YC CEO on AI, Startups, and the Risks of Chasing Trends

Source: DeepThought Circle

Have you ever seriously considered that those who made the worst decisions at the worst times often end up becoming the wisest? Recently, I watched an hour-long conversation between Garry Tan, CEO of Y Combinator, and Anish Acharya from a16z, where they discussed his career beginnings, the opportunities he missed, and how AI is rewriting the rules of entrepreneurship. After listening, I reflected on it for a long time because much of what he said aligned closely with trends I’ve observed over the past few years—but he articulated them with far greater clarity and insight. I believe this content is well worth sharing seriously.

Garry Tan said his biggest career mistake wasn’t a poor product decision, but repeatedly chasing what seemed “hot” at the time, rather than sticking to the areas he truly understood and was genuinely passionate about. When he graduated from Stanford in 2003, he had solid experience in web programming, but because the web 1.0 bubble had just burst and Nasdaq had crashed, with everyone declaring the web dead, he believed them and switched to pursuing Windows Mobile. He said that at the time, Mark Zuckerberg was probably building Facemash and Facebook hadn’t been created yet—and right then, he walked away from web programming. “It’s ironic: I got the timing right but took the wrong path.”

There was an even more expensive one. His Stanford fraternity brothers Joe Lonsdale and Stephen Cohen were interns at Peter Thiel’s hedge fund at the time. When Palantir was founded, Peter Thiel flew all the way to Seattle to have dinner with him and handed him a $70,000 check right then and there, saying, “This is your salary now—come join us.” Garry Tan thanked him but declined, because he thought he might get promoted to level 60 that year. In an interview, he calculated that this rejection now represents a mistake worth $2 billion to $4 billion. But he said the root cause of both mistakes was identical: he was looking at the map, not the terrain itself. He cared about what seemed cool or would impress top investors, rather than simply observing what the smartest people around him were actually doing.

I think it’s easy to say, but extremely hard to actually do. Because chasing trends gives you immediate, positive feedback—people around you call you smart, and the media tells you this is the next big opportunity. But what Garry Tan described as the harder thing is trusting your own direct experience, rather than being swayed by a tweet or an article in the Wall Street Journal. He used the word “courage,” and I think it’s an incredibly accurate choice. Earnestness, in his view, isn’t naivety—it’s a form of courage that only comes with maturity. You need to be清醒 enough to believe in what you’ve personally touched and experienced, even when everyone around you tells you otherwise.

Garry Tan shared a perspective that deeply resonated with me. He said that over the past few decades, the truly great innovations from Silicon Valley almost always began as something strange, marginal, and non-mainstream at the time. The idea that “every household would have a computer” was seen as the outright fantasy of a weirdo; the members of the Homebrew Computer Club were viewed by outsiders as a bunch of fringe weirdo punks. Yet it was precisely this group that defined the next several decades of the computer era.

He also said something that deeply resonated with me: a hallmark of a good idea is that you can’t find its boundaries. The deeper you go, the more you realize the intellectual horizon of the topic is open-ended and infinite—not something clearly defined or countable. From my own observations, this holds true: truly promising ideas are often hard to define in their early stages because they’re inherently creating a new category. In contrast, opportunities that can be clearly defined at first glance are usually areas where competition is already well established.

Garry Tan noted that what YC does is essentially transform Silicon Valley—from a place accessible only through connections and networks—into a space where anyone with a good idea and the ability to execute can enter. He referred to YC’s Startup School as the “birthright for tech,” where seven thousand people flood into San Francisco each year, most for the first time; where your origin doesn’t matter, only your idea and execution do. I believe this positioning is correct, because what YC truly gives entrepreneurs, beyond funding and brand endorsement, is a community where you can speak honestly—not the performative networking of TechCrunch events filled with empty praise like “killing it,” but a real network of people you can call when things are hardest.

This is the most concrete and most worth reconsidering part of the entire conversation. Garry Tan says that today, through vibe coding and agentic coding, a founder can be the equivalent of 400 versions of themselves from two years ago. This isn’t hyperbole—it’s literal. You can now accomplish alone what previously required an entire engineering team.

