Haseeb Qureshi

An AI agent knows how to manage secrets.

You know, that's how it has its anthropic key, right?

If that wasn't true, it would leak its anthropic key and then, you know, boom, you'd have all your credits drained, which, you know, is happening sometimes, but usually not because you're leaking the key.

So all that is to say, I think the answer is actually that crypto doesn't need to move very far for it to be in the right form factor for an AI to use it.

The difference, the difference is that right now, the problem is that AI agents have not been trained to use crypto.

Right.

They've all this RL, all the training that we've been doing inside these big labs is all around coding and mathematics.

And then, of course, just general, you know, be a nice agent, you know, to give me some therapy, answer questions about medicine, whatever, this kind of stuff.

This is what people have been reinforced with learning, training these models on.

It was just last week that you saw at OpenAI the release of EVM Bench.

which is a EVM-based cybersecurity benchmark that kind of shows that all of a sudden, hey, looks like they're starting to train on this stuff.

It looks like they're starting to actually do some reinforcement learning to allow these things to operate in a blockchain environment, right now EVM-based.

And the same thing was true at Anthropic.

Anthropic, they released a paper, I think last year,

of them showing how well models were able to, inside of a harness that they designed, how well models were able to do cybersecurity attacks on EVM-based smart contracts, right?

First time we've seen this.

Now, that being said, you go look at the anthropic model cards.

Every anthropic model, they check to see how well can this thing do Bitcoin transactions.

They've been checking this for a very, very long time inside all the anthropic model cards.

If you go in there, you can see it.