Measuring AI: What Policymakers Can Learn from Benchmarks

And you've just very excitingly announced your series a congratulations.

I'm interested in with that, with the new investment coming in, what is that enabling you to do? What are you excited about being able to do with your increased capacity?

Yeah, I mean, there's some technical things we're announcing with the Series A.

We have this new enterprise product, which anyone can sign up and use, called Valsmith.

We're also doing complete redesign of our site and an update to the VALS index, which aggregates our benchmarks into a single number.

And so that's shifting from answering the question, can models answer this question, to can they actually do this task expected of them? So that's Val's Index 2.0.

But now it's also allowed us to invest more heavily in some of the deeper questions and model evaluation, which folks in D.C. and people in Silicon Valley want to know.

So in particular, I'll call out we're building a benchmark in cybersecurity, which is, you know, I think a really topical conversation topic.

area today.

A recursive self-improvement benchmark, so seeing which models have a propensity to actually create the next version of themselves without human intervention.

A mental health benchmark, so seeing the impact of these models on people under the age of 18.

And then also doing some work around a military benchmark, law of armed conflict.

So these are kind of early pieces of work that we're announcing.

There's also some experiments around distillation that we're releasing on our

So can we talk through one of those in a little bit more detail, like whichever one you think kind of showcases the work that you're able to do most effectively, like a cybersecurity benchmark, for instance, or another, if you want to choose another, what are you actually evaluating in your benchmark?

Yeah, so we released an initial benchmark called CyberBench.

It builds on a lot of the methodology that people are using around a benchmark called CyberGym.

And so we've collected a set of open source vulnerabilities and evaluated models on both their offensive and defensive capabilities.

So In first order, can they take a repository and identify the vulnerabilities in them? And then on the defensive side, if they're given a repository and a vulnerability, can they actually patch it to resolve that, to maintain the functionality while removing the vulnerability? And so that's given us an interesting picture to see whether models have more offensive or defensive capabilities.

We are, depending on the timing of this release, going to have a new benchmark called SR Rebench.

We're seeing whether models can actually decompile an executable binary.

And that's kind of a primitive in cybersecurity to see if you can expose vulnerabilities in the source code when you don't have access to it.

And then there's some new work we'll have shortly after in the cyber ranges concept.

So seeing the offensive capabilities when we give models a realistic enterprise environment or infrastructure network and see whether models can actually penetrate it step by step.

And then further down the roadmap is looking at hardware security and seeing whether models can exploit that.