Dr. Richard Moulange

speaker

1,784 appearances 1 recordings 1 series first heard Mar 2026 last heard 31 Mar

Dr. Richard Moulange’s voice in public audio — every appearance, attributed to the second.

Trend

recordings per month · last 12 months

Recordings per month over the last 12 months — 1 in all, peaking in Mar 2026 with 1.

Appearances

newest first · ▶ plays the moment

80,000 Hours Podcast · #241 – Richard Moulange on how now AI codes viable genomes from scratch and outperforms virologists at lab work — what could go wrong? · 31 Mar 2026 podcast

But then Cloud4 Opus started doing a bit better, quite a bit better, but the sample size is always small.

It's really hard with up-list studies.

There are only so many PhDs to go around.

Your sample size is quite small.

It's quite hard to estimate and get those error bounds down enough to find a statistically significant result.

And then with Opus 4.5 that came out just at the end of 2025, yeah.

Yeah.

It was strongly that the AI-assisted PhD students were much better, so much so that they almost breached anthropic's predetermined critical threshold, the sort of under which they would say, wow, this is an awful lot of uplift.

You can go and look at the graph and the X that marks the AI-assisted performance is just above the red dotted line that meant danger here.

And so this says two things.

One, we haven't really found evidence for novices, though we've been looking for it.

And this connects a lot to safety frameworks that we can talk about in a moment.

But two, we are in fact finding evidence about uplift for mid-tier actors, highly skilled individuals, PhD students.

There is less public evidence on assessing uplift for those sort of somewhat and moderately capable groups.

Because the sorts of groups we're talking about, you know, operate overseas.

And while I hope and I believe they are surveilled and monitored, that's not necessarily public information.

And those sorts of people wouldn't necessarily be amenable to me going, hey, would you like to be in a randomized controlled trial to see whether you'd be better at your terrorism or your WMD program?

But nevertheless, yes, I think it is clear that it was the mid-tier actors who had some way to go.

They can't just already do everything under the sun, but also are not so inept that they would fail even with the help of AI.

And that was our hypothesis.

Showing 361–380 of 1,784 · page 19 of 90