Balaji Srinivasan

So there was a very good counterargument that actually said that this study does not purport to show what people think it's purporting to show and that the task completion, it's like –

it's much more questionable than it actually looks, right?

And this is important because a lot of people's paranoia and stuff is levered on this.

Yeah, that's it.

I think that's it.

Against the meta graph, right?

Yeah, can we pull that up?

Pull that up, right?

So yes, this was like, this deserves more people looking.

But fundamentally, he says, actually, you know, the graph doesn't really look like this because A, a lot of the tasks aren't.

And B, like what is completion?

You know, they're TLDR.

Now, I don't want to overstate this because it's certainly true at a gestalt level that if you use Claude code or Claude co-work or something like that, it is possible for AI to complete long, complicated tasks than it could a few years ago.

That's clearly true.

However, it's also true that you still have to supervise it a lot and check the output.

It's a little bit like, you know, I compare it to a spaceship that can go in any, like, you know, coordinate direction that you point it in and move very fast, pivoting it and give it the route that it's moving on.

It's like a car.

It'll take you there very fast, but then you have to still give it a direction, and now the new thing is everybody's got a car, so now car race as opposed to you just being able to teleport somewhere, right?

Okay.

So this guy is pretty critical of it, and...