Garry Tan on staying above the AGI line

As software becomes easier to build, the scarce skills move elsewhere: knowing what should exist, recognizing what works and understanding what people want.

A white bird in flight against a dark background textured with small, light-colored dots.

At the start of the year, inspired by engineer friends who said they no longer wrote code by hand, I began playing around with Claude Code. I hadn’t consistently programmed in over 10 years, but I was instantly hooked. Coding with AI felt like flying, or moving at the speed of thought. For weeks, I stayed up late building solo projects that would have required dozens of colleagues back when I was a full-time software engineer.

In a month working alone, I rebuilt Posterous, the startup that in the 2010s took me two years, over $4 million in venture capital and a team of engineers to build. Then I created GStack and GBrain, two open-source tools for building with agents. I was so amazed by what the agents could do that I cleared my schedule to spend every possible moment coding.

I had “AI psychosis” and it was incredible.

The most interesting thing wasn’t that AI made me faster. It was that every time the models got better, the scope of what I could attempt expanded. This made me think differently about an idea from the last technological era.

Around 10 years ago, during the software revolution, Peter Reinhardt coined the phrase “the API line” to describe an emerging distinction between jobs that involved directing software systems and jobs that involved being directed by them. As AI became more powerful, it was easy to imagine this distinction hardening into something permanent: a world divided between the people who controlled superintelligent machines and everyone else.

The Humanist Review of AI, Issues 01 and 02

Conversations at the frontier, in your inbox

Subscribe

But my experience with coding agents suggested something completely different. The more capable the agents became, the more I could build. Instead of a permanent class boundary, I began to think AI would create a new and constantly moving frontier: the AGI line.

Above the line are the things that remain scarce even when intelligence itself becomes abundant: deciding what should be done, recognizing what is good, understanding what people actually want. Essentially, judgment and taste.

The people who benefit most from AI will be the ones who continually move themselves above it.

The important thing about the AGI line is that it will keep moving upward. Tasks that require human intelligence today will increasingly be delegated to machines tomorrow. The people who benefit most from AI will be the ones who continually move themselves above it.

I see this now most clearly in the best founders building with AI. They are constantly testing how much more they can entrust to agents without lowering their standards. When a model improves, they don’t simply use it to do the same work a little faster. They ask: what has become possible that was previously impractical? I think of these people as AI maximalists.

Maximalists continually update their understanding of what the models can do. They have strong opinions about what the output should look like, but they are willing to change their approach quickly when reality disagrees with them. They can look at what an agent produces, recognize whether it is good enough and iterate until it is.

Maximalists understand that powerful new tools mean that they should become much more ambitious, not less. With the right model in the right hands, it is now possible to boil the oceans.

Their advantage is not that they trust AI blindly. It is almost the opposite. They have sufficiently strong judgment to delegate aggressively.

My friend Peter Steinberger is one of the clearest examples.

Before creating OpenClaw, Peter spent more than a decade building PSPDFKit, a software platform that was ultimately integrated into over a billion devices. In the process, he spent years making products, putting them in front of users, seeing which parts of them broke and improving them. That experience gave him something the models do not provide automatically: a highly developed sense of what a good product looks like.

With AI, Peter’s exceptional product taste is no longer limited by how many lines of code he can physically write in a day. In fact, he no longer writes any code by hand. And, rather than relying on manual PR reviews, he orchestrates dozens of Codex harnesses to continuously improve his code in a loop. This allows him to move faster than hundreds of engineers at a Fortune 500 company in a pre-AI world.

His role has moved upward. Instead of implementing every feature himself, he decides what should exist, evaluates whether what the agents produce is good and redirects them when it isn’t. The agents supply the execution, while Peter supplies the judgment.

This distinction matters because judgment and taste are not mystical qualities people are born with.

This distinction matters because judgment and taste are not mystical qualities people are born with. They are usually earned through contact with reality.