He spoke plainly: the golden age of pure SaaS (software as a service) is coming to an end. He analyzed that a pure per-seat SaaS model with no data moat and no network effects may no longer be viable within five to ten years. Because AI is driving the cost of creating software close to zero—software itself is no longer a moat; data, user relationships, and network effects are. He said that if you’re building a pure SaaS business in 2026, you’d better ensure it’s a stepping stone to a deeper moat, not the final destination.

Then he said something I found particularly important: "Code is no longer precious." I understand the meaning behind this statement—previously, writing code was expensive, so you had to think carefully before starting: you needed proposals, reviews, specifications, and QA, because every step aimed to reduce the cost of making mistakes. Now, writing code is almost free, so the foundational assumption behind this entire process has fundamentally changed. What has truly become expensive is taste and agency—the ability to know what to do and where to invest your effort. This has become harder and more critical than ever.

This is a metaphor Garry Tan used in an interview, and I think it’s incredibly precise. He said that within YC, their approach to using AI agents for business processes is this: first, do something once, perfectly; then turn it into a Markdown file with code and tests that can be repeatedly called and scheduled to run automatically via a cron job. He said: “A Markdown file is an employee”—one that will never make the same mistake twice. You tell it where it went wrong, and it turns that into a bug fix that stays permanently.

The workflow he describes essentially turns every repeatable business process into a skill file. Engineering has its own skill file, sales has its own, and customer service has its own. It won’t be perfect the first time—you provide feedback, it adjusts, and gradually the skill becomes more accurate. After that, human intervention is no longer needed for the task itself; people only need to set the direction and handle edge cases.

He also mentioned the concept of "token maxing"—maximizing token usage. He said that if you truly want to experience what 2028 will be like, you can do so today by using tools like Open Claw (the open-source version of Claude) or Hermes Agent to extend the context window to 800,000 to 1,000,000 tokens, enabling the agent to consider the full context with every task. He noted that this currently costs approximately $50,000 to $100,000 per year, which is too expensive for most people, but for a CEO or founder, this investment is worthwhile because you’re essentially living in a world three years ahead at today’s cost.

I fully agree with this logic. The Skill file framework addresses a very real problem: as your AI agents grow in number, how do you ensure that the "knowledge" and "processes" they accumulate don’t disappear after each conversation ends? By固化 them into Markdown files and code, you’re essentially giving agents long-term memory and transferable muscle memory—just like training a real employee, except this employee’s marginal cost approaches zero.

One segment in the interview left a deep impression on me: Garry Tan described how Pedro, Brex’s co-founder, uses an AI agent to manage the company. Pedro has the agent read through meeting notes from all of his direct reports, and the agent can access information down two levels below. This means he doesn’t need to attend meetings of, say, the compliance team himself—the agent summarizes what the team has discussed over the past three weeks, identifies points of disagreement, and highlights interpersonal tensions. He can then suddenly appear at a meeting with full context and say, “You’re right, let’s go with your plan,” before leaving.

Garry Tan says this is the deepest management problem AI solves: when a company grows beyond the cognitive limits of a single person, many issues begin to disappear in the middle layers because no one has enough bandwidth to truly understand what’s happening in every corner. Psychology has a concept called “seven plus or minus two,” meaning our working memory capacity is roughly around seven units. But a person augmented with a well-configured agent can hold in their mind the equivalent of three Harry Potter books’ worth of context. This isn’t just an improvement in efficiency—it’s an expansion of cognitive paradigm.

I think this example reveals something more fundamental. Traditional management structures exist largely because information flows with friction within organizations, and everyone can only process a limited amount of information, necessitating multiple layers of reporting and coordination. But if a founder can directly perceive the true state of the organization two levels below through agents, then the very rationale for management must be redefined. Garry Tan put it bluntly: the organizational coordination tasks previously handled by middle managers should, in essence, be performed by agents—not people.

This is the most clear-headed and最容易被误解的段落 in the interview. Garry Tan says he believes AI’s impact will be much slower than most people expect, but he calls this judgment a “white pill”—a hopeful remedy—rather than a disappointment.

