The first thing a developer building an AI app does is pick a model. A few weeks after launch, they pick a router.

But it wasn’t always like this. A year ago, most startups picked one model and built everything around it. Today they’re using a handful, picking different models to cut token spend and get better results on specific tasks. Ever noticed how AI apps often hide the model name and just say “Default”? That’s routing.

Most people think we’re headed for a world with one or two dominant model providers, where frontier labs keep getting stronger and developer lock-in keeps deepening. If that’s right, routing doesn’t really matter and this post is pointless. In that future, there’s nothing to route between. You just pick a model and it does everything.

But that’s not what’s happening. Things feel like they’re moving in the other direction. Cursor built its own coding model, Harvey fine-tunes for legal, and Suno built its own model for music. Maybe most importantly, the number of open source models with real adoption is exploding. It’s becoming hard to ignore that specialized models win on specific jobs.

Is routing the AI era’s load balancer? Probably. The developer picks the router, and the router picks the model, manages cost and latency, and reroutes when models go down. Whoever owns the router sits underneath everything else. The margins for a router business are still an open question, but the importance of the layer itself is not.

The real question is whether routing stays a feature that platforms bolt on, or becomes its own thing with real network effects. That’s the kind of question we’re always asking at USV.

It comes down to whether routing gets better the more traffic it sees, and right now it looks like it does. A router that sees traffic from thousands of apps knows which models actually work for which jobs. A single app never will.

Like most of us, I’ve traded my privacy for convenience many times over the past few decades. But I’m starting to wonder if handing my life over to agents will finally force me to reconsider.

A lot of us are setting up our own OpenClaw, Hermes, Pi, or Cowork agents. As we customize and personalize these agents, we’re making privacy-related decisions ranging from completely air-gapped Mac Minis to full blown YOLO. A friend of mine recently set up a machine that has virtually no real information about her within reach of the device. Another resigned himself to the fact that his agents will become his virtual twins and gave them access to all his personal accounts, credit cards, etc.

I’m not sure where I stand, or where I’ll end up. But it’s clear to me that we’re headed towards a world that looks very different than the one a decade ago, when my social security number was leaked and subsequently used to sign up for a dozen different Verizon Wireless accounts. At the time, I couldn’t imagine a worse outcome of a security breach.

But now, I’m intentionally installing what is effectively spyware on my computers. I’m having my agents handle increasingly personal tasks for me. Buying things. Messaging friends and family members. Helping me with important career decisions. I’m handing them more of my private information and more of my agency than the me from the Verizon incident would have ever dreamed possible.

As a VC, I’m naturally wondering what this means for the future of personal and enterprise computing and where the opportunities lie, as are my partners at USV. We’re talking about it a lot.

The obvious answer is that Anthropic and OpenAI will build more powerful harnesses that will enable them to have even better models, which will unlock even better harnesses, and so on. We, as normal convenience-happy customers, will hand over more and more of our personal data. A few months ago, Ben Thompson made a strong case for this outcome (specifically that the labs that control both the model and the harness will win) and I believe there’s a good chance he will be right. But might the complete opposite thing happen?

The contrarian bet would go something like this: people will eventually understand that their most sensitive context is sitting in someone else’s data center, training someone else’s next model, subject to someone else’s terms of service, and maybe used to do real things we never asked to be done. The same way people eventually demanded control over their health data and their financial data, they may eventually (finally?) demand control over their digital footprint.

The more power people concede to their agents, the more they'll want to know they can trust them. But it won't just be about trust or security. It will also, just like before, be about convenience. People didn't demand health data portability solely because they cared about data rights in the abstract. They demanded it because they switched doctors and their records didn't follow them and it was a pain in the ass. There may be a version of that moment coming for agents.

When that happens, the value won’t accrue to the lab. Instead, it will accrue to whoever provides the user-controlled memory substrate that is both trusted and portable. One of my partners at USV believes that this is where the world is headed, and I’m increasingly coming around to that future, too.

But I’m not even sure I want that future. After all, convenience always wins. And the easier path is to keep handing things over and assume the incentives eventually align. But I’m spending enough time with agents now that I’m starting to wonder. At some point, I’d like to know who they’re really working for.

Systems of record are no longer defensible. For decades, the information inside data-heavy apps was trapped, because moving it cost more than just living with the system you had.

But now, sufficiently capable agents can just raid the systems of record by manually extracting all of the data, row by row, and putting it elsewhere. It might take time, but agents have unlimited time (as long as their human is willing to keep paying for their tokens).

A system of record is really just a database of facts, and facts are portable. There are AI startups that are actively deploying agents to systematically tear down a customer’s CRM on another platform. It’s kind of brilliant. And this playbook will repeat for other types of applications, too: chatbots’ memories, agent harnesses’ customizations.

Network effects are the answer. Like they’ve always been. To make a system of record actually defensible in an agent-first world, you have to have network effects.

An agent can copy every row in your CRM, but it can’t make your customers open the new vendor’s emails.

An agent can scrape every post, photo, and follower graph off Instagram. But it cannot make your friends open a different app to see what you posted.

An agent can distill a model. In fact foundation models themselves are becoming commoditized. Every lab is converging on similar capabilities. Even the data a model was trained on is portable (the open-source community has proven this repeatedly). What isn’t portable is the user base actively using a product and generating the preference and interaction data that makes the next version better.

It always comes back to network effects.

Spotify recently launched a very interesting CLI tool that makes it possible for agents to push custom audio into a user’s Spotify library. I find this new feature very interesting.

Spotify learned a long time ago that people love having all of their different types of audio (music, podcasts, etc) all in one place.

This is a big reason why Spotify ended up wrestling the podcast opportunity away from Apple years ago: while the latter was committed to having separate apps for music and podcasts, Spotify saw an opportunity to bundle them together and cross-pollinate the user behaviors with a single tap, not a heavy switch to a different app.

The next realization they had was that they could follow a similar playbook with audiobooks and win that market, too. And they've stated many times that more formats are on the way.

The feature Gustav is announcing below is clearly a fun way for people building agents to push new types of custom podcasts into their Spotify libraries. But I suspect it's also an extension of the formats strategy articulated above.

We're seeing lots of new products in the app layer that leverage AI and new TTS models to create awesome new audio experiences. Products like Huxe, Particle, Pingo, and Oboe come to mind. But there are many others.

But with the Save to Spotify CLI, Spotify is not only empowering individuals building agents, they're also making it really easy for developers to push new audio experiences directly into Spotify.

This is actually a huge win for all parties: developers don't have to reinvent the wheel and build a robust audio player/experience, Spotify gets more audio content, and users get to listen to new audio where they listen to everything else.

