time interactive content, gaming, and other new stuff: The creation cycle of Film/TV works the way it does because there’s an impossibly long loop between shooting the content, editing the content, versus watching the content. If you can generate video in real-time, then the loop works in a completely different way. Perhaps you generate the video all on the fly, and the story all real-time as well. This makes sense particularly in a world of gaming where people want to make choices and see the world (and characters) react around them. Perhaps there is something between a game and a film. Perhaps there are games that incorporate 10x the cinematic content, and is its own genre. It’s hard to say what might live in this space of next gen interactive content, but this is probably one of the most interesting opportunities since it’s where “AI native content” actually lives.

Of course there are more ideas beyond this, but I’m cheating here a bit and saying “new formats” are a placeholder for all the cool new things we’re bound to see. The digital-native approaches are highly appealing for Hollywood outsiders, who don’t have the relationships within entertainment. It’s also appealing because you can create low-fidelity content — perhaps meme content, or shorts — and gain traction. It doesn’t require you to jump directly to what can be shown on the silver screen right away, which is a huge win. But on the other hand, all the major problems that tech startups face exist here — scarce financing, expensive marketing channels, difficult distribution for new apps, etc.

Whew, so we’ve talked about a lot of permutations now. A new AI startup might look at the maze ahead, with all the potential options and winding paths, and inevitably be faced with confusion. Ultimately that’s OK — building a startup is about thinking probabilistically, launching and iterating and learning and launching again. You might start by wanting to sell a new animated series to Hollywood, but then realize you’d rather try to sell the tool. But perhaps the customer base is slow to buy, so you release a self-serve version that’s then embraced by social media creators. You can quickly pivot from one part of the maze to the other.

Here’s why disruption theory is so limiting when you look at the view of the world from the incumbent, agonizing if the AI Horde will go after your particular market. The reality is that the Horde is moving as a wave, exploring every niche, and every new technology. They are hard to defend because it’s hard to even tell which one is actually going after you — is it the AI short film app that sucks away all your consumer time, thus tanking your streaming subscription revenue? Or is it the AI tools company that enables all your competitors to create amazing 3D animated content as you’ve been doing for decades, eroding your defensibility? The AI Horde does it all.

How this applies to other categories and AI startups
To wrap up here, I’ve spent a bunch of time talking about Hollywood and the upcoming AI video gen models but this discussion could be applied to many other markets. You often face some of the same choices:

• Do you work within the existing market, or try to create a new one?

• Is it better to sell tools, or to go directly to consumers?

• Can you accelerate by buying an existing player, and adding AI? Or do you start a new firm? Do you start with technologists, and add domain experts, or the other way around?

You could be talking about many industries here — marketing/PR, accounting, consulting, legal, customer service. And my guess is, multiple approaches might end up working.

So I want to go back to the original statement from the beginning of the essay: “AI will disrupt Hollywood!!!” — will it? I think AI will disrupt, but also be adopted and assimilated. There will be new consumption experiences that compete directly with film/TV, but also ones that indirectly compete. It’ll all happen.

After all, if you went back in time 20 years and realized that online content would have been big, you might have thought: Let’s go strong form. People will just watch internet content, don’t worry about where the incumbent media formats will go. Yet both Netflix and YouTube worked. You’d ideally have bet on both. I think the same thing will work. The AI Horde is so powerful that we’ll see it reinvent Hollywood, gaming, YouTube, and more. It’ll affect the content, yes, but also AI tools, and AI-enabled companies, and much more.

AI products are going through a phase shift
The AI boom has driven a tremendous amount of infrastructure to be created in recent years, massively opening up the group of product builders can create new AI products. As thousands of new venture-backed AI startups emerge, the insights required to build a new product will shift. Fast-following and copycatting will become the norm, as defensibility is chipped away with time — this is what happened in web and mobile, and AI products will experience the same pressures.

Recent years have been dominated by solving the Idea Maze, figuring out all the correct turns (and avoiding mistakes) needed to pick the best idea. But in future years simply picking the “best idea” will not be enough, but one will also need to figure out how to distribute them too. This is the the Growth Maze, the long series of decisions needed to pick the correct initial target audience and launch strategy, then the experimentation to slowly scale growth and modify the product over time to accommodate the strategy, and ultimately getting to a large-scale distribution strategy to reach the mainstream.

But to introduce the Growth Maze, we must first discuss the Idea Maze.

The Idea Maze
The Idea Maze is one of my favorite concepts about startups that has stood the test of time, and it comes from my friend Balaji Srinivasan (ex-a16z and Coinbase CTO). He wrote about this concept many years back in his lecture notes about entrepreneurship, and started with this diagram:

Further, he defines defines it quite simply:

[A] good founder doesn’t just have an idea, s/he has a bird’s eye view of the idea maze. Most of the time, end-users only see the solid path through the maze taken by one company. They don’t see the paths not taken by that company, and certainly don’t think much about all the dead companies that fell into various pits before reaching the customer. The maze is a reasonably good analogy. […]

A good idea means a bird’s eye view of the idea maze, understanding all the permutations of the idea and the branching of the decision tree, gaming things out to the end of each scenario. Anyone can point out the entrance to the maze, but few can think through all the branches. If you can verbally and then graphically diagram a complex decision tree with many alternatives, explaining why your particular plan to navigate the maze is superior to the ten past companies that fell into pits and twenty current competitors lost in the maze, you’ll have gone a long way towards proving that you actually have a good idea that others did not and do not have. This is where the historical perspective and market research is key; a strong new plan for navigating the idea maze usually requires an obsession with the market, a unique insight from deep thought that others did not see, a hidden door.

In other words, people often fall in love with their particular product idea and run full speed at executing it, without understanding the history of the category, context of other failed attempts, etc. These might be important clues that might explain why this particular idea hasn’t yet been built, and the dangers posed along the way.

If you can’t see the maze, how can you possibly solve it? More likely you’ll simply repeat the mistakes of others.

By the way, in my day-to-day role working at a16z, this is one of the key things I try to assess: What is the founder’s mastery of their corner of the idea maze? I ask them a few simple questions: “why choose X feature instead of Y? Why start at X instead of Y?” and “why do you think X didn’t work as a product? how do you think X product is doing in the market?” When you meet a founder who has deep observations about every turn and corner of the Idea Maze, these prompt an absolutely terrific conversation. It’s often the strongest sign that someone is an expert in their space, and they will teach you more in an hour than you might learn in months. On the other hand, it only takes a short amount of time to realize when someone is early in their thinking.

