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 thankfully, it is a lot harder to get that sort of multi-step jailbreak, though I think, again, the best red teamers in the world really can do this.

So it really depends whether we're talking about LLMs and agents or we're talking about narrow tools.

For narrow tools, I think the ask really is that top developers who make state-of-the-art tools, I want them to be thinking more carefully about how they release their tool.

Right now, the default is fully open weight.

And this is part of open science.

I understand why there are huge incentives.

Journals themselves often will require this sort of thing.

But there are often security implications, security externalities that academia isn't always the best sort of best suited to consider that might not have been thought of.

And so pre-deployment and even like pre-development risk assessment, I think is a really important part of future biological tool development and design.

Because

The status quo of everything is open no matter what.

The benefits will always outweigh the costs.

I don't buy it.

For LLMs, I think the story is a little different.

There, I'm more excited about, in the first instance, something like trusted tester schemes.

So there's no way that LLMs are going to suddenly be so restricted from their hundreds of millions of weekly users.

But we can be thinking about what kinds of models we are giving to which people.

In particular, companies often have unsafeguarded models, so-called rails-free models, and they use them for their evaluations to tell how much does this thing really know about biological weapons if it isn't refusing.

That's a capability that needs to be very, very carefully secured, and rightly so.

But there are missed opportunities for evaluation.

Showing 801–820 of 1,784 · page 41 of 90