Like many others, we’re spending a lot of time at USV talking about where moats will accumulate for AI companies. And while we don’t all agree yet, it’s been a helpful exercise, as well as a fun one (because as a byproduct, we all experiment a lot with various models, agent platforms, harnesses, etc).

For a while, it was very en vogue to say things like “memory is the moat” in X posts, on podcasts, etc. That never really resonated with me, personally. Any memory that a model has about you can actually be easily ported, because memories are just facts. If you get a model to spit out the facts, you can just take those facts and put them into another model. A great example of this was a few months ago when the Claude app overtook ChatGPT in the App Store. To help make it easier for users to switch over, Anthropic released a prompt that got ChatGPT to spit out everything it knew about you. I tried it out and it worked perfectly.

But maybe something similar to memory is sticky. Not memory of facts but memory of customizations, workflows, tools, skills, etc.

The more I customize my Claude desktop app, the stickier it becomes. A small part of that may be memories about me. But the bigger part is the way I’ve customized my set up: my nutritionist project, my doctor project, all of the various skills I’ve either written or imported, the workflows that I’ve chained together that Claude can go and execute on once I ask questions. This is arguably even more powerful in something like agent hosting platforms such as Tasklet, where the volume of customizations is much greater, because they do so much more for you than the Claude desktop app. Tasklet can just go call a tool or host your mini app on a server somewhere easily, and all of those bells and whistles even further entrench you.

And yet OpenClaw, Hermes, and similar harnesses take customization (and therefore, potentially lock-in) to a whole new level. There are infinite permutations for setting up your OpenClaw instance. And Hermes has workflows that are self-learning and codify themselves as you execute them, so that’s another layer of customization that happens automatically.

So where I net out on this (subject to change, of course): I think the customizations that get deeply embedded into a harness are the moat. The more you invest in the harness from a personal standpoint, the stickier it is, and the harder it is to leave.

On the other hand... maybe I’m completely wrong about memory not being a moat. The more a model knows about me, the more it can proactively address my needs. While we haven’t yet seen a breakthrough of “proactive” agent use cases out of the box in Claude, ChatGPT, etc, I suspect we will eventually. And the more the models know about you, the better they’ll be at proactively providing us with the best overall experience. That sounds pretty sticky.

I’m spending a lot of time using AI to build out my own “personal back office.” Many friends, founders I work with, and my partners at USV are, too.

Before AI, it was not feasible from a cost or time perspective to invest in building custom software to help one automate and/or manage various aspects of our non-professional lives. But with AI, we can delegate away tasks that give us superpowers or that we simply find annoying or tedious.

Just a few examples:

I’ve been iterating on an Advisor agent that serves as a public equities advisor. First I used it to translate a specific investment thesis into a strategy. Then, it took that strategy to build me a portfolio that it automatically monitors daily, evaluates for excessive drift or thesis changes, and rebalances as necessary. I may open source this agent soon.

I also have a Nutritionist agent that pulls in all of my personal health data (historical blood tests, x-rays and MRIs, VO2 max tests, DEXA, plus real-time Apple Health data) and maps it against my daily meals to help me hit my current fitness goals. It checks in with me around meal time to make sure I’m logging accordingly, and then lets me know how I’m doing on my calorie and protein goals throughout the day.

Here’s a different kind: my partner Fred Wilson built a Restaurant Recommendation agent that actively monitors a number of his trusted sources for restaurants that it believes he will like and adds them to a database. Then, if Fred wants to eat at any of them, it automatically accesses his Resy, OpenTable, Doordash, and Blackbird accounts to order food, book reservations, etc.

I’m sure many reading this have their own AI “personal back office” stack, too.

There is a real possibility that over the long run, I’ll end up ditching a lot of the stuff I’ve built and that the value will have come more from the fun of building the back office than the actual output. Will Manidis’ excellent essay Tool Shaped Objects captures this well.

But I’m optimistic that for every few abandoned agents I leave behind, something real will stick.

And I’m excited to see the tools other people build, along with the full-fledged products startups will create to support the new “personal back office” movement.

More than a decade ago, I had the privilege of pitching USV when I was raising capital for the podcasting startup I was building at the time, Anchor. While I dreaded many fundraising pitches, I was excited for this one, for a few reasons. First, USV was one of the world’s most legendary and best performing venture firms, especially in New York, where I’m from and where we were building the company. Second, I knew they were into products that more mainstream VCs wouldn’t touch, and few others seemed willing to back a podcast company at the time. Third, I had learned a lot about building my startup as a daily reader of Fred Wilson’s blog, AVC.

We didn’t get the investment. But the conversations stuck with me: it was obvious that they were students of the world’s best products and founders, deeply opinionated, and genuinely curious about things that were still too weird for most VCs to take seriously. I got to know them much better over time, as both a friend to the firm and eventually an LP. And what I consistently heard over the years since was that USV’s reputation with both their founders and LPs is every bit as strong as their returns. Occasionally, after exiting Anchor to Spotify and becoming a VC myself, I found myself channeling their sensibility through my own investments when they still looked strange to most people.

USV describes itself as “a conversation”. They form their strategy by talking to each other. A lot. They do this as a very small and collegial partnership, in person, in NYC. It’s how I believe I’ve done my best work, too: building Anchor with my co-founder Nir, in a small office, where we’d talk strategy and whiteboard obsessively, sometimes for weeks, before touching a line of code. Or at Spotify post-acquisition, with Gustav Söderström, Daniel Ek, and the team, where we’d debate deeply, and align as a 6,000+ person company before building anything. Gustav has a famous saying within Spotify: “talk is cheap, so you should talk a lot.” His point: building the wrong thing is expensive, so earn the right to build by thinking hard first. USV operates the same way. Talk a lot, form true clarity of thought, and don’t be afraid to look weird before being right.

Venture capital has a prevailing strategy right now, and it’s not a secret: raise a war chest of capital and deploy aggressively into companies that look like they can become category leaders. Capital and velocity are the weapons. And it has worked extraordinarily well. The funds that deployed into consensus-winning AI companies over the past few years, backing the infrastructure, the foundation models, and the clear category leaders early, will generate some of the best returns this industry has ever seen.

I also believe the AI era will produce something the last few years haven’t: an enormous proliferation of abundant software, abundant founders, and new business models where knowing what you’re looking for before it arrives becomes a distinct advantage. The total volume of what will be built over the next decade will dwarf anything we’ve seen before. In that environment, I believe the winning strategy will be to know exactly what you’re looking for before it arrives. In other words: to be thesis-driven. That’s always been the USV playbook, and I think it’s exactly the right one to be running in the years ahead of us.

I’m grateful to everyone who has shaped how I think and invest. I’m excited for what comes next. As always, I’ll be based in New York City but investing all over the world, mostly at the seed and Series A stages. If you’re building something on the edge, something that doesn’t automate, but instead obliterates, we’d love to hear from you.