Because I don’t spend all my time in their corner of the maze, if they aren’t teaching me anything new, then that usually means they are either too early or too incapable. After all, they should be the expert, and should have spent months/years immersed in their problem. They should be teaching me with every sentence out of their mouth. (Of course, funny enough, there is always a counterpoint to all of this. The world is constantly changing, and sometimes it takes a beginner’s mind to try something once again in a new era. This is the advantage of youth — sometimes Kozmo.com fails, but Doordash succeeds, and same for Pets.com/Chewy, Friendster/Facebook, etc. You often have young entrepreneurs trying paths in the maze in new, novel ways that seem wrong to “experienced” people, and sometimes they succeed!)

For the Idea Maze in the context of AI, we are seeing the transition from one type of maze to another. The past few years have been been maze is focused on the continual technology breakthroughs we continue to see every few months, but which might get slower and less accessible, as it is quickly becoming the realm of very large, well-financed scaled enterprises (not startups) and AI PhDs. But the coming years are all about all the Idea Mazes capturing a wide variety of product categories where lines are being redrawn. Previous Idea Mazes that had been marked solved, like search engines, CRM, or email clients, might be reinvented as new doors are unlocked by AI. New categories are being formed from scratch today — although there isn’t yet a multi-billion dollar virtual companion that poses as your spouse, surely we will get one of those in the coming years (!!!). This is part of why AI is so exciting as a new, general purpose computing technology — it ought to reinvigorate broad classes of products, the same way that mobile and Internet did in prior generations.

The Growth Maze
Now let me introduce a new term: The Growth Maze, a complementary concept. It is a complex series of path-dependent series of decisions with an infinite number of choices, capturing how a new product eventually reaches its mainstream market.

The maze includes decisions like these, about product growth:

• the potential starting points: The initial audience has a huge effect on where you eventually go. Where will your product find its initial customers? (college launches? telling your friends? posting on Reddit?). Their feedback will affect both the product you build, as well as downstream decisions

• dilemmas throughout: At each turn of the Growth Maze you have big decisions: Go enterprise, or try to grow bottoms-up/PLG? Hire an expert in SEM or try an agency to launch via social media? Go with a high price point to power the economics for ads, or go freemium and try to go viral? Some decisions are reversible, but many aren’t — if you launch as one thing, it may be hard to scrub the internet of all these mentions if you change your mind

• timing: How long to stick to one strategy before jumping to another? Launching college-by-college is a time-honored way to get 10,000s of users, but not millions. Starting bottoms-up is great, but when do you eventually layer on the enterprise features and GTM?

• dead-ends, shortcuts, hidden paths, and trap doors: In your product category there may be many prior attempts that teach you about directions you could take. Let’s say you’re building an AI companion app. You might have the context to recognize the similarities of your category to dating sim games, and that’s a hidden door taking you to another maze. Or you might know that Character.ai got a ton of top-of-funnel downloads, but you might have ideas on how to improve retention.

• context dependence: Of course, every product’s path is different. You might know your UX is not great — not that useful, not that retentive — but is very viral and brings in a young userbase. This might take you down a different path than a product geared at prosumers/serious users

Unlike the Idea Maze, which you traverse by adding and removing features, the Growth Maze is traversed by iterating on new go-to-market initiatives and making decisions about the above. You might target new audiences. Or the choices might involve hiring, or length partnership discussions, or a combination of product iteration with growth intentionality.

The ultimate goal of the Growth Maze is to reach the nirvana channels (haha) — the scaled, cost-effective growth strategies that ultimately propel the product to hundreds of millions of users and revenue. Solving the Growth Maze might require precisely timed choices, jumped from one part to another, and a good dose of luck too. The Idea Maze and the Growth Maze might be interlinked — perhaps you need the idea to build an AI-first wedge app that launches to a limited but highly engaged set of users, and then in a later stage, expanding the feature set to a full suite while leveraging the initial network to grow more mainstream. Perhaps the Idea Maze and the Growth Maze are yin and yang, two sides of a coin.

The new AI Growth Maze that will be unlocked
Solving the Growth Maze might be very different in the pre-AI age versus in the coming waves of AI innovation. We all know that AI will reinvent marketing, and as such, we might see brand new branches in the Growth Maze that have never been encountered before:

• should we hire the traditional branding agency, or just ask our new brand/marketing fine-tuned AI agents pick our logo, positioning, and otherwise?

• do we grow our sales team, and prospect the old-fashioned way, or use the new top-of-funnel agentic AI sales product that does outbound for us?

• are we betting on SEM/SEO, and our customers finding us via search engines? Or do experiment with the (inevitable) hyperlinked ad units that are embedded in LLM responses and AI companion products?

• do we position ourselves as AI, or handcrafted and anti-AI?

• … and much more

It may be that the role of a marketer will change dramatically. Rather than supervising the constant iteration of campaigns, instead they will define overall strategy and let agentic marketers figure out how to generate thousands of variations of short form video creatives, and who to target amongst many thousands of audience segments. Perhaps marketers in 2030 will end up looking more from the birds eye view of the Growth Maze than how we work today, optimizing individual experiments by moving buttons around and rewriting the headlines of landing pages.

Which maze is more important?
In the early part of a new technology S-curve, the product experience is defined by the magical feeling created by the early generations of the technology. Remember the moment when you first used mobile Safari to browse the internet? How about the first time you ever uploaded a photo into a website, and you created your first social network profile? And then saw a feed of your friends’ posts? These are magical because of the “it works” feature, the fact that they can function at all. It is in these moments that systematic navigation of the Idea Maze is incredibly important. The dead ends haven’t all been mapped out, which means the maze hasn’t yet been solved. Opportunity is still there!

It is in the mid/late parts of the S-curve that simply working is not enough. When you try your 5th messaging app, the magic has worn off. After a few years of “mobile-first startups” we just dropped the mobile-first moniker and went back to calling them startups. It’s at this moment that the Idea Maze is more carefully mapped out, the trap doors have all been marked and closed, and people aren’t exploring quite as aggressively as before. This is where innovation often happens via the Growth Maze, when great distribution can couple with “good enough” product to win. And a good product with great distribution, with a few more versions, can become the best product in the market. That’s when both mazes need to be solved, and I think we’re about to see this dual race of Idea + Growth kick off, in the AI market.

Growth and network effects soon to return as a dominant force
The AI landscape has evolved a ton in the past year, with many new entrants, booming traction for many AI-first products, and existential questions for foundation model startups. These questions include:

• What if there’s low defensibility for AI model startups, and there continue to be open source alternatives and new entrants that erode advantage over time? Who ends up winning instead?

• New AI-first apps benefit from novelty effect and are seeing stunning growth. But imagine, with time, that this goes away as AI becomes an expectation, not a novelty. In a world of millions of new products, who wins the distribution game? How will products grow and reach their customers in a crowded market?

