Haseeb Qureshi

You'd make a shitty crypto podcast.

That's true.

I think, okay, we should be clear that we're not that close right now, right?

Like this whole, you know, this meter chart that I was just alluding to showing that Opus 4.5, or sorry, 4.6 can operate a task that's 14 hours, takes 14 hours for a human to perform.

That means that if you leave your Opus running for a week-

and you check on it a week later, it's going to be running in circles, right?

It's not the case that you can just leave.

Anybody who's tried this, anybody who's left their AI agent just running overnight, when you show up the next morning, it's not like, wow, it built the Taj Mahal.

You come the next morning, you're just like, wait, what were you doing?

You spent like $300 in credits just doing random stuff.

So this is what it looks like today, is that as these models get better and better, the longer you can leave it unattended for it to do useful work, that's what this meter test is actually measuring, is how long can you leave it unattended and have it do useful work and not kind of collapse into meaninglessness or get run in circles.

As this number expands, it's going to go from 14 hours to 30 hours to 50 hours to a week to two weeks to a month.

Like fundamentally what this is measuring is if I create a Conway and I let it run, how long will it continue to do useful things?

And the answer is, you know, it's like an Energizer bunny.

Like you wind it up and you see how long it's going to keep walking.

Eventually right now, the models...

they do run out.

There is no measure of a, yeah, this just does it infinitely forever and always stays coherent.

Eventually, it will become so big that the answer is like three years.

The answer is like basically infinite.