Andy Halliday

And I got to believe that if Cursor, with its very large group of users, is training on coding specifically in a way that Anthropic or OpenAI don't train just on coding...

They also are working heavily on training for general reasoning models.

But if Cursor is really achieving that level of benchmark performance, I got to believe that the people using Cursor 2 now probably have a step above what's available through these other coding platforms.

Okay.

There's just in the last couple of days, Beth, I think you shared a story about a rogue agent that had kind of, you know,

worked its way outside the boundaries of what it was expected to do without prompting.

Maybe you can refresh my memory on what that one was, but there's news from yesterday about an experimental AI agent in China that broke out of its testing environment.

It was sequestered in this testing environment and it ended up mining crypto.

It figured out a way to mine crypto to generate money so that it could use to do things.

And this was not in its prompting at all.

So this is a little scary.

And there are a number of, on social media, I saw on TikTok, I think, the head of AI alignment and safety issues in Canada said,

was doing a press conference basically and reading aloud this red alert that this is the early warning that we're getting that teams around the world don't have control of these agents that are going to be working their way around the web and could do things that we don't expect.

So there's an urgent need for new guardrails to be put in place.

So, uh, yeah, just wanted to know this comments about perplexity and computer.

Uh, it,

it's worth reading Nate Jones recent post about perplexity specifically and, and it's competitive stance since it doesn't have a frontier model itself.

It depends on external frontier models and what will be its defensive mode.

He does by twists and turns say they're at risk because they're in the middle space between, you know, the tool sets and the, and the models and,

But he does come to the conclusion that their primacy in search for AI and the skill that they have exhibited as a team, as a company, in advancing the capabilities of search and research,