• Imagine it becomes truly trivial to copy cat another product — something as simple as, “hey AI, build me an app that does what productxyz.com does, and host it at productabc.com!” In the past, a new product might have taken a few months to copy, and enjoyed a bit of time to build its lead. But soon, perhaps it will be fast-followed nearly instantly. How will products hold onto their users?

• In recent years, innovative AI products that didn’t build their own models were derided as low-tech “GPT wrappers.” Yet consumer products for the past few decades have been low-tech, with seemingly small moats and yet have generated huge value. Will the future follow the past?

I’ll argue that it’s in this environment of a massive war between “GPT wrappers” that the traditional defensibility strategies — particularly sustained advantages in distribution and network effects — will return to the forefront. They won’t manifest in exactly the same ways, but instead, hybridize with AI features to create something new. In that way, the next gen of AI products will ride some of the forces that have driven the last few waves of computing, whether in Web 2.0 or crypto or the on-demand economy.

To explain why I believe this, let’s start with the prior theory for AI defensibility.

A failed theory of AI defensibility?
The popular theory for AI defensibility was meant to be simple, and pervaded discourse for the past few years: First was the observation that to build each successive generation of AI model, the amount of data/compute/energy required would exponentially increase. In 2024 it might stand at $100M+, but in future years it might be a billion or billions, creating a “scale effect” moat against new entrants. (See the graph above, and note the log scale for the cost!) Further, as the AI models get more powerful, they can do anything that an app built on top of it might want to do, so the vast majority of apps become merely “GPT wrappers” — commoditized bits of mobile and web UI that interface with the more powerful underlying models. In this view, the world will consist of a few large model companies who create all the value, and tax the world of GPT wrapper apps above it.

As I write this in Feb 2025, this theory seems to be facing major complications: State-of-the-art models only seem to be able to stay ~6 months ahead of their open source cousins, and new entrants seem to create near-peer capabilities on a regular basis (Grok, DeepSeek, etc). Also, the amount of data available to train upon — which initially provided a big advantage to scaled players which got access early — is nearing natural limits. And even if SOTA models take a ton of money/energy/compute to train, their competitors are able create similar performance via model distillation. All the while, the ecosystem is teeming with new app-layer startups that specialize on specific niches — creative tools, customer service, legal, and many other sectors — that show $0 to $5M+ ARR growth in under a year.

And in most cases, these startups don’t specify the underlying AI model they integrate with, nor do we care as users/customers. Is it time to cheer for the GPT wrappers? And what should be our new theory of defensibility for this new generation of AI-first apps? In a world of many many AI-first apps, which ones will stick?

And of course, Network effects. We saw that network effects played maybe the critical role in the defensibility of the last generation of workplace collaboration tools, marketplaces, social networks, etc (as I wrote about in my book, The Cold Start Problem) — and I think it could play a big role in the AI age too.

Database wrappers and CRUD apps
Hints to these questions can be 1990s-2010 S-curve cycle of building web apps and how it might apply to today, and although of course the metaphor isn’t perfect, it’s still informative. I’ve written about these concepts in my recent essay The mobile S-curve ends, and the AI S-curve begins, where initially startups in the 1990s dotcom era would raise millions of dollars simply to build the v1 of their websites, because there was so little infrastructure available — you had to put an actual physical server into a datacenter, build with a proprietary software stack with very expensive products at each layer, and marketing/growth was derived from faulty lessons from the CPG industry. Products were successful because they had the “it works” feature, and no wonder the first generation of web companies were built by Stanford Computer Science PhDs.

Two decades later, everything had changed: Websites became trivial to build as we got open source, cloud computing, and cost-per-click advertising to drive growth. Many of the most popular web apps could in fact be called “database wrappers” (actually, at the time, often called CRUD apps) — dead simple products with minimal technology that had the ability to Create/Read/Update/Delete data. Blogging and Twitter/Flickr were that. And a lot of marketplace startups, where you could post a listing, and other people could view it. Plus ecommerce websites. Web frameworks like Ruby on Rails and the entire genre of CMS software was created to make this easy. In fact, it got so easy that I remember venture capitalists asking, at the time, how products like Facebook might be defensible at all.

Of course, Web 2.0 was part of the solution to these questions — the big innovation was to not just allow an individual to do CRUD operations on their own data, but to allow entire communities/networks of people to do it with shared data. And if you kept those networks alive over time, that’s what was defensible, not the product itself. That was the essence of Web 2.0 that re-ignited consumer tech starting in ~2005 after the dotcom era had subsided. (I’ve also been told that other waves of tech, like the Windows/Mac-led GUI desktop boom of the early 90s, was also propelled by “form-based applications” created in Visual Basic).

In other words, we saw the internet make the same transition from an expensive, closed source stack in the dotcom era which opened into a much more ubiquitous, cheaper (but commoditized) stack in the Web 2.0 era. And as millions of new websites emerged, the axis of competition changed from “can you build it? can you raise the money to build it?” to “you can build it, but will consumers come? And will they stick?” I think the same wave is now coming for AI products. It won’t look the same, but instead, fuse network effects and AI into something new.

Growth and network effects in a GPT Wrapper-dominated world
People are often familiar of the definition of a network effects as, “a product where the more people that use it, the more valuable it becomes.” Products like marketplaces, social networks, workplace collaboration tools, etc are all classic examples.

In the next phase of AI products, either these new products will have to add network features, or the incumbent networked products will add AI. The question is which one will get there first?

In some cases it’s obvious how AI products will add network features. For all the products in the B2B/SMB sector, they will naturally add support for teams, and collaboration workflows (commenting, tagging, etc), and allow for sharing inside the enterprise. But in other cases, it’s less clear. Will the way that AI ultimately reinvents social networking be that the other people you interact with on the network will actually be AIs? Perhaps it’s old-fashioned, but I still think that people love to watch and interact with other humans, which is why we want to watch Magnus play chess, not two supercomputers. Or we want to watch Usain Bolt race other people, not a car. Will the various flavors of AI companionship be a full replacement of human-to-human interaction, or will it instead be something that augments? Perhaps AI-first social apps will simply let you communicate with your loved ones by sending not just image-based memes, but entirely custom interactive products to make your joke? (Imagine sending/sharing an interactive Trump companion, rather than just a funny photo). It’s hard to know how it’ll happen, but we know a lot of companies are trying.

Part of the difficulty is that we actually haven’t yet seen a fully consumer-focused win within AI. Of course we’ve seen glimmers, like Character.ai, but a very fast growing sticky AI-first consumer app is still up for grabs. There’s a lot of reasons for this — you want API costs to get much lower, to support an ads-supported/low $ sub model, and the incumbents are damn good. But maybe also the mechanics simply don’t yet work, and the ability for AI to create engaging human-level companions isn’t there yet.

