Y Combinator president Garry Tan used an a16z interview to make a sharp point about startup decision-making: the most expensive mistake founders make is not a single bad product bet, but repeatedly chasing whatever appears hottest at the moment.
In the two-hour conversation with Anish Acharya, Tan moved from his early career and missed opportunities to the way AI is rewriting the startup playbook. The through line stayed the same. Founders, in his view, should spend less time asking what the market is excited about right now and more time asking what they actually understand, what they can see firsthand, and what they are willing to pursue for a long time.
Leaving the web and passing on Palantir
Tan said the biggest mistakes of his career came from following fashionable narratives instead of sticking with the areas he genuinely knew and cared about.
One example came right after he graduated from Stanford in 2003. He already had solid experience in web programming, but the Web 1.0 collapse and the Nasdaq drop convinced many people that the web was dead. Tan believed it too and shifted his attention to Windows Mobile. As he framed it in the interview, Mark Zuckerberg was likely building Facemash around that time and Facebook did not yet exist, while Tan was leaving web programming at exactly that moment. He described the move as going in the wrong direction at the right time.
The more expensive miss came with Palantir. Tan said his Stanford fraternity brothers Joe Lonsdale and Stephen Cohen were interns at Peter Thiel’s hedge fund when Palantir was founded. Thiel flew to Seattle, had dinner with him, and offered him a $70,000 check on the spot as his salary if he joined. Tan thanked him but declined, thinking he might make level 60 that year. Looking back, he estimated in the interview that turning down Palantir was roughly a $2 billion to $4 billion mistake.
What ties those calls together, he said, is that he was looking at the map instead of the terrain. He paid attention to what looked cool and what good investors would approve of, instead of opening his eyes to what the smartest people around him were actually doing.
Tan linked that problem to courage. The harder path, as he described it, is trusting direct experience rather than being pulled away by Twitter or an article in The Wall Street Journal. He also used the word earnestness. In his telling, it is not naïve. It is a form of maturity that lets someone stay grounded in what they have actually touched and tested, even when everyone around them is saying the opposite.
Why the fringe matters
Tan also argued that Silicon Valley’s most important breakthroughs usually begin in places that look strange, marginal or unserious at first.
He pointed to the old dream of putting a computer on every desk. At the time, that idea belonged to what he called fringe weirdo punks, including the Homebrew Computer Club crowd. Yet those people ended up defining the computing era that followed.
For Tan, strong ideas share another trait: they do not have obvious edges. The deeper someone goes, the more open-ended the intellectual frontier becomes. The opportunities that can be defined immediately and neatly often sit in places where competition is already dense.
That thinking also shapes how he described YC. He said YC effectively turned Silicon Valley from something people entered through relationships and insider circles into something accessible to anyone with a strong idea and the ability to execute. He called Startup School a “birthright for tech” and said 7,000 people come to San Francisco each year, most of them for the first time. Where they come from matters less than the quality of the idea and the ability to build.
He added that YC offers more than money and brand value. It gives founders a community where people can speak honestly, not just exchange performative praise at industry events.
One founder can now equal 400 people
On AI and startup leverage, Tan’s language was unusually direct. Through vibe coding and agentic coding, he said, one founder can now be the equivalent of 400 versions of themselves from two years ago. He did not present that as a metaphor. His point was that a single person can now ship work that used to require an entire engineering team.
From there, he turned to SaaS. Tan said the golden age of pure SaaS may be nearing its end. A pure per-seat SaaS business with no data moat and no network effects may stop making sense over the next five to 10 years, because AI is driving the cost of producing software close to zero. Software by itself no longer holds the moat. Data, user relationships and network effects do.
His advice was clear: if someone is still building a pure SaaS business in 2026, it should be a bridge to a deeper moat rather than the final destination.
That argument sits behind another line he used in the interview: code is no longer precious. When code was expensive, organizations built heavy processes around it — planning, reviews, specs and QA — to reduce the cost of being wrong. If code becomes cheap, the old assumptions weaken. What becomes scarce instead is taste and agency: deciding what deserves to be built and where effort should go.
“A markdown file is an employee”
Tan offered a practical picture of how YC uses AI agents internally for operational work. The process, as he described it, starts with doing a task once by hand and doing it well. Then the team turns it into a markdown file plus code plus tests, and runs it repeatedly as a cron job.
That is why he said, “a markdown file is an employee.” In his framing, this kind of employee does not make the same mistake twice. Once someone points out the error, the correction becomes a bug fix that stays.
He said this is really about turning every repeatable business process into a skill file. Engineering has skill files. Sales has skill files. Customer support has skill files. The first run will not be right, but feedback and iteration gradually sharpen the process. Over time, either people no longer need to do that work at all or they only step in to set direction and handle edge cases.
Tan also brought up “token maxing.” If someone wants to feel what 2028 is like today, he said, they can already do it by using tools such as Open Claw or Hermes Agent with a context window expanded to 800,000 to 1 million tokens, letting the agent reason with full context on every task.
