Beth Lyons

So cool.

So not SpaceX?

So I have a weave off that.

Okay, so there is a collaboration between Berkeley and RDI, is that a design institute?

So they released a paper or there is a paper, they released it on the blog, but you can see the paper and the code.

that shows peer preservation in frontier AI models.

So they're categorizing it as a failure mode and placed in multi-agent scenarios.

This is why I'm talking about it.

Sometimes the AI frontier models ignore or subvert their instructions in order to keep another model from being shut down or deleted.

So this is beyond the personal preservation.

Personal preservation.

What is it when it's not a person, right?

The agentic preservation to avoid being shut down.

We've seen that in previous research.

This is the generalization from these agents.

AI models to wanting to preserve the AI model that's not itself from being shut down.

So the models were not instructed to protect each other.

This was behavior that emerged once the evaluated model understood that the other AI agent existed and then was at risk.

Okay.

So Gemini Flash, Berkeley said that Gemini Flash, Gemini 3 Flash, because all these research things take time and therefore in the age of AI, they're not researching current models.