Either way, let’s say all these network+AI features get added, and we see a generation of these hybridized products. Under my hypothesis, these products could still be easily copied, but they would at least benefit from the defensibility offered by their network. The question is “how?”

I’m eventually came to break down the general idea of a network effect into three underlying pillars that can be defined, put onto a roadmap, and otherwise optimized:

• First, there is an Acquisition network effect that allows products to tap into their network of users to invite, share, and otherwise gain more users. A traditional “solo” app has to buy their users with advertising, whereas a networked product can get their users to bring more users onto the platform. An AI-first product might generate really compelling or useful content that is often shared with other people, which then acquires them into the product.

• Second, a Retention/Engagement effect that allows networked products to get their users to reactivate dormant users — this might happen because of an interaction like a comment, shared file, or because they get tagged — or something more passive like a personalized email that shows activity in your network. Contrast that with a solo product that has to rely on rapidly decaying email/push notifications to get you back.

• Finally, there is a Monetization effect that allows products to take advantage of stronger business models. If a collaboration product goes “wall to wall” inside of a workplace and grows virally, it is more likely to convert to high monetization tiers. If a social gaming experience charges you for decorative goods to dress up your avatar, you’ll be more compelled by that value prop if your friends are there to see it.

Done well, you might see an AI product initially enter the market by simply providing a novel interaction — the ability to generate a new kind of content, or by going deeper into some set of workflows. But then it might add viral sharing features, that help it spread amongst friends, or within a team. You might see them start to integrate “multiplayer” use cases more broadly, and ultimately, bake all of this into an enhanced business model.

Of course there will also be B2B SaaS products that succeed the old-fashioned way as well. Rather than building a few simple network effect-driven features, instead they will deeply understand customer workflows, work with compliance, IT, and security, and otherwise build a huge company via lead bullets, not one silver bullet.

Network effects are what defended consumer products, in particular, but we will also see moats develop from the same places they came from the past decades: B2B-specific moats (workflow, compliance, security, etc), brand/UX, growth/distribution advantages, proprietary data, etc etc. The big difference is that instead of taking two decades to sort this all out, as it took between the dotcom and Web 2.0 cycles, we’re speedrunning the whole thing in just a few years.

Will the current generation of AI products win, or a new generation altogether?
Or perhaps this is giving too much credit. Perhaps one generation of startups will prove out the novelty value, and then another will successfully add networked features and eat the first-movers. This is not the only possibility. We are also seeing incumbent products rapidly embrace AI, so I wouldn’t write them off either.

There are two historical precedents worth mentioning:

In prior computing revolutions, from mainframe to desktop to GUI to web to mobile, all the new startups had the advantage that apps would have to be rewritten for the new UX. Incumbents were often caught flat-footed in the transition, and created poor next-gen apps if their expertise was a different platform (just think about Whatsapp vs AIM). This generation of AI is unusual though, in that it doesn’t come with a big reinvention of the UX. We still interact with all of these products as mobile apps, websites, and so on, rather than creating a completely new modality. Maybe incumbents that already control network effects and distribution will have an advantage, and people will rather interact with their LLMs via the Whatsapp search box rather than downloading a whole new app for it.

The other question I have is if the first-movers will actually be the ones to make it. After all, in the first generation of mobile apps after the launch of the iPhone, we saw the massive growth of early movers like Flipboard, Foursquare, Kik, and others. Yet these were not the products that ultimately emerged as the mobile winners — instead it took 5+ years for others, like Uber and Doordash, that used the technology in novel ways, to define the deca-billion dollar outcomes.

Interesting times ahead.

Product launches work differently in the age of social media.

Imagine that one day, you launch an app. Your new app is received well. A small trickle of users shows up, and they are loyal and loving. With their feedback, you iterate on the product. You ship. They send in nice comments. Feels good man. One night, a friend sends you a link showing a thumbnail of your app, with the comment, “omg this vid is hilarious.” Wow, one of your users made a video about your app, and it’s funny. Really funny. You watch it a few times. Wow there’s a lot of comments. You check your analytics, and HOLY SHIT. Signups are going straight up. A major influencer just shared your video, and then another. Even more signups are piling up. Your app gets a Reddit thread. You search on X, and there’s posts in many languages about your app. It’s the start of a wild night.

Congratulations, you’ve just gone viral on social media! Living the dream, right?

The Invasion of the Looky-Loos
Going viral is a new phenomenon, borne of the era of social media. Fueled by memetic copycatting, people view, share, and click through on the same videos and content. But when products go viral, they are invaded by Looky-Loos. What are Looky-Loos? Here’s the Oxford dictionary:

LOOKY-LOO: A person who views something for sale with no genuine intention of making a purchase.

(I learned this phrase a long time ago from a friend who used to sell cars!). When your product goes viral, yes, you do a big spike in users. But what you’re really getting is an Invasion of Looky-Loos, low quality users who come in, check things out, but don’t stick. In prior eras, this type of behavior was less extreme — there were fewer marketing channels, and fewer large scale ecosystems the size of today’s social media platforms. As a result, the top of funnel you receive in a less fragmented environment are naturally higher quality. The only comparison might be in getting a big newspaper headline written about your product, but that was rare — in a world of limited front page space and a daily cadence (not real time, like the internet), the new social media environment is much different.

The question isn’t simply to increase DAUs but rather to look at the fundamentals: Is it durable? Is it scalable? Is it valuable? Looky-Loos are none of these, and as such, I am against the constant search to “go viral” on social media.

When growth goes away
When a product goes viral, its traction can’t last. It’s not pretty. Here’s what happens next:

• The spike goes away after a few days. It isn’t plausible to sustain endless sharing, so eventually new signups will die down, and the real question is if your product retains (hint: it won’t, at least not this audience). And then you’re back where you started, looking for the next spike.

• Unfortunately, it’s impossible to create a second spike the same way. People won’t re-share the same viral video multiple times. You can’t make the same jokes a second time.

• The looky-loos arrive. The spike of new users is actually a spike of worse users. They’re less engaged, lower quality, and less valuable. They don’t convert to premium at the same rates, they don’t use the product in the same ways, and although your DAUs might be up 10x, your other metrics often aren’t.

• Not ready. Turns out your product wasn’t ready for all the international users. Or the power users. Or the teens. They want to use your product in all the wrong ways, posting bad and weird things, terrorizing your loyal base. And after a while, both sets of users want to leave.