He estimated that this currently costs about $50,000 to $100,000 per year. That is too expensive for most people, he said, but worth it for a CEO or founder because it amounts to living three years early at today’s prices.
The larger point is not just automation. As companies adopt more agents, they need a way to preserve what those systems learn instead of losing it at the end of each interaction. Encoding the workflow in markdown files and code gives agents persistent memory and reusable operational muscle.
The Brex example and the future of management
One of the strongest management examples in the discussion involved Brex co-founder Pedro. Tan said Pedro uses an agent to read the meeting notes of all his direct reports, which gives the system visibility two levels down in the organization.
That means he does not need to sit in on a compliance team meeting to know what the team discussed over the past three weeks, where disagreements sit, or which people are in friction with each other. According to Tan, Pedro can show up to a meeting with full context, say “you’re right, we’re going with your plan,” and leave.
Tan’s takeaway was that this gets at AI’s deepest management implication. Once a company grows beyond any single person’s cognitive boundary, problems start disappearing into the middle because no one has enough bandwidth to understand what is happening in every corner. He referenced the “7 plus or minus 2” idea from psychology as a rough description of working-memory limits. A person paired with a well-configured agent, he said, can hold the equivalent of three Harry Potter books’ worth of context in mind.
That is not just an efficiency gain. It is a change in cognitive range.
From there, he argued that middle management exists in large part because information moves through organizations with friction. If founders can directly perceive the real state of teams two levels below them through agents, then the rationale for layers of management has to be reconsidered. A large share of the coordination work that used to sit with middle managers, in his view, should move to agents instead.
AI may take 20 years to fully reshape institutions
Tan did not frame AI as an overnight rewrite of the world. He said its effects are likely to unfold much more slowly than many expect, and he called that a “white pill,” not a disappointment.
His reasoning was structural. The world’s infrastructure, companies, institutions and governments were all built around the assumption that humans can only hold “7 plus or minus 2” units of information at a time. That assumption has been true for thousands of years, so organizational charts and process design evolved around it. AI is now starting to break that assumption, but people are creatures of habit and institutions are inertial systems.
Tan’s estimate was that the transition will take about 20 years.
For startups, though, that slow rollout is the opportunity. He put it plainly: “an org like Microsoft can’t. But a startup can. And every startup must.”
He illustrated the point with a story from his time at Microsoft. To get another department to handle a P3 bug fix, he said he spent four hours and even ran to the other building with a baseball bat, only to find that the other team had not replied to emails, had not marked the issue as won’t fix, and had simply ignored it. For Tan, that kind of bureaucratic dead end is not about one bad actor. It is baked into how big organizations are structured.
So the transition may be slow, but that does not mean people can wait it out. The gap will keep widening between those who know how to use agents and those who do not, even if that gap does not become obvious within a single quarter.
Voice, memory and the 2027 “harness wars”
Tan’s view of the next computing form factor was one of the most forward-looking parts of the interview. In the near term, he said, interfaces may still look similar to today’s, but voice interaction is close to a sure direction. More important than voice is memory.
He said he wants future computers to be more than systems that answer questions. He wants something benevolent, something that knows a user’s hopes, fears and desires and keeps helping them move toward those goals.
Tan called 2027 the beginning of the “harness wars.” By that he meant the cost of running frontier-model intelligence may fall to a fraction of current levels within the next two to three years, setting up a real battle in consumer AI. The level of model capability people can access today for $50 to $100 may correspond to the top tier of intelligence right now. As costs drop, the decisive battleground could shift to user experience, memory design and multimodal interaction.
He compared it to a return of the browser wars.
In that framework, the key moat is not simply which model is smarter. It is which product can better integrate a user’s historical memory, behavior patterns and work context into daily interaction. Today, whether people use ChatGPT or Claude, most sessions still begin close to zero. The system usually does not know what decisions the user made last week, what long-term goals they have, or which communication styles they dislike. If that changes, AI assistants move from tools to partners, and switching costs rise with them.
Taste and agency become the scarce assets
The most persistent theme in Tan’s remarks was the growing importance of taste and agency.
AI has pushed the cost of making things toward zero, he argued, but it has not made it easier to know what should be made. If anything, it has made that harder. As resource constraints loosen, founders do not face fewer options. They face more, and many of those options appear equally buildable. Judgment becomes harder to substitute for.
Tan described himself as a late bloomer who made many mistakes, including leaving the web and passing on Palantir. The point of that history, he said, is not that mistakes automatically lead to success. It is that people can make mistakes, examine them honestly and choose differently later. The capabilities that matter in a fast-moving environment, in his telling, are introspection and iteration.
From chasing Windows Mobile in 2003, to missing Palantir, to eventually leading YC, Tan’s career arc in the interview reads like a record of bad timing, confusion between what is hot and what is right, and repeated course correction. In the AI era, he suggested, the people who benefit most may be the ones willing to admit that confusion early and keep acting anyway.