His logic is this: The entire world’s infrastructure—all companies, institutions, and governments—has been built on the assumption that humans can only hold seven plus or minus two items in their minds at once. This assumption has been true for thousands of years, so all organizational structures and process designs have essentially been created to accommodate human cognitive limitations. But now, this assumption is being overturned by AI. The problem is, this shift won’t happen overnight, because humans are creatures of habit and institutions are inert systems. He estimates this transition will take about 20 years.

But he said these 20 years aren’t a bad thing—for entrepreneurs, this is an opportunity window. Because large companies can’t change, or change extremely slowly. Garry Tan said something I found incredibly powerful: "An org like Microsoft can't. But a startup can. And every startup must." While working at Microsoft, he spent four hours trying to coordinate a P3-level bug fix with another department, even going over to the opposite building with a baseball bat—yet the other team didn’t reply to emails or mark it as "won't fix"; they simply ignored it. This kind of bureaucratic nightmare is almost structural in large companies—not the fault of any individual, but an inevitable result of the organization’s design.

My own feeling is that Garry Tan’s 20-year prediction serves both as a comfort to those who are overly anxious and as a warning to those who think they can simply wait passively for the next big trend. It’s slow, but it will come—and before it arrives, it will continue to create uneven opportunities: the gap between those who understand how to use agents and those who don’t will grow wider, though this gap won’t be clearly visible within a single quarter.

I think Garry Tan’s perspective on the next generation of computing form is the most forward-looking part of the entire conversation. He says that in the short term, interfaces may remain largely unchanged, but voice interaction is almost certainly the direction forward. More importantly, he emphasizes the dimension of memory. He envisions future computers not merely as tools that respond to questions, but as benevolent entities that understand your hopes, fears, and desires—and continuously help you achieve them.

He refers to 2027 as the beginning of the "harness wars"—a term meaning that the computational cost of today’s most advanced models will drop to a fraction of their current price within two to three years, triggering a true battleground for consumer-grade AI products. The capabilities you can access today for $50 to $100 may soon match the intelligence level of today’s most advanced models. At that point, user experience, memory design, and multimodal interaction methods will become the true differentiators. He likens this to the return of the browser wars.

My own view is that the core of this harness wars isn’t just about whose model is smarter, but whose system better integrates users’ historical memory, behavioral preferences, and work context into everyday interactions. Right now, when people use ChatGPT or Claude, each conversation essentially starts from scratch—the model doesn’t know what decisions you made last week, what your long-term goals are, or which communication styles you dislike. Once this challenge is truly solved, AI assistants will shift from being mere “tools” to becoming true “partners,” and user switching costs will skyrocket. That’s the real moat.

After listening to the entire interview, the two words that kept coming back to me from Garry Tan’s repeated emphasis were “taste” and “agency.” He said these two things have become off-the-charts important, but many people still haven’t realized it.

My understanding is that the emergence of AI has lowered the barrier to "creating something" to nearly zero, but it hasn’t made it easier to "know what to create"—in fact, it has made it harder, because the number of possibilities has multiplied. Previously, resource constraints automatically filtered out most options. Now that those constraints are gone, you’re faced with countless nearly equally "viable" directions, and what’s truly scarce is judgment.

Garry Tan says he is a late bloomer, made many mistakes, turned down Palantir, and left the web—but he argues that his life proves you can make mistakes, reflect deeply, and then make different choices. I find this more authentic than many motivational speeches, because it doesn’t claim you’ll inevitably succeed; it only emphasizes one thing: the ability to introspect and iterate is what truly matters in a rapidly changing environment.

From chasing Windows Mobile in 2003 to missing Palantir, to later leading YC, Garry Tan’s journey has, in some ways, been a microcosm of Silicon Valley over the past two decades: filled with misjudgments of timing, confusion between what’s “hot” and what’s “right,” but also filled with those willing to admit mistakes, adapt, and persist even when everyone else said no. I believe this trait is more important now than ever in the AI era, because never before has any era given people like this such immense leverage.