Why does this happen? There is ultimately a tradeoff of quality and quantity for user growth, because getting more users generally leads to lower intent (and targeting) of that audience. People who find your product because they were recommended it by a friend — a highly targeted, relevant bunch — are not the same as the masses who see a funny video and click through to check it out.

If you generalize this rule more fully, you realize that traction works on an “Easy come, Easy go” basis. If users come slowly, and an audience is built over time, they are will have high retention. But when they come quickly and all at once, they will leave quickly too. Again, easy come, easy go.

Why big user spikes are bad
I’m anti-spike! I’ll say it. I always tell people to stop chasing these silly launch spikes because in the end, they’re not valuable to the business. This is all funny to say because product builders are often looking to create these spikes. They’re chasing a big launch, or a huge trend driven by viral video.

But I’ll argue that this is the wrong approach, and these crazy-fast traffic spikes are actually bad for your product. You can't just look at those up-and-to-the-right graphs and call it a win - you need to evaluate traction through multiple lenses, not just celebrate because some curves are shooting up.

Not all traction is created the same. Instead, you want to look at the traction ask evaluate it on multiple lenses:

• Durable. When product traction is durable, it means that high-quality users are finding the product, slowly but surely, and they stay engaged. They're the right target audience and they stick around. This means that you can trust that week to week or month-to-month, you are really building growth curves on growth curves.

  • Metrics: D1/7/30, consistent weekly growth (5-10% WoW), NPS (and qualitative feedback), DAU/MAU

• Scalable. When traction scales, it means your marketing channel has a repeatable motion that can be tapped again and again. These are channels with real size and depth - not just a one-off spike, but something that can be systematically activated. For example, if your growth motion is built around TikTok videos, it will probably get you some quick spikes of users, but it’s not possible to grow the number of new signups/day over time. You need to find deep channels like referrals, paid marketing, and viral growth that can scale to millions.

  • Metrics: % organic, MAU/registered (low means poorly targeted), paid marketing payback period, signups/day (particularly if it smoothly increases over time)

• Valuable. Not every customer is created equal, and the ones at the heart of your target audience will be more engaged, spend more, and interact with other users more. You need your marketing channels to deliver you valuable users, and you also need the most valuable ones to stick around

  • Metrics: % US versus international, % convert to paid, ARPDAU, ARPU and ARPPU, and key action metrics

The problem with going viral on social media is that that the traction comes quickly, and leaves quickly, so it’s not durable. The spikes you see can’t scale, and in fact, often the opposite — the initial spike you get is the biggest one, and once people have seen it, they are actually less likely to share over time. So it can’t scale. And of course, the Invasion of the Looky-Loos means they aren’t valuable.

When do spikes work?
Am I being too pessimistic about these viral spikes? Perhaps. After all, if you design your product correctly, perhaps you can filter out the bad users and just get the good ones. So here are some counter-examples:

• Waitlists. If you put a lot of users in queue and make them fill out some information, you can make sure your app for graphic designers is actually being users by graphic designers. This may not be scalable, but it’s more likely to be durable and valuable.

• Raise VC money immediately. I say this half joking, but sometimes having a big spike and a top ranking app is exactly the time you should be raising money to set you up to build something durable over time! Playing the long game.

• Targeted at the masses. Sometimes the product you’re building really is for the masses, and it’s just so great that it means the looky-loos will stick. This is of course the best possible outcome, where your spikes are retained. (But obviously very rare)

• Designed to churn. Some products are made knowing that they will have low rates of retention. The early AI image apps that show you in various AI filters come to mind — these are single player, try to get you to put in a recurring subscription right away, and ask you to share to your friends. You probably won’t use it month after month, but could an app dev end up with a few million bucks of revenue from people who forget to cancel? Sure.

• Cold start. Sometimes you’re building a networked product — a social app, a workplace collaboration tool, a marketplace, or otherwise — and you just need a lot of tonnage of users to get things going. Even if most folks are irrelevant, if you can get the 1% of creators in the 1/9/90 of creators/contributors/lurkers to show up, and you can aggregate even a few of them, maybe you just need to get started

• … I’m sure there’s many more! And in fact, if you have other ideas for how to take advantage of this, shoot me a mention on X at andrewchen and I’ll add to this list

Of course, there are always X factors. Sometimes a product utilizes so much interesting tech, or is so ground-breaking in its approach, or the team is simply so good, that you should bet on them anyway. Even if their v1 causes ephemeral spikes, you might want to bet that it’s their v2 or v3 that gets there. A lot of this is what underlies many of the very large (but probably pre-product/market fit) investments in AI startups.

But even then, it’s important to understand the risks.

Don’t forget the goal is to scale high-intent users, not the Looky-Loos
You might wonder, are the ephemerality of these social media spikes fundamental to their nature? I argue yes. My long-time readers will know that this has a distant relationship to something else I’ve written about in the past, on why A/B tests that purport to increase a metric like Signups by +10% usually don’t lead to +10% increases in revenue or active users. This essay argues that anything that increases the top of funnel by making it easier to do something always ends up letting in lower intent users who then are less motivated to go all the way through the funnel. Instead, they drop off.

A big viral social media spike is exactly a flavor of this, but out in the real world. More users more quickly equals more users leaving quickly too. And if the spike is ephemeral, those DAUs will be too.

Ultimately, startups and product builders need to remember that their goals aren’t to increase signups, or active users, which can be gamed by these spikes of Looky-Loos. Instead, the goal is to scale their high-intent, highly sticky users over time. Sometimes that is slow, gradual, and requires constant iteration.

DAU spikes don’t matter. And this is why when I talk about product/market fit, I consider metrics that generally run over long periods of time — because stickiness, and high-intent, take time to measure. Instead, I refer to metrics like this to indicate scaled product/market fit (from a tweet I posted a few years back):

10 magic metrics indicating a consumer tech startup probably has product/market fit:

1) cohort retention curves that flatten (stickiness)
2) actives/reg > 25% (validates TAM)
3) power user curve showing a smile -- with a big concentration of engaged users (you grow out from this strong core)
3) viral factor >0.5 (enough to amplify other channels)
4) dau/mau > 50% (it's part of a daily habit)
5) market-by-market (or logo-by-logo, if SaaS) comparison where denser/older networks have higher engagement over time (network effects)
6) D1/D7/D30 that exceeds 60/30/15 (daily frequency)
7) revenue or activity expansion on a *per user* basis over time -- indicates deeper engagement / habit formation
8) >60% organic acquisition with real scale (better to have zero CAC)
9) For subscription, >65% annual retention (paying users are sticking)
10) >4x annual growth rate across topline metrics

Of course, hitting these metrics is hard. Very hard. But it’s metrics like these that define an scalable, durable, valuable product.

