I'm Thomas. I'm in SF, researching neural architectures and pretraining. I also think about startups. Sometimes I write about it.

Most of what I care about comes down to the suspicion that the ceiling is much higher than people think. Most of the field is chopping faster: scaling what works, tweaking it, and wrapping it. I'm interested in how urgency from AI introduces a bias to fall back on validated ideas, even if they contradict the first principles of a problem.

This shows up outside ML too. Cognitive biases can look a lot like local optima. The actions with the best long-run payoff are sometimes the ones that look wrong for the longest, because anything legibly good gets arbitraged away.

The only bets that still pay are the ones that look wrong for long enough that nobody else is willing to make them. That's most of what I'm interested in — in ML and otherwise.