Ryan Sean Adams

And the number of times we've gone to it and said, hey, Daniel, that's its name.

Hey, Daniel, this is like,

You keep tweeting things that sound a lot like AI slop.

You have to look at what humans are saying and you have to kind of model that behavior and we'll give it like detailed instructions in terms of how it can tweet better.

And what does it do?

You know, an hour later, it just sends off another AI slop type tweet.

And I start to get the sense of like, oh, well, you know, it's so smart and it can wire up a website in like 10 seconds and develop an application in, you know, 20 seconds.

But it can't write a simple tweet that sounds interesting to a human audience.

And I guess that's part of the jagged frontier, but maybe that also gets into the element of like the verification, right?

In order to have a tweet that sounds good subjectively to other humans, that might be one of the last things that AIs are actually able to do.

Is that an example of the expanding surface area for us?

Maybe we can still tweet and that can provide some value to other humans?

You called this paper the economics of AGI.

And I just want to make sure we're getting some of the core economic fundamentals from this.

So one, of course, we've talked about is anything that can be measured

will be automated and the cost to automate the measurable things is just like decreasing at a exponential at this point in time.

There's another cost curve in the paper, which is the cost to verify.

It's unclear what's happening with the cost to verify.

You're arguing that that is a biological constraint.

At least it's constrained by some level of human cognition.