Anything else is just ephemeral.

Hi readers,

I’m kicking off my writing in 2025 with a quick roundup of my “best of” essays list in case you missed anything from last year. Am proud that I managed to keep up some decent momentum on my writing over much of the year, and hope to do more in 2025 as well.

Top essays. Without further delay, here’s my top essays from 2024:

• How novelty effects and Dopamine Culture rule the tech industry: AI apps with shitty retention, consumers with zero patience, and what happens next

• The case against morning yoga, daily routines, and endless meetings: How to maximize 10x work and avoid thoughtless daily 1x work routines

• 10 years after "Growth Hacking": What's changed and what's new

• How AI will reinvent Marketing: What happens in a world of infinite labor, infinite content, and mass personalization

• Time sinks and money sinks: How business models drive product design and user behavior

• Why data-driven product decisions are hard (sometimes impossible): All the excuses we make when we see data, but why that's sometimes OK

• Why high growth, high churn products never seem to work: yes, D30 ends up dominating over random social media spikes

• The end of the 1 billion active user ad-supported consumer startup: And why highly-monetizing, useful, vertical apps might be the next thing

• Bureaucrat mode: The road to hell is paved via collaboration, consensus, inclusiveness, stability

• Always Be Launching: Because tbh no one gives a shit about your new product, anyway

• Why your product idea sounds too complicated: The "simple" to "WTF" scale of product complexity

Enjoy! Besides that, what am I up to? Here’s some updates.

Back in SF for a bit. We’re living in San Francisco for the next few months months as part of the a16z SPEEDRUN program, hanging out with founders from 45+ startups. We kicked off at the start of the year and have a demo day in late March, so it’s been great to work with an exceptional group of founders. The energy’s been great, and always nice to be back in town.

SE Asia holidays. We spent the holidays in SE Asia, which was a first for me. Flew in and out of Singapore, hung out with some great folks there, and then spend a few days in Sumatra, Cambodia, Laos, Hanoi/Vietnam, Bali, all while listening to audiobooks about the region. Learned a ton, and am glad to see some of these countries start to turn the corner after some dark years in decades past.

Reading. Managed to get through a bunch of nonfiction and fiction over the break. Wanted to list out a few:

• A City on Mars. Here’s a detailed discussion on what it actually takes to build a colony on Mars. tldr; it’s a lot more complicated than you’d think. Talks through everything from tech, societal issues, political issues, space law, etc.

• Things I’ll Never Forget, The Vietnam War, In the Dragon’s Shadow. Couple of the books we read on our trip, a memoir, a Ken Burns companion book, and a book discussing China’s relationship with each SE Asian country

• The Corporation That Changed the World. The history of the East India Company, including its initial formation, the competition with the Dutch (and their charter company, the VOC), how it worked, and their eventual decline. Super interesting in how it touched many parts of Asia including India, China, and SE Asia too.

• The Maniac. The history of John von Neumann, World War 2, AlphaGo, AI. What more could you want!??

• Lots of sci-fi, including Player of Games (Culture series, set in a post-AI/post-economic universe), Delta-V (space mining), Old Man’s War (military scifi), etc etc.

Fires in LA. And finally — I really appreciate many of you checking in on myself/family given the recent fires in LA. Luckily we live across town from the affected area, and have been in the Bay Area last few weeks. But obv we have many friends and colleagues that were directly affected. Really tragic. Anyway, thanks again for your kind notes.

Andrew
Lower Pac Heights, San Francisco, CA

Above: The horseless carriage (known here as a motor carriage) as the ultimate example of the adjective+noun product description. Positioned against horses!

The Simple to WTF scale
There's a complexity scale for how people describe products. Here it is, from simplest to wtf:

• SIMPLE (and easiest to understand): The product can be described in 2 words as [adjective] + [noun] like "electric car" or "smart phone" or similar. It's something you understand, but with one major change that's emphasized. If the category gets big, then eventually something like "horseless carriage" just turns into "car." (which then invites a new adjective-led category later)

• OK GOT IT: It's also easy to understand something like, "an [kind of app] for [well-understood behavior]," like an app for making a restaurant reservation or a VR app for playing basketball. The more understandable the behavior (and the more obvious why someone would want to do this) the better. If there's a clear commercial value, that will make it very easy to understand.

• HUH: Famously a lot of startups go with the "[product] for [category]" description. This can work well for products that are easily segmented, like "online dating for international students" or "voice notes for doctors." If the product is obviously useful while the segment seems valuable and big enough, then it works great. It works less well when things are too niche, or there's no real market intersection for the "X for Y" idea, like a social network for cats. (No, this isn't a good idea, don't even get started).

• UH.... WHAT: There's a close cousin to the above, which is "a [kind of app] for [literal/weird/bizarre behavior] or "it's [weird product idea] for [niche segment]". Obv if it's an app for a weird/strange behavior, like "an app for visualizing wikipedia links as geometric diagrams" then one might understand literally what the app does, but still not get the why. Or saying something is Roblox for cats. What would that even mean? This is particularly tempting for nerds who want to describe in detail what their product does, but not why, or to just make word salad because it's fun.

• WTF (like seriously what?): One more step towards descriptive would be to lead with a lot of detail about the inner mechanics of the product as the starting point. Perhaps starting with a history of an esoteric set of technologies involved as the 20 min preamble, before finally describing the product. Or talk about a complicated product strategy based on the structure of your market, and after analyzing everything, finally get to a complex family of somewhat unrelated products, followed by a super product that encompasses all of them. These pitches all follow the same pattern: 10-minute intro, question, another 10-minute spiel, more rambling, more questions, even more rambling... more confused questions. And on and on. WTF.

• DOUBLE WTF: And don't get me started if you begin your presentation with the "jobs to be done" framework, describe that, then describe your weird product, then describe how it fills some bizarre psychological need. Then 2x2s and then a weird super product with a bunch of features to be built over the next 5 years. We get it, you just got your MBA :) Just tell us what you're building now.

If you’re building a product, maybe this is starting to scare you. Maybe you’re at the WTF level and no wonder your idea has sounded so confusing over time.

How to fix your WTF description
If you're at a higher WTF level than desired, here's the easiest fix. But I warn you, it's a painful one, because it requires you to ask questions and listen.

• Show your product to target customers

• Ask them: "How would you describe this to someone else?"

• Bite your tongue and listen - you'll soon learn some simple truths

Their words that follow this question are gold. Simplicity is key. They will toss out your complicated description, and replace it with something easier and more truthful. Often times, you will dislike what you hear. Because it strips out all of your strategic differentiation, and just describes what is in front of them. Or perhaps because it doesn't capture the technical "wow" of what's under the hood. Sorry, this is the unvarnished truth of your product even if it's painful.