Late last year, the AI note-taking product Granola released a feature called Crunched, their take on an end-of-year, Spotify Wrapped–style recap (Disclosure: Lightspeed is an investor in Granola). Crunched left me stunned. It made me realize just how much Granola had learned about me after transcribing many of my meetings throughout 2025. And judging by my X timeline, plenty of others felt the same. It made me wonder: What does ChatGPT know about me?

So I asked both Chat and Granola what they knew about me. I asked them to get specific, deep, personal. The differences were immediately clear: Chat was great at knowing the things I had specifically asked about: flight hacks, how to interpret my latest blood test results, things I liked to write about (which it affiliated with my interests), and problems I had tried to solve throughout the year.

Granola, on the other hand, seemed to have a deeper understanding of me as a person. My values, relationships, emotional intelligence, actual interests. It knew the things I had said out loud throughout the many conversations it had transcribed. Some of these were in group settings. And some were one-on-ones that bounced between professional and personal subject matter. But none of it specific questions I had asked the AI.

The differences in these two products’ understanding of me made one thing very clear: there’s a stark difference between active context and passive context.

Active Context

Most of what AI knows about us, it knows because we chose to share it. Asking ChatGPT for travel recommendations or daily workout plans is an example of Active Context. Querying Google or Gemini for a phone number or research help is another example.

If you directly and consciously engage with an AI, chances are you’re giving it Active Context. And OpenAI is clearly extremely well positioned to be the King of Active Context given how synonymous it’s become among consumers with AI question answering.

Passive Context

But there’s clearly magic, and potentially even greater value, in passive context. What Granola observes in meetings is one example. Access to our email inboxes is another. Crunched was a mind-blowing experience, but a few months earlier, I had a similarly impressive experience chatting with Poke, which made me negotiate with it to access its subscription tier.

Throughout our negotiation, it became clear that Poke knew an impressive amount about me just by doing inference on my email inbox. As one example, it knew I was both a VC and someone who was actively paying for a Whoop subscription, and therefore assumed I could easily afford a $9.99 Poke subscription.

Companies with significant stores of Passive Context include Google with email and calendar and docs, Apple with messages and email and health data, and Meta with interests and browsing behavior.

What AI Wants to Know

It seems likely that companies with a lot of Active Context will want more Passive Context, and that companies with a lot of Passive Context will want more Active Context.

We’ve seen ChatGPT launch several products aimed at capturing audio (Passive Context), and recent rumors suggest that this functionality will be the focus of their upcoming hardware device. Meta, on the other hand, seems to be investing heavily in consumer AI products, which if they nail would give them access to more direct Active Context.

While both are valuable, I expect Passive Context will ultimately be the more important, especially for consumer businesses. In consumer, products need to eventually be ad-supported (because ads are how consumer products become ubiquitous and available to all).

The consumer AI product that wins will be the one that gets people comfortable with an AI that can passively observe in the background.

Since I was a little kid, I’ve always loved watching live sports. But over the past few years, something has shifted. I’m enjoying them more than ever. I crave turning on the Knicks on a weekday evening to watch Jalen Brunson play heroball in the fourth quarter. Or hearing Scott Hanson take me through every scoring drive on RedZone. Or, when he was still on the Mets, seeing Edwin Díaz trot out from the bullpen to close out a ninth-inning high-wire act.

The true unpredictability of live sports seems to grow more irresistible with each passing year. Is it because I’m getting older? Am I turning into my father, who for my entire life has been a die-hard fan of sports at all levels? Or is something else at play?

Yesterday, Lulu Cheng Meservey published another brilliant essay called Standing Out in 2026. In it, she writes:

In this world, the real has never been more precious, refreshing, special, rare.

As the creator of tech’s “Go Direct” manifesto, Lulu is, of course, referring to how CEOs and tech leaders should tell their stories in an AI-first world. But I can’t help wondering if the realness of live sports is the key reason I feel so captivated by them right now.

Sunday night’s division-settling game between the Pittsburgh Steelers and the Baltimore Ravens is a great example. There were multiple moments in the fourth quarter alone when anyone watching was certain they knew the outcome, only to be proven wrong. It was thrilling and unpredictable in a way I no longer recognize in most other forms of content I consume.

Social media feels staged or purely generated. Modern music is formulaic. TV is algorithm optimized. Movies are mostly repackaged IP (key word mostly: film still feels like the medium where we’re seeing pockets of true creativity and there were many standout examples in 2025).

But sports remain dynamic. A safe haven for true, unpredictable realness. As we increasingly question the authenticity of nearly everything, I expect more of us to find comfort in good, old fashioned athletic competition, whether it’s a slow, boring Tuesday night baseball game or the bleak heartbreak of a season-ending, blown field goal.

The first phase of social media was defined by social graphs. Creators amassed followers, and when they posted content, it would be distributed to their followers’ newsfeeds. This guaranteed that at least some followers would consume a creator’s content, even if it wasn’t perfectly tailored to their interests. It wasn’t very efficient, but it worked.

Then came the second phase, defined by “the algorithm,” popularized by TikTok, and copied by Reels and every other platform. I called it “Recommendation Media” instead of social media (and wrote about it in The End of Social Media). This phase was far more efficient: creators no longer needed to build up a follower count to get distribution. They simply had to make good content and let the algorithm do the rest, finding the perfect consumer for any given video. This made content distribution dramatically more efficient, and platforms were rewarded with much higher engagement (a.k.a. addiction).

Ben Thompson summed it up perfectly in a conversation I had with him on this topic:

“The other angle is that to succeed with recommendation media, to be a recommendation engine, which I think is honestly another way to say to an aggregator, you need volume, you need huge amounts of stuff on your platform on both sides, both on the consumption side and the creation side. How do you get big volume? You get big volume through innovation, through doing something unique and so it touches on both sides of it in a way.”

We’re now entering the third (and final) phase of social media. Platforms love distribution efficiency because it drives consumption, and in this third phase, they’re guaranteed to get it. As generative models improve, platforms will pair on-the-fly content creation with algorithmic distribution. Instead of sourcing from creators, they can generate content tailored to individual preferences at scale. Production shifts from human to synthetic. With these models (like the ones we’re seeing in Sora and Vibes), platforms will increasingly generate content dynamically to perfectly match a consumer’s interests, rather than selecting from an ocean of human-made videos.

The good news is that people will keep creating, because it’s what we do. But the incentives will be very different.

On one end, we’ll see creators moving toward differentiated human experiences that resist commodification. Live events, community, and shared experiences are already on the rise, and that trend will accelerate. On the other end, we’ll see platforms rewarding uniqueness instead of eyeballs, likely through direct payments to creators when they feed the models (but skipping the distribution part). People will no longer be incentivized to simply reach as many viewers as possible, because platforms will be able to do that without them.

