There's Only A Few Problems Left
and they're a lot harder than they used to be. why I think it's a particularly bad time to start a startup
23.09.26
Robotics, biology, healthcare, brain-computer interfaces, manufacturing, space, and energy. That's seven, there's maybe some more I don't know about. Probably no more than ten.
As an ambitious young person working on technology these days, it feels like you kind of have to pick one of the above (I'm personally on bio and health!). To be clear, there's tons more billion-dollar and million-dollar problems around, but there's not that many trillion-dollar problems, and especially not that many that are aesthetically appealing or conducive to a bright future (in contrast to e.g. video games, trading, porn etc).
You may ask why we're not interested in solving billion-dollar problems anymore, and I think it's partly because of how much higher the stakes are now. In 2012, the main character startup was Uber - all the talented people were there, the news was talking about it, it was the place to be in Silicon Valley. It was, if we're being honest, just a ride hailing app. Nevertheless, it changed almost everyone's lives a little bit. There are plenty of billion-dollar startups being started now - AI for lawyers, AI for doctors, AI for restaurants and hairdressers. All of these problems are undoubtedly valuable and worthwhile pursuits that will make a bunch of money, but they're not going to feel like they change the world for most people - they might just change the lives of hairdressers.
Now the main character startups are the frontier labs, where I've been lucky to spend some time. However, I joined OpenAI in 2024, when I was 19. I couldn't have practically joined any earlier, and yet I felt even then that I was a bit late. When I talk to younger friends, they all feel a bit late. Not that we can't have a big impact in AI - many of us have and will - but that the key parts of this story, the key players, have been decided before our time. This isn't our story, we're all employee number 500 or 1,000 or 5,000. Fundamental AI research startups ("neolabs") face a similar problem: there will surely be more to do on the way to abundant intelligence, but it doesn't feel like that much more. AGI will happen with or without us.
One thing tying together all these big problems is that they feel like they don't quite take care of themselves in a world with abundant intelligence. Many of these problems involve the physical world and long timescales. Many of them involve somehow distributing AGI; robotics give it a body, bio and health give it the tools to work on human disease, neurotech gives it an interface to minds. Almost all of these problems were too ambitious to even dream about for our forebears 10 years ago, so in a sense I'm grateful to be able to even consider them. In fact, one genuinely good side effect of recent advances in coding is that most of my friends in college no longer see computer science as a safe or "default" option with high expected upside, and so feel happy taking the opportunity cost and studying mech eng, neuroscience and so on.
For a long time the received wisdom in Silicon Valley was that you had to be technical. It didn't matter if you didn't have an MBA or any experience running a company or what a board meeting was; you could just figure that out, but you had to be good at mathematics, physics, computer science or a related field. You had to be an engineer. Machine learning research, for example, is a quintessential technical problem in computer science - you need the smartest people working on it, doing genuine research, which happens to take the shape of writing code on a laptop. I still want to be technical, and I appreciate other technical people. I have trouble squaring this belief with the fact that technical intelligence will soon be abundant. In fact, almost all of the big problems I mentioned are not actually bottlenecked by technical problems! In robotics, everyone wants to be working on giant vision-language-action models, but most agree that the problem is orchestrating a complex supply chain to scale from heavy-metals to actuators to produce the robots themselves cheaply. Healthcare is clearly not a technical problem at all, rather it is full of purely operational questions about margins, reimbursement and care. Biology research is technical, but once AI is able to do it (soon!), the bottleneck becomes the operational complexity of running clinical trials.
Weirdly, then, as a generation of technologists we find ourselves signing up to tackle a ton of execution problems - in Tyler Cowen's words, humans will be the bottleneck, and so our job will not be to solve hard technical problems or to do difficult research, but to go after all these humans! Email them to make sure they're doing their job fast enough, hire some more if not. Many of these chunks of the economy have not been looked at seriously in decades, since they haven't been under this much stress - the energy grid did not put on power in the last few decades since it didn't seriously need to; manufacturing did not advance because it was offshored; healthcare didn't advance because of raw systemic inertia. I think they'll benefit from some fresh eyes.
These problems are also all gigantic. They'll probably be able to host several giant companies, and it's not clear if it's even possible to "win" something like space; they will all require new infrastructure and products, and shit tons of capital to get it all built. Most of all, though, they require actual domain expertise. You can complete the chain of logic required to start Uber pretty quickly; everybody's got phones, everybody's got cars, let the people with cars come to the people with phones. Again, ten years ago we were still in the throes of the app boom, where it felt like you could just have an idea and maybe be right. By contrast, you can't just come up with the correct approach to fusion, or humanoids, or colorectal cancer. You have to actually be an expert. Certainly you shouldn't like, freak out and get a PhD because of this - you can now just talk to Claude for a couple weeks and you can become knowledgeable pretty fast in anything - but you can't bluff it either. Reality is the ultimate arbiter and even if your intuitions aren't as developed as the older people, if you don't at least demonstrate that you know your shit, you won't be able to convince them to join your side. These kind of problems therefore lend themselves to a different kind of startup; a startup that starts with an insane quantity of capital, and hires the experts, and puts them to work on a particular bet on a new technology. If the experts don't believe in the bet, they won't come, and they actually have the leverage in this interaction, not the founder. Because of this, the bar to start these startups is so much higher than it used to be. 2022 or 2023 may have been the best time to start a startup; there was so much green potential everywhere even in the industries we young people knew about. In 2026, we're going to have to look further afield to have an impact, and that's going to be scarier and harder.
I wrote about this before, but I think this actually all results in a relatively simple procedure for young people right now. Pick one of these problems - whatever appeals to you the most, whatever you feel most excited by - and start getting to know the crew of people who are going to join you on this journey over the next few decades. It wouldn't hurt to join a startup in the field. Try and plan out what might happen to the industry over the coming years, and what you can do about it. Finally, once you have the team and the plan, you can start working together.