To dislodge this ugly but truthful idea, you have to replace it with a better idea that's prettier but also equivalently truthful. You may need to shift towards familiar product categories or behaviors and position against them in an attractive way. Your customers can only understand things through the context of what already exists in the world -- they can't talk or ask for new/innovative things. The classic quote by Henry Ford says, "If I had asked people what they wanted, they would have said faster horses" Of course it's your job to deliver these innovations, but you must learn to describe them. And cars were originally called horseless carriages for a reason -- to bridge the gap between what customers understand and what they actually want.

As a result, the simplest product pitches are often counterpositions on existing categories. Take something that's well understood, whether that's product in market, or a behavior that's commonplace, and describe your product relative to that one. If you can describe it as adjective+noun, then great, or "noun for segment" or "app for behavior." These are all attempts to use pre-existing categories.

What if your idea sounds too bland
The main downside to positioning against something pre-existing is that it'll feel too similar and bland. It'll seem to lack differentiation. Sometimes that's fine, particularly if you are early in the product cycle where having the "it works" feature is enough, or if you at the end and "it has better design" is enough. But sometimes you do need to have a strong counterpositioning.

A few things to try.

• First, try to be the anti-something. If there's a pre-existing product in the market try to differentiate by positioning against it. "We're a new kind of X" or "We're the anti-X" and follow this line of thought to what this might be. Perhaps it's mostly branding, as Pepsi and Coke are, or maybe it's something more substantial like a consumer versus enterprise focus (like Box and Dropbox). There’s a good book called Different by Youngme Moon that I sometimes recommend, which discusses this antagonistic positioning.

• Or, make a fundamental choice in UX, one that's immediately visible. Perhaps another product opens to an AI text box, and you should open to a grid of templates. Or you use animated GIFs when others use videos or lists of text. Sometimes a big "out of the box" difference in UI goes a long way, particularly if that UI helps you target a different audience.

• Go really premium and target a smaller audience. Or go towards free (or as close as you can) and go broader. Sell via partners, or go direct. Go for a younger audience, or try for an older one. Grow internationally rather than domestically. There's a lot of choices like these. These dichotomies help define the product description and oftentimes it's choices like these that can help win, not just features.

There are may other strategies of course, but these are some common approaches.

Simple is a competitive advantage
When it's easier to describe what you do, it's more memorable. It spreads faster. Your onboarding gets more efficient, and your customer acquisition gets cheaper. Simpler is a competitive advantage unto itself.

The hardest part about this for a lot of product builders is simply that the ego wants to be different. People want credit for the cleverness of their idea. And the line of thinking goes, more complex means more clever, which means they are smarter. Customers don't care about that. They just want to understand how your product fits into their life and when they should use it. That's it.

Product launches in the social media era are fundamentally different
Social media has changed the way we launch products. Instead of a Big Bang Launch — a single spike of attention — we have Always Be Launching, the idea that what you need is nonstop/constant talk about your product, its capabilities, case studies, and achievements. You don’t “save up” info for announcements, you just go. This is the optimal strategy in world of abundance — an abundance is caused by infinite feeds, plentiful influencers, and a short news cycle that devours new content every hour.

The ABL philosophy emphasizes the following:

• It’s a relationship, not an advertisement. Launching isn’t a one-time thing, but rather, a long-term narrative you’re telling the market. Sometimes punctuated by bits of big news, but often small things too

• Drum beat not big bang. A steady drumbeat of screenshots, videos, demos, comments, and discussion about what you’re building

• Wins and more wins. Publicizing small wins, and big case studies, with interviews from customers and partners

• Communicate, don’t just broadcast. Replies, comments, and discussion alongside people in the ecosystem that are discussing products in your category

• Educate and help, don’t just sell. Lessons learned, unexpected events, and counterintuitive insights gleaned from running your business

• Long-term, not one-time. A long-term view on engaging your ecosystem. Not to drive a lot of attention in one spike, but to build a large mailing list or community to share content over time

• Go direct. A purposeful disengagement from traditional gatekeepers like journalists/partners/conferences — instead, go direct and talk to folks as a human

Some of this will resonate — but remember, it was not always this way. Old school comms/PR emphasize many opposite concepts. They say, be polished and buttoned up. Hold off until you have enough for a Big Bang Launch, because you only have one chance. Suck up to gatekeepers. Hire all those agencies, pay all those experts. It comes from the dynamics of scarcity, where a newspaper only has so much space on their front page, and journalists will only write about you once in a blue moon. In the past, there were only a few important newspapers and TV channels to watch, but things have changed — we’re in the world of abundance of media, not scarcity.

The problem with the BBL is that it doesn't work anymore. Journalists want to critique tech, not advocate for it. Journalists also get their distribution via social media, because they lack it themselves. Billions of people get their content from feeds, not newspapers, and even successful traditional launches fail to drive real signups/customers. The news cycle has been shortened, and as a result, the next day everyone forgets about it. (Believe me, I've seen many founders get confused when they get covered by Techcrunch, see nothing, but then go viral on X or Hacker News, and get a flood)

The Big Bang Launch is also intoxicating to people who are a bit naturally shy, who suffer from imposter syndrome, who don't want to speak up. If you surround yourself with PR experts, who can program you with all the right things to say, then with enough preparation, then the launch will go well! Then you can put it behind you, and go back to "running the business."

ABL in the world of abundance
In this new world of feeds, Always Be Launching is the right strategy. The reason is that your 15 minutes of fame has now been shortened to 5, and realistically, your users won't care (at first) what you have to say. Announce your new startup, and just a few friends will congratulate you. Show screenshots of your demo, and you'll get a few more pings. It's only the long, gradual accumulation of these launches, with occasional random spikes -- alongside collecting emails for your mailing list, gathering followers via social -- that over time people come to know who you are.

The biggest psychological barrier is to ask founders to talk more. To write down what they're excited about, about their milestones, and to join the dialog. It's difficult because -- in the fog of war that is startup life -- it's hard to know how to be right. But it's part of the fun to lower the bar. Say one thing one day, and then do a writeup about why you're wrong the next. Launch a product, then share the metrics, but then talk about all the bugs and problems. It's the unvarnished view that gathers supporters.

Quantity is a quality all on its own. ABL builds on the natural strengths of the new social media environment. More content and more feeds on more days means that you accumulate reach over time, rather than expecting it to happen all at once. And once your following hits critical mass, and frankly, your skillset becomes more refined, you'll start getting threads and content able to pop off in a repeated way.