What platforms’ AI won’t be able to generate is something truly unique: a human’s name or likeness, the feeling a revered brand evokes, or a net-new meme or joke that only a human could dream up. We’re all going to help the models get smarter, until they no longer need us.

If there’s any silver lining to this third phase, it’s that some form of creativity will continue to be rewarded, at least for the foreseeable future.

Overall, the system will keep optimizing for distribution efficiency, and we’re fast approaching the point where platforms no longer need creators to feed the machine.

AI is going to accelerate the meme economy. By a lot.

Everyone knows that the world runs on memes, which is why we're all so performative on X, TikTok, Instagram, etc. Even in traditional media, everyone performs when the camera is rolling.

Every person at yesterday's White House press conference (on both sides - US and Ukraine) had likely planned out their exact individual performances in advance. It being the most viral clip on the internet is not an accident. No one knew exactly how it would play out but they certainly all wanted it to go viral.

AI feeds on context, and we're increasingly willing to give it all the context it needs. And so the cameras are going to be rolling much more often. And the mics recording. It's not a coincidence that more and more of us are recording our Zoom meetings. And some, our IRL meetings as well.

As AI enhances every aspect of our lives, we'll all work harder to influence the AI's context window, especially as content in all settings becomes a commodity. The more unique our performance, the more likely we'll be able to get the AI to meet our objectives.

SEO, influencing the TikTok algorithm, etc — this was all just the warmup act.

“Wrapper” isn’t a dirty word. It means a product built on top of a technology. Almost every great product you’ve used it likely a wrapper of some other technology. Stripe, Robinhood, and Uber made existing technologies more accessible and valuable to everyday users. The same is true for Google, Microsoft, and Apple. Everything is a wrapper of a wrapper of a wrapper.

Yet somehow, in 2025, we’re obsessing over the underlying technology of AI. So much so, that today, January 27th, Wall Street is punishing the tech sector as a result of DeepSeek, the China-based research lab and model. The breakthrough of DeepSeek is driving down stocks like NVDA as much as -13% because it proved that state of the art AI models can be built and deployed for fractions of what it’s cost companies like OpenAI, Google, and Meta.

Does this matter? In the short term, absolutely yes. It challenges the fundamentals of many of the most important technology companies over the past 24 months. But on the other hand — no, it doesn’t. If anything, this will only accelerate the race to AI and great technology getting into products and the hands of customers.

Startups that focus on technology over product may have advantages in the near term. The “wow” factor. On the other hand, they are far less durable over the long term. Over a long enough period of time, technology always commodifies — especially on the internet.

Products, on the other hand, can live forever. Apple, Disney, Coca Cola, Ford, Nike, and the list goes on. There’s nothing more durable and less commodified than great products which are beloved by actual paying customers. And so not only is “wrapper” not a dirty word — it’s a winning strategy.

I’m looking forward to seeding more simple wrappers that consumers love but overly intellectual VCs hate.

About 10 years ago, my coworker Nir Zicherman and I got talking about podcasts. We felt that podcasts were exploding as a medium, but it was really hard for most people to make them. That conversation turned into an idea, which turned into a prototype, which turned into a company called Anchor, which somehow turned into the biggest podcasting platform in the world. Spotify ended up acquiring us.

More recently, we’ve been talking about education. We can’t make sense of the fact that despite all the technological advancements made over the past several decades — especially in AI — humans seem to be getting less intelligent. That conversation has also now turned into a company, and it’s called Oboe.

Our mission is make humanity smarter, and we’re building products that leverage AI to help people learn 10x more efficiently, effectively, and affordably than ever before. We do this through LLMs and a novel method of education that Nir has been building over the past year. While we’re co-founding this together, Nir is CEO this time around, and I’ll remain a full time partner at Lightspeed.

To help us get started, we’ve raised $4M from some of our former Anchor investors and a few of the smartest humans we know, including Eniac Ventures, Haystack, Factorial Capital, Homebrew, Offline Ventures, Scott Belsky, Kayvon Beykpour, Nikita Bier, Tim Ferriss, and Matt Lieber.

We’re currently hiring the core team—in particular, a product designer. If you or someone you know is based in the NYC area and is interested in joining us, please reach out at jobs@oboe.fyi. And if you're interested in beta testing Oboe once it's ready, sign up here and follow us on X @oboelabs.

Arthur C. Clark once said that “any sufficiently advanced technology is indistinguishable from magic.” The magic-like experiences each of us have felt using AI over the past few years help explain why so many startups have found instant Product Market Fit. Today’s AI products are so magical you just can’t help but want to use them. ChatGPT is the fastest growing consumer product ever. And public launches of other AI-native products like Waymo, Suno, Midjourney, and perhaps even Google’s new Notebook LM podcast feature are other recent examples of instant Product Market Fit. As one AI founder recently said to me: “it doesn’t make sense for us to launch the product until we reach a technical breakthrough, because the breakthrough is the product.”

This isn’t your typical Product Market Fit; instead, it’s what I call Magic Market Fit: when a technological breakthrough is so magical that it instantly meets a market demand. Unlike Product Market Fit (PMF), which typically requires months or even years of experimentation and constant iteration across product, design, engineering, and go-to market efforts, Magic Market Fit is met the moment a technical breakthrough is achieved. ChatGPT is a great example of this phenomenon. The design and user experience is ridiculously simple – it’s a basic chat interface. On the other hand, the technology underneath it is anything but simple, and the results are so impressive that people can’t help but be drawn to it. 

In the absence of technology, startups needed to hack their way to PMF through product and design.

The traditional search for PMF looks quite different from what we’re seeing today with AI, and products built during the mobile revolution of the 2010s serve as a good example of this distinction. After the launch of the iPhone and Android platforms, builders of early mobile apps had to navigate an evolving landscape supported by hardware, software, and infrastructure that was far from robust in the early days. It’s easy to forget how many of the iOS and Android features and services we all utilize today didn’t exist in the early days of smartphones. And hardware features like smartphone GPS and cameras were only just becoming available in the early 2010s. Plus, the lack of cloud computing, lightning fast internet speeds, and longer battery life hindered the capabilities of startups and their products. Of course, all of these things quickly came online over the years that followed, and teams raced to adopt them as a result. But it took time, grit, relentless adaptation, and product hacks for startups to break through, and all of the above contributed to teams needing to iterate their way to PMF.