You build something. You show it to users. They misunderstand it. They refuse to pay for it. Something breaks. You change it. Eventually you begin to recognize patterns before other people do.

This is one reason why experienced founders and builders can become disproportionately powerful in an AI world. They have spent years accumulating judgment through thousands of decisions, and AI has suddenly removed many of the execution bottlenecks that once constrained what they could do.

At YC, we are already starting to see this shift.

Tiny teams can now build products that would have recently required much larger engineering organizations. Some of these products are reaching meaningful revenue before they have raised much capital at all. Their founders come from software engineering, design, product, research and specialized industries, but the pattern is similar: increasingly, implementation is not the thing slowing them down.

The hard part is knowing what to build. Is this a problem worth solving? Does the product work the way users expect? Is this feature important or distracting? Are we learning something from the market, or merely rationalizing a bad idea?

AI can generate thousands of possible answers. The founder still has to know which one is right.

AI can generate thousands of possible answers. The founder still has to know which one is right. This changes the economics of ambition.

Historically, many ideas were not so much impossible as uneconomical. A product might have required 50 engineers before anyone knew whether customers wanted it. A new market might have needed years of infrastructure work before a startup could begin experimenting. Entire categories of problems were left alone because the cost of trying was too high. AI lowers that cost.

This means tiny teams can attack markets that once belonged almost exclusively to large organizations. You no longer need another giant company to challenge a slow or unsatisfactory incumbent. In more and more cases, a handful of people with good judgment and an army of agents may be enough to get started. And the effect will not be limited to software.

The internet created extraordinary abundance in the world of bits. AI has the potential to help extend that abundance to the physical world because many physical-world companies entail enormous amounts of cognitive work: engineering, simulation, design, procurement, logistics, software, documentation, analysis.

As the cost of that work falls, founders can attempt problems that previously required much larger organizations. YC founders are building aircraft, factories, robots, exoskeletons, drones, energy systems and other products that would once have looked impossibly capital- and labor-intensive for a startup.

AI does not make these things free. You still have to manufacture the aircraft and build the factory. But it changes how much a small number of people can coordinate, design and understand.

This applies inside existing companies too. A business can use AI to accomplish the same goals with fewer people. But it can also decide that the old goals were small because execution was expensive. It can now ship 10 products instead of two. Enter markets it previously ignored. Give customers things that were once too costly to build. Attempt projects that management would have rejected five years ago because they required too many engineers.

The interesting question, then, is not simply whether AI reduces the amount of labor required for a given task. Of course it does. It’s what people choose to do with this abundance of intelligence.

The history of technology is largely a history of people moving up levels of abstraction. Programmers stopped writing machine code. Designers stopped typesetting pages by hand. Founders stopped running their own servers. Each time, work disappeared below the line. And each time, people found larger things to build above it.

AI will accelerate this process dramatically. The AGI line may move faster than any previous technological boundary. Things that feel difficult today may seem trivial a year from now. Anyone who wants to remain above the AGI line should internalize the lesson that the world is not zero-sum. Human desires are infinite and ever-expanding. There will always be work to do and problems to solve.

As AI capabilities continue to improve, humans will move further up levels of abstraction, liberated from ever more of the execution that once consumed their time. What remains scarce will be the things that have always mattered most: judgment, taste and the ability to recognize what is worth doing. Choosing the right problems and making something people want.

Garry Tan

Garry Tan is President and CEO of Y Combinator. He cofounded the venture capital firm Initialized Capital and the blogging platform Posterous, which was acquired by Twitter. Previously, he was an early designer and engineering manager at Palantir. He is also the founder of Garry’s List, a political news, commentary and online community organization. He holds a degree in computer systems engineering from Stanford University.

English (United States)
Your Privacy Choices Opt-Out Icon Your Privacy Choices
Consumer Health Privacy Sitemap Contact Microsoft Privacy Manage cookies Terms of use Trademarks Safety & eco Recycling About our ads