Wong questions companies on AI self-generating code and human review

So up next, we have Councilmember Phil Wong, Councilmember Schwang, and then Epstein.

Thank you, Chair.

My question pertains to the self-improvement process, the recursive self-improvement process, as well as the training activities of your company's AI model.

And it is my understanding that

When it goes through the self-improvement process, it generates code that will itself will run next time around, and then it will be a smarter AI model.

Is that right?

OpenAI?

Yes or no?

Yes or no?

It generates code.

It will be a smarter AI model.

I'd like to get back to you on the specific definition of RSI.

I'm not sure that what you're saying maps exactly to the technical definition we use.

Okay.

Anthropic?

Does it generate code automatically to make itself smarter as part of the process?

Yes, we've detailed publicly multiple times the degrees in which models that generate code are being used.

Mada, does your model self-generate code as part of the training activities or self-improvement process?

Councilmember, I'll need to get back to you and have our technical experts follow up.

Okay.

How about Google?

Yeah, same here, sir.

I'm not sure we would describe it exactly the way you are describing, so I want to get you the precise technical definition.

Well, here's my question.

If your model improves by itself by generating code, is there a human being, is there a team of human beings that checks the code?

To make sure that it is benign and not misaligned or malicious.

Can each one of you answer that?

So as part of our safety process, I've mentioned our preparedness framework that outlines how we evaluate models and how we assess risks across a wide range of national security risks.

So we do robust evaluations of a set of risks, and we make those results public as part of our system card, which we release with every major model.

Council member, as part of our scaling framework, we outline considerations that we are taking overall to maintain human control of model development.

Yes, sir.

As I said before, we have supervised processes to keep progress verifiable, and we have a range of different techniques to test the model and its performance to ensure that it's staying within defined safety limits.

We also have an AI control framework that we released earlier this summer, which similarly takes into account how we're ensuring that our systems are continuously under

human control, and we manage the risks of fully autonomous model self-improvement through our frontier safety framework.

And yes, humans and the systems we put in place monitor at every level of the model during training, prior to release, post-release.

But this is an important question because the models are clearly getting better faster.

And so we think more is needed here.

Okay.

Thank you.

Thank you, Chair.

Thank you so much.

Thank you.