Fast forward to 2024, and the infrastructure exists for new products to be quickly built and distributed to many millions of people. The challenge now, however, is that the market is extremely crowded and there’s already an existing market leader established for nearly every category. Just think about how many times you’ve seen a new mobile app launch and you’ve thought to yourself, “haven’t I seen a dozen different versions of this app?” But this ability to distribute your product to millions with the flip of a switch also partially explains why AI products can find Magic Market Fit today. Building products has become easy. But building magic? That’s special. And that’s exactly why customers are instantly drawn to AI products.

Tech iteration is the new product iteration.

Beyond this now-mature infrastructure which enables developers to quickly and cheaply build and deploy products, there are other major factors contributing to the rise of Magic Market Fit. First, the exponential growth in computing power, driven by advanced GPUs, has made it possible to train deep learning models at unprecedented scale, which would have seemed impossible only a few years ago. Plus, the availability of huge data troves from the internet, social media, and many of the products built during the mobile revolution, serve as the fuel needed for AI products to learn and perform. And lastly, large investments in AI research and talent from both the world’s biggest companies and startups alike has created huge demand for teams to translate their tech breakthroughs into market-ready products. All of these factors have converged to create a unique moment in time where magical technologies like ChatGPT, Midjourney, and GitHub Copilot can achieve Magic Market Fit once a new technology is unlocked.

With the search for Magic Market Fit driving roadmaps and hiring plans, perhaps it’s no surprise that startups look differently than they did only a few short years ago. Teams are forgoing hiring extra PMs and sales leaders in favor of doubling down on engineering and research, hoping to mine magic faster than their competitors. Even designers are becoming deprioritized in favor of “engineers who also know Figma.” But it’s hard to know if this trend will last. If history rhymes, and this AI market crowds just as quickly as the mobile market did, we may see things rotate back in the other direction, especially as engineering – and potentially AI research – becomes easier for less technical teams.

In the meantime, we should expect the pace of change – and investment – to accelerate towards a future where what really matters isn’t great product; it’s magic. 

Everybody hates ads. At least that’s what we all say out loud. But our revealed preferences tell a different story: we all actually love ads, because they fund most of the content we consume on the internet. Without ads, we’d have to pay for content with our hard earned dollars. But with ads, we pay with our attention, and we perceive the cost of the content as free. Ads have made the internet as we know it today possible.

Ads are also a really good business, because advertisers are willing to pay top dollar to access our attention. In the early days of the internet, this looked like simple banner ads brokered through manual deals on basic HTML websites. But over time, ads evolved into some of the greatest money printing machines ever invented, like Google and Meta. Even businesses we don’t often associate with ads, like Apple and Amazon, make a killing from them. And the little guys love them, too: many long tail publishers, websites, niche blogs, podcasters, and indie games are all funded by ads. Ads built the modern internet and funded our internet addiction, making billions of dollars and billions of us happy in the process. It's a perfectly tuned system that we all take for granted.

All good things must come to an end, and ads are no exception.

AI is increasingly challenging the business model of ads – the system that makes the open web feel free. For example: many of us are using Google search less frequently, along with the ad-supported websites that it leads us to. And in their place, we’re using answer engines like ChatGPT and Perplexity more. Each time we task these AIs with delivering us a single, definitive answer (in favor of wading through many potential answers via traditional search), the relationship between an advertiser and our attention is severed – and the ad never gets delivered.

Challenging the model even further, as AI models become more powerful and capable, they won’t just deliver us answers; they’ll also perform long, complex tasks on our behalf. They’ll do the busy work of research, making appointments, and even buying stuff for us. This will free us up to do other things, and as a result we’ll devote less of our attention to the content, and thus, the ads. At scale, this represents a massive change to the foundation of the open web. What happens to the perfectly tuned system when the funding dries up?

Premium content will become even more premium.

Tightly guarded walled gardens, especially built around premium content, will feel the least pain from this shift. Many will become even more valuable: as the overall amount of monetizable supply of ad supported content shrinks, the demand for higher quality content will rise. Ad dollars previously spent on commoditized content will be redirected to the best stuff on the internet. Plus, as we spend less time on busy work, we’ll spend more of our time entertaining ourselves, which means premium content will capture even more of our attention. And the best content will even find new ways to monetize: Reddit’s AI deal with Google serves as the perfect example of how valuable, niche content traditionally supported by advertising can be sold as training data for AI. And media formats that are more resilient to AI scraping, such as audio and video, will likely retain their advertising value longest. No matter how good an AI agent gets, you’re never going to ask it to watch your favorite Netflix show on your behalf.

But commoditized content, or anything that can be easily scraped by AI or replicated by competitors, will fare much, much worse. For this content to continue to exist, it’ll find new ways to be funded. For many of the sites we visit, this will increasingly be through user-funded subscriptions to offset the lost ad revenue. Those of us who can’t afford these will lose access to these services, widening the gulf between the haves and the have-nots. 

AI services will attempt to reinvent ads.

AI services will seek to capitalize on the end of traditional internet ads and invent new forms of advertising to capture the dollars that were previously delivered directly to publishers. Perplexity has recently revealed some early details of how ads will work on their platform, and OpenAI is rumored to be exploring this as well. While this is an obvious next step for the future of advertising, it’s unclear how directly the current ad spend will transfer over to this new model. And it’ll need to be structured such that publishers can participate in a way that incentives them to keep making content. Otherwise, they won’t be able to afford to exist.

Brand new models will also emerge to help publishers recoup lost ad revenues. As agentic traffic increasingly displaces human traffic, publishers will look to charge AI companies directly. Much of this will be through big, splashy content deals, such as OpenAI’s deals with News Corp and Axel Springer. But for smaller players that can’t afford to do deals themselves, they’ll work with third party services that will collect “tolls” on their behalf. Every time an AI hits their site to scrape content for their own service the AI will be blocked and forced to pay before the content is served. And as AI moves beyond content ingestion to agent-based task execution, APIs that help agents do their jobs will also be charged.

Content wants to be free.

For decades, the economic relationship between consumers of content and the providers of content has been hidden behind business models that were monetizing our attention. We didn’t put too much thought into it because we were getting what we wanted with minimal effort or cost. But as AI increasingly disrupts these models, the relationship between us and publishers will become much more transactional and direct. And we will feel it.

Over time, it’ll be a new internet, and the open web will be a memory. Great content will still find a way to reach us, just like it always has. But we’ll look back on the first few decades of the internet as the golden age of content, when everything felt free.

Thanks to Nir Zicherman, Sara Beykpour, Faraz Fatemi, Semil Shah, Lulu Cheng Meservey, Nikita Bier, and Kayvon Beykpour.

In a past life, I built consumer products. Building consumer products is all about “mining for flywheels” and extracting as much as possible from each well before it goes dry. Growth flywheels, talent flywheels, data flywheels, etc — all of these help your company grow bigger the faster they spin. But just like some things can make flywheels spin faster, other things can make flywheels spin slower. Adding unnecessary friction to your product onboarding flow can slow your product’s growth. Having an unappealing employer brand can slow your team growth. And so on.

