Toward the impossible, one day at a time.
← Journal

· 4 min read

It's Time to Work on Impossible Problems

요약 · Summary

The best high school mathematicians in the world need more than 100 steps of reasoning to crack the hardest olympiad problem. AI now thinks 1,000 steps deep. That changes which problems are worth our time.

Updated

It's Time to Work on Impossible Problems

In my last letter I asked who ends up paying for the AI buildout, so you may have read me as an AI skeptic. I'm the opposite. I'm very optimistic about AI. Where the money gets stuck is worth arguing about, but I don't have many doubts about what the technology can do. If anything, I keep coming back to one thought lately: it's time for us to go after the impossible.

Two videos helped me put this into words.

The first is an interview with Eric Schmidt on Blackstone's podcast. He ran Google for a decade and now runs the rocket company Relativity Space. He says that as you get older you should take on more risk, because you have less to lose, so you might as well try the things you thought were impossible. Coming from someone who is 71, that landed.

In his view, the biggest change in AI over the past year is long, deep reasoning. Models used to start thinking and then wander off, from math to French literature, as he puts it. Now they can stay on one problem for eight hours, with chains of thought that run 1,000 or 2,000 steps. "I can't think a thousand steps," he adds. Asked what young people should study, he says he used to say biology. Now he says deep reasoning.

The second is a conversation with Seo In-seok, a math professor at Seoul National University. He won a silver medal at the International Mathematical Olympiad (IMO) himself, and he helped review Gemini's IMO answers. By his count, the hardest of the six problems, Problem 6, takes more than 100 deductive steps, depending on how you split them. Only about ten of the best high school students in the world solve it in a given year. When AI first reached gold-medal level at the IMO in July 2025, Problem 6 was the one it couldn't solve.

The conversation is in Korean, with English captions.

Put the two side by side and the numbers jump out. There is a problem the smartest kids on the planet can only solve by pushing 100 steps deep in their heads, and AI now thinks 1,000 steps deep. The two men aren't counting "steps" the same way. Still, I think the direction is clear. AI can already think about ten times deeper than the deepest human thinkers, and within a few years that could be 100 or 1,000 times.

If that's true, the problems we're working on may be too small. Problems that fit inside human cognition will increasingly get solved by AI on its own. The human job is to pick up the problems that used to be close to impossible.

We're not there yet. Schmidt says real discovery hasn't happened. The test the industry uses is this: give a computer everything known up to 1902 and see whether it can come up with Einstein's relativity on its own. Today the answer is no. People in San Francisco think it will take two years; Schmidt thinks maybe within a decade. Professor Seo also says that problems humans can't solve because they need too much depth are probably still out of reach for AI. Problems that need more fields than one person can master in a lifetime might be a different story.

AI is changing one more thing. The moats companies spent years building are dissolving one by one. Accumulated technology and know-how don't protect a company the way they used to, and I think you can see it in valuations. So what moat is left? I think it's the belief itself that the impossible is possible. AI will think very deeply in whatever direction you point it, but a person still has to choose the direction.

Schmidt says something similar. If you're not technical, figure out how to use these tools to scale your dreams; the only limits now are your curiosity and your willingness to take risks. My version: draw a very concrete picture of the future you badly want, then whip AI toward it, and you can get close.

That's why I picked the eye. My left eye has both glaucoma and a cataract. A cataract can be fixed by replacing the clouded lens, but the optic nerve cells that glaucoma kills don't grow back. Medicine today mostly treats the eye as optics: swap the lens, lower the pressure, keep it from getting worse. Bringing cells back to restore sight biologically is still close to impossible.

So here's my picture. First, to use my own body to help bring forward the era when we can restore the body biologically, starting with sight. Beyond that, to freely upgrade what our bodies can do, until they're bodies that can one day go to space. I couldn't come close on my own, but in an era when AI thinks 1,000 or 10,000 steps deep, I believe that changes. So I'm starting with my eye, studying it, and writing it down here.

From here on, I want to run at the impossible with boundless optimism and not worry about how it looks to anyone else. If you aim at a future that feels 100 years away, you'll probably start seeing the first clues in three to five years. That seems worth it to me.

I'd love to know what future you want badly enough to draw.

Sources