Is AI Breaking the Lean Startup Playbook?Patrick Collison × Harj Taggar · Guided reading notes by theme

Merged reading: in Harj's view, Patrick's most impressive achievement at the time was Croma, the Lisp dialect he designed. Compilers now do that work for us and we do not mourn it much, so perhaps in the same way we should not mourn the disappearance of source code; we only need to transcend the plane of writing instructions ourselves and hand it to Claude and the rest. Emotionally, though, Patrick misses it

Merged reading: how much should students in this AI world learn and derive from first principles themselves, and how much should they outsource to AI? Patrick's model for this is the computer cache, cache and not cash. Jeff Dean has a famous set of "numbers every programmer should know", including bandwidths, latencies and the constants you have to reason about when building systems. Obviously, when you think about building any system or distributed system, the various lookups between components and their bandwidths differ enormously. Retrieving from L1 cache is very different from retrieving from RAM, which is very different from retrieving across the network. He thinks knowledge works the same way. Of course you can have an agent compute or look something up for you, but querying an agent is much slower than having the knowledge in your own cognitive L1 cache; the number of round trips of thought you can complete in your head is far greater than what you can do by whispering a question into Superwhisper or typing it out. So even granting the models their full capabilities, he still thinks that for a long time to come, neuronal lookups will be much faster

Merged reading: if you look at the revealed preference of companies like Stripe and the AI labs, that is, at how they actually act, strong cognitive ability still commands an enormous market premium; giving up on training it before there is evidence that we have saturated its benefits would be premature

Merged reading: Patrick still writes himself; both philosophically and specifically, in terms of the substance, he dislikes the models' writing. It is interesting, because these models can do things like prove the Jacobian conjecture and are clearly capable of monumental feats, yet somehow he still has not read an LLM essay he found truly compelling. It may simply be that writing is very hard to optimize with reinforcement learning, because the utility function is hard to define. Still, he thinks writing and interpersonal communication are utterly fundamental; they require a person to reason sensibly in the multidimensional space of reality, and in a way that is hard to describe, he feels the models still fall short there. So he has never sent one of those pre-written suggestions: every tool now tries to offer them, whether Gmail or, reportedly, WhatsApp with a similar new feature, and he thinks he has still never once sent a suggested text in his life

Merged reading: back then, in what felt like the pre-Cambrian era of startups, startups were far less widely known than today, even on university campuses. Patrick felt a very strong sense of urgency at the time, but looking back, that urgency was a bit unnecessary. Marc Andreessen has also talked about a version of this urge to speed-run life

Merged reading: at the time Patrick thought many of the opportunities in startups, Silicon Valley and similar fields were ephemeral and fleeting; if they did not build it right away, three or four years later it might no longer be possible and every opportunity might be gone. In hindsight that was a poor intuition: over many decades it has been robustly and reliably the case that Silicon Valley has a surfeit of opportunities

Merged reading: humanity has always been drawn to millenarian models, the belief that everything as it exists will soon end and society will undergo a permanent, total transformation. There is a very good book called The Winged Gospel: after the airplane was invented, people believed civilization and even the human species were entering a new age in which everything would be completely different. Aviation was obviously a big deal, but it was not the sociological rewriting that the excitable proponents of the time imagined. So anything is hard to predict, especially the future, but on the claim that these are the last couple of years to create a company, Patrick would take the under

Merged reading: it is very easy to hallucinate a customer problem that is not actually felt viscerally by anyone who would pay money. In building Auctomatic together, they ran into a problem first-hand: moving money and taking payments on the internet was a real pain. On one hand it looked like an obviously good idea, because nobody liked the existing ways of doing it; they were universally disliked, antiquated and full of legacy baggage. You had to fill in paperwork, go to the bank in person, the documents might as well have been written in Latin, and everything about it was terrible

Merged reading: they felt like the proverbial squirrels in a trench coat, trying to pass as a real company or as serious adults while obviously knowing nothing about the space when they entered it. The banks and partners did not literally laugh them out of the room, but you could almost see them feeling under the table for the button to call security, because the whole thing looked so implausible. What really saved them was that what they were doing was grounded in a very concrete, actual, real user problem