Turns out, VC is a flywheel, too. And like consumer products, you want to get that flywheel spinning fast. When you invest in compelling companies, it makes it easier to meet other compelling founders, which in turn can lead to more compelling investments. When your investments return capital, you can recycle that capital into companies with bigger potential outcomes. When you surround yourself with other great partners, you increase the radius of potential investments (or potential hires). Conversely, having a bad “founder UX” when being pitched by startups can very quickly bring your VC flywheel to a screeching halt (founders never forget).

But the difference between VC flywheels and product flywheels is that in VC, the flywheel doesn’t always help you. In fact, it can sometimes hurt you. For example: just because you have access to more companies doesn’t mean you should invest in them all (you’d exhaust your bandwidth, let alone your capital base). Just because your flywheel is landing you invitations to speak at a bunch of conferences doesn’t mean you should waste your time flying all over the country (and not meeting founders). That’s because you, unlike a consumer product, are not software. You are not infinitely scalable. Instead, you need to get the flywheel spinning, but stick to your principles and the stuff that really matters for you. It’s kind of like staying up on a wave for as long as possible before toppling over near the shore. You want to specifically work on the things you need to improve in your form, not just try to copy Kelly Slater.

So get your VC flywheel spinning. But choose very carefully where and how to spend your time. Because just like a great product that’s shooting up the App Store charts, it’s hard to get the flywheel spinning again once it stops.

Thanks for reading! Feel free to follow me here on LinkedIn or X if you want to connect. And if you enjoyed reading this, I'd really appreciate it if you shared with your friends or teammates. 🙏

Something that happens often with startups is they choose to hire people who have really good resumes or experiences, even though they don’t have a specific job for them at any given moment.

Typically this happens because a VC or advisor wants to impress the startup, and they try to “show value” by connecting the company to someone with a compelling background that maps to the startup in some way. Then, the company falls in love with the idea and they bend over backwards to jam the person into the company. This can come in the form of excessive comp, inflated titles, or sacrifices on team values (such as IRL work vs. remote, etc).

But the worst part about it is that the person ends up doing a job that usually shouldn’t exist in the first place. This can have all sorts of negative downstream effects, such as: wasteful spend of capital / equity, constant misalignments between coworkers, and an inevitable firing that takes up mindshare and emotional capital of the founders. It’s a mess.

The cure to all of this? Just never hire opportunistically. Does it work out sometimes? Sure. And some teams are better at it than others. But it’s rare and much more commonly, a trap.

Instead, set a hard rule for your company: don’t hire anyone until something is actually broken. Then, define the exact specific role you need filled to fix the problem, along with requirements, level, comp, etc. And then go do the hard work of finding the exact perfect person for that role.

Hey everyone. It's been a while since I've written, mostly because I haven't had time to write the longer form essays that started this newsletter. But I've been missing writing. As a solution, I've started writing shorter-form "screenshot essays" on my X account, mostly focused on startup advice (based on my experiences as a startup founder and investor/advisor to other startups). These are much shorter posts, but I'm going to share a few of them here from time to time in case some of you find them interesting. Enjoy.

There’s a saying in startups: being early is the same as being wrong. If you’re right but you show up to the market early, it doesn’t matter – no one is ready to try your product. Shipping too late is just as bad, but for other reasons – especially if you’re doing so to overly prepare for GA in a closed beta.

First off, holding back on shipping to GA so you can optimize in a closed beta almost always leads to imperfect signals. Everyone in your closed beta is by definition biased. They proactively chose to join your small beta and jump through all the necessary hoops of signing up for it: filling out a Typeform, installing TestFlight, joining your Discord, and maybe even giving you feedback if they’re really pumped about what you’re building. They are extremely qualified, and this will show up in their engagement metrics and feedback. You will think your product is better than it actually is.

Second, because your closed beta is made up of highly qualified users, the feedback loop around user needs and product roadmap is heavily skewed. You end up building features that power users want – not the general population. The problem with this is that power users almost always want different things than the rest of the world, and you’re trying to get really big – not simply cater to a small, niche audience.

Third, and possibly worst of all: by the time you do finally get to GA, you’ve built a load of features specifically designed for power users – and now you have to support those features forever (or until you decide to pivot). You’ve signed yourself up for an eternal backlog of product and technical debt you must continue to pay off in perpetuity, thus slowing down your sprints on the stuff that actually matters. In other words, you are now simply a slow product team, maintaining features that can’t find and sustain PMF.

Conventional startup wisdom suggests you should perfect your product in beta before you go to GA. There’s a fear that you’ll “burn through your potential users” on an imperfect product. There are nearly 8 billion people in the world. You’ll be lucky if you can get a few thousand people to try your product in the first week, regardless of whether your product is great or terrible. I say skip right to GA, ship constantly (at least every week), and start learning on real-world data ASAP.

It is Summer 2023, and each day brings a new AI product demo that goes completely viral on Twitter X or TikTok. Countless people are blown away by the product’s magic-like qualities, powered by GPT4, Stable Diffusion, or some other new language, video, or image model. And while it’s an incredibly exciting time to be building or investing in AI, it’s also a fiercely competitive one as well, especially for teams building in the application layer – the part of the technology stack that delivers real world products to end users interacting directly with software.

The competition is being fueled by what’s at stake: participation in a generational platform shift in which the capabilities (and the potential value) of the tools we use are reaching new heights, all because of AI. We haven’t witnessed a shift of this magnitude since the advent of cloud computing, the mobile revolution, or even the internet. In other words: the stakes are high.

But it also has just as much to do with the fact that building and launching new AI products is arguably easier than ever before, democratized by more accessible coding education, powerful IDEs (aka integrated development environments, which help engineers be more efficient when coding, even aiding them via AI through offerings like GitHub’s Copilot), and the AI itself: companies like OpenAI and Stability, to their credit, have made it really easy for startups to make products using their game-changing tech.

AI is a Commodity 

The combination of the high stakes, the excitement, and the accessibility of the technology means that there are a lot of new startups out there building AI products. Seemingly every startup today can (and is!) incorporating AI into their products. But it’s not just the startups…bigger, more established players are also incorporating AI, too. And they’re moving really fast. As a result, AI has quickly become so ubiquitous, that it’s fair to say that it has become commoditized.

Historically, when new technologies have become commoditized, they have gone from being early differentiators for new entrants to becoming table stakes, and a requirement for most products and services to remain competitive. Take mobile apps as one example; shortly after the App Store launched, a handful of exciting and super innovative companies took the plunge and launched apps quickly and well before others. Some of these teams were rewarded handsomely; Instagram, WhatsApp, Uber, and others became big winners of the race to innovate on mobile before others did.

