Wilhelm Hedenskog in a red Mario cap, sticking his tongue out

Wilhelm Hedenskog

I'm 19 and one of three founders of Petrichor, in Y Combinator's F26 batch in San Francisco. We build the infrastructure that lets AI do wet lab work and learn from it, so the best models can run experiments and discover things faster.

This is what I believe about the next few years.

01

Most people have no idea what the best models can do.

Most people I talk to have only used the free tiers of ChatGPT or Claude, and they judge AI by that. The best models perform at the level of the smartest people alive in nearly every cognitive field, which makes them a different technology from the one most people have tried.

human ability what most people think AI is the best modelstoday

02

Progress is accelerating.

Models now write much of the code and run many of the experiments that go into training the next models, so each generation arrives sooner than the last. A plan built around what models can do today is out of date by the time it's carried out.

time model generations

03

Ambition arbitrage.

For the next one to two years, almost anyone can build things nobody thought were possible, because the tools arrived faster than people noticed them. Most people are still planning their careers by the old rules. That gap won't last, because people are catching on too. Their curve is exponential as well, and nobody knows how fast it bends.

What's possible now What people attempt how fast? nobody knows time

04

Don't train to do what a machine will do better.

Spending years learning to solve hard technical problems by hand is a bad bet now. The more your goals overlap with what models will do, the more likely you are to be replaced, so aim for the layer above them, where you decide what gets solved and direct the models solving it. In 1820s Britain, handloom weavers kept training for work that power looms did faster and cheaper, and within a generation their wages had collapsed.

what models will do your goals

05

Up your ambition.

A lot of people feel threatened because models look like they'll take their jobs. Models don't get tired or bored, so they will take the spreadsheet work, and the average person gets to spend their days on work that matters to other people. As models push down the cost of making things, the cost of living falls with it, including healthy food, which can end up cheaper than fast food in a country where 40% of adults are obese.

cost of living someone's week tedious work meaningful work

06

AI will cure most major diseases within 20 years.

That includes cancer and autoimmune diseases, and it means better vaccines and better antibiotics. Drug resistant infections directly killed about 1.27 million people in 2019, and for decades bacteria have evolved resistance faster than we've found new drugs. Models will design new antibiotics faster than resistance spreads, and we'll win that arms race for good.

time resistance spreading new antibiotics

07

No one should ever have sole access to the best models.

No single person, company or government should ever control the best models alone. In the 2019 documentary iHuman, Ilya Sutskever warned that AI could make dictatorships infinitely stable, so that nobody could ever escape them. I think that's the most important risk there is, and the defense is keeping access to the best models spread out.

Elsewhere

Email me. I read everything.