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

And this has just been the story for years now.

The one that people are most excited about is data filtration.

which is you don't train it on the scary CBRN, chemical, biological, radiological, nuclear stuff in the first place.

Now, this is something that frontier companies sort of say they do.

For a very long time, for example, in the open AI model cards, they've said, we do data filtration.

We filter out, say, child sexual abuse material, which everyone agrees no AI should ever be trained on.

And we filter out CBRN material because we think there's even legal questions about whether you'd be allowed to train on this stuff.

And yet they seem to have lots of knowledge.

So I'm not always clear how that's worked.

There was this great paper recently, O'Brien et al.

There was something outside of the labs.

It was civil society, third party actors who...

built a smaller LLM, but they explicitly took out all the CBRN information.

And this is not usually possible because it's really expensive to train models from scratch.

This is usually why only the companies do it.

It was very impressive that these folks on the outside were able to get enough compute together to do it themselves.

And then what they did is they spent a lot of money, maybe something like a million dollars of compute on training and then some fraction of that on fine tuning, trying to put all the stuff back in, the scary stuff.

And they found that even after hundreds of thousands of cycles of fine tuning, I mean, this doesn't sound much, but it's something like weeks of fine tuning, they couldn't fully recapitulate the dangerous biological capabilities that a model that hadn't had that data filtration would have otherwise had.

And this was really promising.

However,

Showing 961–980 of 1,784 · page 49 of 90