Merged reading: from the first lines of code to the public launch took nearly two years; had they been going to YC meetings every week, they would have been knocked on the head and told to launch long before. What really saved them and kept the development from becoming a blind walk in the wilderness was that they had real production users almost from the very beginning. In January 2010, about two months into working, they got their first live production user. The product did very little then; it was larval and incomplete. All it could do was charge a credit card. After charging one, Ross asked a very reasonable question: "How do I see all my charges?" So they built a small dashboard. Then he said, "I want to refund a payment", so they added refund support. A few weeks later he asked, "Will I eventually get this money?", which was also reasonable, so they built payouts. It was very much just-in-time development driven by demand. In short, they had production customers very early, and although the product stayed in private beta, the number of customers grew every month right up to the public launch. Every week they had real customer feedback, feature requests and new users coming in; they were learning from reality rather than from their own hypothesized or extrapolated picture of the market. Patrick thinks that as long as you have a significant stream of grounding like that, it is probably fine not to have done the big, capital-L public launch yet

Merged reading: traditional lean startup says to buy traffic such as Google ads first, find a small gap in the market, then iterate and expand from there. Today's internet is far larger than it was twenty years ago when these ideas appeared. Perhaps it is more worthwhile to start from genuinely different points, entering territory nobody else is trying to occupy; in short, in the era of AI you may have to de-correlate your choices from everyone else's more aggressively

Merged reading: many of the successful companies of the past decade are very anti-lean startup, whether the AI labs themselves or Anduril; a lot of them share this characteristic. And now you can start much more aggressive and ambitious things from the outset

Merged reading: within YC, and probably the whole startup world by now, Patrick is famous for what Paul Graham called schlep blindness. At least on the surface, Stripe involves a great deal of tedious grunt work that is presumably not the most intellectually interesting. Patrick thinks PG did latch onto a real phenomenon, that there are all these menial tasks, but taking Stripe as a whole, he finds it fascinating

Merged reading: 25% of newly formed Delaware corporations are incorporated through Stripe Atlas, after which Stripe can keep working with these companies, hear their feedback and requests, and watch some of them grow into standout successes such as Shopify and OpenAI. Every business is a theory applied to reality, explaining how some aspect of the world or some market works; a new company with a new model is a contrarian thesis about some counterfactual

Merged reading: people worry that their startup idea will be steamrolled by the big AI labs. Would that outcome come from rapidly improving AI capabilities, or from the labs themselves deliberately entering the space

Merged reading: when they were doing Auctomatic, the question for their company and every other was always "What if Google does this?" Google at the time seemed almost omnipotent, with huge amounts of top talent and nearly unlimited access to capital, servers and every kind of resource. But human organizations are complicated; it is very hard for one company to aggressively prosecute a hundred different priorities while handling all the issues and interference that arise among them. Google has done incredibly well in many specific areas, but it has not done all the things, even if in some basic material sense it perhaps had the ability. So judging by the historical record, the results of big companies entering these spaces have been checkered

Merged reading: LLMs alone may make a set of specific verticals or tasks unnecessary, and agentic capabilities may directly eliminate the need for some products or tasks

Merged reading: back when Harj and Patrick were starting out, reaching one million dollars of annualized run-rate revenue quickly was a major achievement. In YC's day zero to day 90, growth is now mainly driven by enterprise customers being more willing to buy from startups, so a company can sign new contracts within the batch. Harj asks whether other factors are also steepening the growth curve, taking companies from 1 to 10 and from 10 to 100

Merged reading: businesses everywhere are spring-loaded, more ready than ever to adapt and try new things. They have a real terror of being left far behind with archaic, antiquated ways of operating. People know that maintaining the status quo carries a very high risk, so even though adopting something new is risky, staying on the old path looks quite dangerous too. For startups there has never been a better time to sell, and a product just out of the gate can be adopted by customers at a pretty meaningful scale

Merged reading: some fear AI will be a hegemonic, centralizing, totalizing force in which a small number of companies gobble up a very large share of the economy. Many of the companies at the AI frontier have already done extremely well and will surely keep doing so. But Patrick thinks there will be many thousands of winners; based on the trend lines visible today, we are heading toward a more decentralized world with more broadly shared prosperity