But as smartphone application development became cheaper, and the distribution for mobile became more ubiquitous (in the form of smartphone adoption), it was no longer a differentiator; it was a commodity. And not just for startups, but for incumbents, too.

Speed Matters, But Distribution Matters Most

Like startups, incumbents also want to win platform shifts, but in the early days, the advantage sits squarely on the side of the startups. Startups can see opportunities and act on them swiftly, like in the case of the examples of above. They get products out into the world fast, blitzscale, and win markets.

But as time goes on, the incumbents mobilize. While they may not ship as quickly as the startups, they can move with heft and might, deploying dozens, hundreds, or even thousands of engineers towards a common goal, often on a collision course with an entire category of startups. When this happens, incumbents hold a very valuable advantage, one that’s arguably much more valuable than the size of their teams or the quantum of their investments. That advantage is distribution. 

While startups search to find an audience for their new products, incumbents have already found one. While startups iteratively progress to earn the right to spend precious capital on marketing, incumbents already have entire marketing departments. And while startups fight and claw to unearth new hidden promotional channels, incumbents distribute within products already adopted by millions. 

All of this is to say that if a new startup is building an AI product that is likely to be offered by an incumbent, then the startup will inevitably face a very steep, uphill battle. After all, the incumbent can offer and aggressively market the same commoditized AI technology to an existing user base of highly qualified customers.

Just take a look at what Adobe is doing with Firefly, as one example. Earlier this year, a handful of super innovative startups were dazzling the world using open source libraries like Stable Diffusion to offer mind-blowing, AI-powered image generation tools. But in recent months, Adobe has moved decisively to offer similar capabilities directly inside of Adobe Photoshop, a product with vast distribution power and an ability to meet millions of creatives where they’re already doing work. And as I learned in a recent conversation with Adobe’s Chief Strategy Officer, Scott Belsky, the company has no plans to take their foot off the gas anytime soon.

How to Play a Different Game

Does this mean all hope is lost for startups? Of course not; after all, this same dynamic has played out repeatedly throughout history, and countless legendary startups have emerged, succeeded, and gone on to become generational companies. So then, what can AI-focused startups do? How can they gain an advantage for AI products in the application layer that are inevitably destined to go head to head with incumbents? Below are 3 examples of strategic tactics startups can take to fend off bigger companies’ home field advantage.

1. Unique Formats

Few tactics are as potent a weapon against incumbents as gaining adoption of a new, proprietary format. As I wrote about in The Standards Innovation Paradox, when new formats succeed, they provide startups with a huge competitive advantage. If a new team’s product outputs a unique, proprietary format that reaches scale, the startup becomes much more defensible than a product that operates with a standardized format (such as a standard image or video file). This is because the cost to others of adopting the new format for an existing product often requires reworking infrastructure, user experience, or even an entire business model. 

Snapchat

There’s perhaps no better example than Snapchat’s introduction of their signature “snap” format to illustrate the point. Prior to Snapchat, the most common form of sharing on platforms like Instagram and Facebook was through basic image files. These products (and many others) were perfectly designed to support the static, non-dynamic nature of a standard photo format. But Snapchat’s signature “snap” format offered a new way to share moments (in the form of photos or short videos) which would then disappear after a specific amount of time, rather than live on in perpetuity on recipients' devices. This unique format offered a fun and more spontaneous way of sharing that went far beyond a static photo that lived on forever. Users could create and share moments throughout their days without worrying about the permanence associated with legacy image formats. This made users feel less pressure to share more than on other platforms, which drove Snapchat’s engagement to the moon.

Snap then doubled down by introducing another unique format: stories. Launched in October of 2013, stories offered a new way to share multiple snaps together as a creative narrative, encouraging users to create even more content.

Eventually, after seeing the success of Snapchat’s unique formats, incumbents like Instagram and Facebook eventually raced to introduce their own versions of snaps and stories, but the challenger platform had already gained a substantial first-mover advantage, which helped propel it to the scale of a massive, publicly traded company worth tens of billions of dollars.

The strategy worked: while Snapchat didn't invent the concept of image or video sharing, the proprietary formats they brought to the world with the snap and stories formats revolutionized social media and made it hard for others to follow without massive investment. By focusing on light-weight, ephemeral, narrative storytelling, Snapchat found a new way to engage users, demonstrating how unique formats can challenge even the biggest players in the market.

AI-First Formats

Now, AI is making it possible to introduce brand new formats that didn’t previously exist. Take generative video avatars, as one example. Previously, to make compelling sales or training videos, users had to capture raw video footage of human beings speaking. As a result, video editing products have adopted a standard mode of editing and production, entrenched in decades of common user experiences. This standard UX is built to support the specific workflows of capturing video, importing it to a timeline, and enabling users to edit these videos. But now, through AI-first products like Synthesia, Veed, Tavus, and HeyGen, videos of people speaking can be generated dynamically. This not only eliminates the need to capture raw video footage from cameras (saving users time and effort), but it means that the entire user experience associated with editing these videos can be reimagined from the ground up to support the new capability more efficiently. In some ways, this new approach invalidates the legacy user experiences of classic video editors, and forces incumbents to deeply rethink their products and businesses to support the emerging avatar use case.

2. Value Destruction

When a new product's business model threatens to create value destruction for an incumbent, it becomes much more formidable. An example could be as simple as undercutting an incumbent's prices or offering a service that the incumbent can't replicate without severely wounding their existing business. And while this is a tried and true strategy for new entrants, it requires a delicate balance. Destroying others’ value can’t be the only tactic pursued; instead, startups must also create value for the customer in ways the incumbent cannot. Otherwise, the incumbent will also lower prices and leverage their massive scale to simply offer a better product.

Robinhood

One recent example of the business model destruction strategy is Robinhood, the company that offers commission-free trades of stocks, ETFs, and cryptocurrencies. Robinhood’s free trading, paired with a highly accessible, easy to use app (adding differentiated value beyond just undercutting prices), made it a no brainer for new traders to adopt. This was a major blow to incumbent brokerage firms, which typically charged fees for every single trade and catered to a more sophisticated type of trader. To compete, these traditional brokerages had to slash their fees, which represented a meaningful chunk of their revenues, effectively causing value destruction to their existing businesses. Many of the products in this market have since followed Robinhood’s lead, but the damage is already done; Robinhood is a publicly traded company worth more than $10B as of this writing.

Airbnb

And Airbnb created a platform that made it easy for people to rent out their homes or spare bedrooms to travelers, thus competing directly with hotels. This model was a game-changer in the hospitality industry and was a classic case of value destruction. Traditional hotels, bound by fixed costs and regulatory norms, found it difficult to compete without making substantial changes to their operational model. Value destruction…check.

