Matt Freeman

I have the, like, saw the tweets in passing level of awareness of this.

So, I mean, I agree with you that I didn't love it when I saw...

that we're basically taking like a core document that governs their approach to scaling and deployment and making it weaker.

I mean, I think just facially that that's a bad look, right?

Easy, easy answer from me there.

But as to like, as for like, what are the actual changes between the previous version and this version?

Honestly, I'm trying to read this 20-page document while talking to you guys, and it's not going well, so I'm going to stop that.

I was trying to get an overall sense of what did they actually change concretely, and I'm unable to figure that out on air while talking to you.

It may be the case that it's not so bad.

If you look at the fine print, I admit I have a tendency to want to believe the best of people.

I don't know.

I mean, do you, do you, can you by any chance summarize like what exactly did they change with the new RSP?

Yeah.

Let's see if, let's see if Claude can summarize.

anthropic is releasing the third version of its rsp reflecting on two plus years of experience key structural change separates what anthropic can realistically achieve on its own from what it believes the broader industry needs to do collectively acknowledging that higher safety levels may be impossible for one company to implement unilaterally basically what you said steven

introduces three new mechanisms, a published safety frontier safety roadmap with publicly graded goals, a periodic risk reports with external expert review, and an industry-wide capabilities to mitigations map, all aimed at increasing transparency and accountability.

So my interpretation, tell me if you think I'm being overly generous, is they basically said, hey, we're not going to hold to the RSP, but we're going to introduce three new transparency mechanisms that will at least hopefully give some feedback loop where other people can see where things are going and, you know, be mad at us if we're doing the wrong thing, something like that.

So just to be a little technical here on the RSP, you know,

It's fairly specific where basically they test, you know, any given model on purely on capabilities and they kind of see what it can do.

And they're like, you know, how useful is this as a tool for building chemical weapons, for building biological weapons, for doing, you know, high leverage, extremely damaging acts.