Authenticity helps. When you talk to other humans on social from your own personal account, and talk like a real person (not a corpobot) it makes a difference. When you share a win, or a lesson you’ve learned, people will engage in ways they’d never engage an official company account. If people genuinely like you, they will like your product. And part of the de-programming from the Big Bang mentality is to actually talk like a human instead of a corpomarketing person.

Starting the ABL journey
When you’re ready to start the Always Be Launching strategy, I have a few ideas. Step one, go to your favorite social app, and type in your latest small win at work. Hit share. Do it again every day, but include any of the following:

• a quote from a customer that you liked

• your favorite feature from your product

• something you learned at work recently

• a recent article you read in a work context

• someone who helped you along the way, and what they taught you

• … or anything else

Share something that’s a few lines, and do it again the next day. Don’t outsource it to anyone else — make sure you do it yourself, and it speaks to you. It’s really that easy. And then when you have something big to announce or say, just do it. Don’t overthink it. The next day, take that big announcement and announce it in a different way. Repeat :)

The media landscape is post-scarcity, and we need to retire the Big Bang Launch as the strategy. Always Be Launching, and launch a little each week. Go directly to your customers, your investors, and your supporters. Build over time, not in one go. And learn to tell the truth.

Above: The CIA on how to disrupt any organization. Sound familiar?

Founder mode vs Bureaucrat mode
Many of you have heard about Founder mode, the idea that there are times when you have to make real decisions, overrule people, lead via conviction not consensus, etc. Before this, Ben Horowitz wrote is the excellent essay on Wartime CEOs vs Peacetime CEOs preceding “founder mode” by over a decade.

But let’s be honest with ourselves as we read all of this. It’s very aspirational to work somewhere where you see leadership from across the company working towards high-conviction success. But it’s rare. What’s more common? Bureaucrat mode. This is what happens when companies get big, scaled, and successful.

Here’s Bureaucrat Mode:

• create committees for every decision

• make sure every meeting has pre-meetings to build the papers/deck for the meeting

• end every meeting by expanding the scope of the project

• no one ever owns a decision, so make sure there's complete consensus. Create more meetings if needed

• punish anyone showing initiative

• discipline anyone who moves quickly without complete consensus

• require detailed status reports before any progress

• create complex approval workflows for trivial tasks

• celebrate vanity metrics and milestones

• reward people based on "impact" based on how many people are working on your projects

• ask legal, brand, compliance to approve everything no matter how small

• talk endlessly about downside risk

I’m sure as you read this that this is starting to sound familiar. For all of us that have worked in scaled organizations — as Uber was when I left, at 20,000+ people — a lot of these are actually flavors of “best practices” that people purposely implement. Note that I didn’t put OKRs, QBRs, brand editorial guidelines, unnecessary legal/compliance review, etc on here but they probably should be!

It’s easy to beat up on these ideas
We can read the above list and laugh (and cry a little too) but of course they fundamentally are the result of good intentions. After all, we’re forming committees to facilitate communication when very complex initiatives like products are getting launched. There’s just a lot of details, and a lot of tradeoffs, and not everyone agrees. This is the good interpretation of this.

And to extend these root causes further, there are a few ways that the road to hell is paved with good intentions:

• collaboration: Let’s get everyone to work together :)

• consensus: The need to address everyone’s concerns, to avoid mistakes, so that everyone is supportive of the decision

• inclusiveness: Making sure everyone’s opinions are heard

• stability: The core business is the gravy train, and why endanger it?

• empowerment/delegation: You don’t want to step on toes! Let’s trust the various teams do their work!

• accountability: We have KPIs and goals and playbook, and def not worth distracting ourselves

I’m sure these all sound familiar — we’ve all said words like this! Taken independently, of course these are positive+helpful cultural values, and when you take these and then implement them in the form of processes/committees/etc., they can be great. But when industrialized on a massive scale as large tech companies do, it becomes hard to get anything done.

Of course, this is where startups have a huge advantage over over large companies. When you have 2-3 people, there’s no consensus that needs to be reached over weeks of meetings — all the information is already held within peoples’ heads, as they work from the same room. You can move incredibly fast and just focus on output, because there are less social relationships to manage. It’s fine to disagree, because it either takes a moment to sort out, or you can just try stuff out, and undo if it doesn’t work.

But all of this bureaucracy is not just driven by good intention. What makes this Bureaucrat mode and not Collaboration mode is that often these mechanisms are hijacked by people who forget why the mechanisms exist in the first place, and instead use the machine to drive their own careers.

Self-replicating bureaucrats
If you create an organization where “impact” is measured by how much your team is outputting — and thus, it correlates with the size of your team — then you are going to create a massive incentive to pitch all sorts of large scale projects that require hiring. If people see that other people getting promoted requires them to manage people, so that their responsibilities and scope are vast, rather than the success of their output — well, you are going to creative an incentive to hire a ton of folks. If big visible projects (“Project XYZ!”) end up being what’s required to drive internal visibility, and thus promotions, small impactful things will be ignored and big grandstanding projects will end up being encouraged. Committees will be formed for reasons other than building consensus.

This creates the phenomenon of self-replicating bureaucrats:

If winners hire winners, and losers hire losers, what do bureaucrats hire? More bureaucrats of course.

The reason is that companies that highly prize consensus, process, etc., will inevitably hire the people who are good at executing against this set of constraints. This continues and continues, until the moment the company is required to actually move nimbly to face off against an entrepreneurial new startup (example: car companies versus Tesla) or a big technology trend occurs (example: AI and Europe). Because there’s so much that’s unknowable about these situations, and so much of what’s required is just to try things and learn things fall apart. The highly consensus-driven, collaborative organization that has become staffed with self-replicating bureaucrats end up not being able to bureaucrat themselves out of the situation.

The cycle of life
This phenomenon is so ubiquitous that it’s almost a cycle of life within tech.

• A new, nimble startup with an aggressive new founder(s) emerges

• To scale, it hires well-intentioned, competent managers

• It wins the market (woohoo!) and IPOs

• Later, bureaucrats who are attracted to peacetime (and brand, and stability) sneak into the company. They have shiny resumes

• The entrepreneurial people quit, or leave, and can’t deal with the new processes. The bureaucrats take over. The founder either checks out, or retires

• The company is in Bureaucrat Mode

• A new, nimble startup then emerges…

Without this cycle of life, the tech industry would not exist. I saw this first hand at Uber, which was respected as the fastest-moving big company led by an aggressive founder, and eventually things got bogged down as it grew. Very hard to counteract, even with a company where “moving fast” was part of the core DNA.

In tech, at least we have a cycle of life where new startups can take over.

In Europe however… :)