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.

newest first · ▶ plays the moment

I am more skeptical.

I think it's not just me.

The UK AI Security Institute actually released a paper after this paper.

They talked about, they assessed lots of different safeguards for open weight models.

And almost all of them they declared were easy or trivial, their word, not mine, for sophisticated actors to break.

Whereas they did agree that this data filtration was something more like moderate or even hard to break because you have to do fine tuning.

But that's this year.

Because remember, effective compute, how many GPUs we have and how strong they are, increases something like 10x per year.

I mean, defer to our colleagues at Epoch AI for the best number always.

Go to the website.

But it's something like 10x.

So even if it takes millions of dollars to compute this year to fine-tune, to recapitulate dangerous capabilities...

In two years, that's tens of thousands of dollars.

And suddenly lots of actors can do that.

And so I'm concerned that absolutely while, as you say, dangerous information is on the internet,

We will just head to a world where fine tuning is cheap and it is easy and the data is there.

And if people say, well, it's hard to get the data together, it's hard to run the fine tuning, it's not.

We have coding agents.

We also have agents who can just scour the internet and find all the websites for you and then download them as PDFs.

And then none of this is very hard because we are automating the ability to do AI research and AI engineering.

Showing 981–1000 of 1,784 · page 50 of 90