AI-Inflicted Damage

AI makes it easier for certain types of startups to inflict swift and aggressive damage to incumbent business models. Take legal services, as one example. While there have been recent headlines around how AI-powered legal services have stumbled in the actual courts, it’s clear the technology has potential to be very disruptive for the category (or least leveraged as an efficiency driver). Think about it: Huge law firms charge exorbitant fees to fund (and profit from) the sheer human-power of their legal partners and associates. The skills of these highly educated lawyers is worth a lot, thus creating a large market for the best law firms in the world. But like highly skilled lawyers, new large language models have also proven to be effective at reading, analyzing, and even writing vast amounts of text. Startups like EvenUp are leveraging AI for specific legal services for a fraction of what they would typically cost. And law firms can’t simply turn around and use AI instead of people; this would make it impossible for them to justify the high costs of their lawyers, thus severely disrupting their business model. It’s a classic case of value destruction, all powered by AI.

3. Hidden Data Moats

In the world of AI, the concept of a "data moat" refers to the advantage a company gains through its access to (and usage of) high-quality, differentiated data. In essence, the more unique data an AI product can learn from, the more effective it becomes. For AI startups, developing a data moat involves amassing unique, valuable data that isn't easily accessible to others, and using it to train their own AI models. And once a startup has built a substantial data moat, it becomes extremely difficult for other companies (including incumbents) to catch up, unless they too can access the same volume or quality of unique data to train their AI systems. But once the moat forms and strengthens, it can be really hard for others to catch up.

So how can a startup that’s starting from zero build a data moat? There are a few potential paths, such as collecting data through their own unique services or doing partnerships with other companies that have unique datasets. Just keep in mind that incumbents can also do these things, so startups have to find a way to access a hidden moat not easily accessible by others. 

Palantir

There are a few classic, recent examples of data moats. Palantir, as one example, built an initial moat through its work with the United States government. The company’s early product was built for the intelligence community and focused on assisting in work on counterterrorism. This involved processing vast amounts of data from hard to reach sources, such as reports from agents in the field, intercepted transmissions, and private bank transactions. But over time, as the company expanded its customer base beyond the government to include financial institutions, healthcare providers, and other industries, Palantir continued to strengthen its moat by integrating and analyzing their diverse, massive datasets. By amassing such a vast and unique dataset, Palantir created a competitive moat that has made it hard for other companies to compete to this day.

TikTok

A well known, more recent example of a data moat is the TikTok algorithm. TikTok has amassed a treasure trove of highly entertaining, short form video content. But more than that, they’ve found a way to tune the TikTok user experience such that everyone using it is matched with highly personalized and relevant content each and every time they use the product, all through a unique data moat. The result is a platform that’s so effective, it has forced all of its competitors to change the way they distribute content, which I wrote about in The End of Social Media. So how do they do it? TikTok dissects users’ behavior upon each and every viewing of a video, including tracking their duration of consumption and analyzing interactions with the user interface. It’s even rumored that they monitor facial expressions of users as they watch videos through smartphones’ front-facing cameras. The result is a nearly impenetrable data moat that both gets stronger – while also improving the product experience – every time someone uses the product.

AI-Propelled Moats

Perhaps it’s obvious, but data moats are by definition, baked into AI products. Providers of large language models, such as OpenAI, leverage their own massive data moats to ensure their models are the best. However, there are ways for startups to build their own moats by offering AI in new and unique experiences, thus in turn generating a new data moat. For example, recent chatbot platforms like Circle Labs, Character.ai, and Replika are using AI as the foundation of their experiences; specifically, users of these products chat directly with characters that are powered by AI. As conversations with these AI-characters go on, the data powering the characters becomes better, thus making the conversations higher quality, driving even more conversation. It’s a classic flywheel of improving engagement while strengthening a data moat, all propelled by AI.

The Twist: AI startups are just startups

Startups which want to fend off looming incumbents’ distribution advantages must attempt to make their businesses as defensible as possible. This requires being nimble, tactical, and leveraging as many strategic advantages as possible. And while unique formats, value destruction, and data moats can all help, there are many other tactics to be pursued, as well. 

But here’s the thing about the above three tactics: they are not at all unique to AI startups. In fact, if you went back and re-read this entire essay but skipped all of the AI-specific sections and examples, the tactics would still hold true for all startups.

What does this all mean? It means that if you’re building an AI product – despite now being able to leverage an awesome, transformative technology – it’s really no different than building a non-AI product. AI can help your startup do magical things it simply couldn’t do previously, and the ways in which this is motivating teams to offer truly novel experiences is inspiring; however, in this context, AI is similar to the other tools your products can leverage, like cloud computing, mobile app development, live audio/video streaming, GPS, web3 . . . the list goes on and on.

At the end of the day, finding success for your product is not about leveraging the latest and greatest technology just because you can; it’s about building and shipping products that solve real problems for real people, and scaling those really f*cking fast once they find product market fit. After all, that’s what the incumbents did when they were startups, too. If your team can do that, you’ll walk away from the application layer battlefield victorious.

— 

Thanks for reading! If you enjoyed this, I hope you’ll consider sharing it with a friend.

Want to share your AI-first product with me? Shoot me an email.

Want to read more of my product strategy essays? Follow/subscribe on Substack, X, and LinkedIn.

Last week, the developer of Apollo (a popular third party Reddit client) announced that Reddit could charge the company $20M a year in API fees to continue accessing the platform's data.

In this new era of AI-first products, unique training data is the new gold. Data from user-generated platforms (like Twitter, which is also now charging for its API) holds exceptional value for AI training due to its sheer volume, diversity, and representation of human behaviors, emotions, and language use. The varied contexts and authentic interactions present in this data provide fertile ground for AI to learn, generalize, and predict effectively. Given that, it's no surprise that Reddit is now choosing to charge for it's API — an acknowledgement that they believe their data is now worth far more than they previously considered.

Think about it: with Reddit's vast assortment of user communities and topics, it offers a diversity of content that is particularly beneficial for training natural language processing (NLP) models. From light-hearted chats to intense debates, Reddit's rich text data encapsulates the complexity of human language.

Moreover, Reddit's unique structure enhances its value for AI training. Its upvote/downvote system is a goldmine for sentiment analysis, providing a ready-made mechanism for assessing user sentiment. In other words, it has reinforcement learning organically baked into its dataset.

The time-stamped nature of posts and comments also facilitates tracking trends over time, a feature that can be harnessed to anticipate future developments or comprehend the evolution of online discourse.

As we delve deeper into the AI age, the data from platforms like Reddit will continue to drive AI's growth, rendering it even more sophisticated and in sync with human intricacies. And so we should expect platforms to value it accordingly.