Research Scientist

tl;dr: Research Scientist, Berkeley, $230k–930k/year

In early September, AI solved the Navier–Stokes equation, one of the seven Millennium Prize Problems. The proof itself ran to 165 pages, and its verification would have taken years. But mathematicians got lucky, they have Lean. But how can humans verify the results of very smart models — now, and especially in the future?

For the record, humanity doesn't have good answers to this question yet! But there are some — for example, summarization. If we can't handle a 165-page proof, let the model summarize it in chunks, then summarize the summaries — and at the n-th step, you're left with text that a human can check themselves. Resolution founder Jeffrey Irving developed another approach: debates. Two models argue for opposite answers and look for holes in each other's arguments, and a human acts as a judge to decide whose argument prevailed 🏬

The field of science that comes up with all this is called scalable oversight, and it's really cool to work in — the experiments are understandable, technically challenging to implement, and the answers to the questions you face are desperately needed by humanity today.

Scalable oversight is one of several research agendas at Resolution, and they are very actively looking for researchers for it. They are looking a bit less actively for other agendas, but still — so if theoretical computer science, philosophy, statistical physics, or something else is more your speed, write to us, and we'll tell you all about it.

Experience in safety will be very useful, but, in general, not mandatory — they are also looking for strong scientists from other fields 🤓

Approximate list of what you'll need: 🟡PhD or equivalent research experience: in ML, CS, mathematics, physics, statistics, philosophy, or something close; ⚪️papers, preprints, or comparable results + experience in clearly describing them; 🔴strong mathematics; 🟡you are prepared to use AI extensively but carefully in your work; ⚪️you are genuinely motivated by alignment and want to contribute to it personally.

Specifically for scalable oversight, the following will be a plus: experience with debates, prover-verifier games, or large-scale frontier RLHF, a background in game theory, complexity theory, mechanism design, or decision theory.

Work is in the office in Berkeley; visa and relocation assistance are provided. Pay is $230k–930k per year.