Alex Wissner-Gross

If agents become one million times smarter than two and so on, isn't there a diminishing return at some point?

I think yes.

So, Seth Lloyd at MIT was studying the question in the early 2000s of the physical limits of computation.

Does the physics that we have right now impose a universal limit on the fastest or smartest, for that matter, computer that you could possibly build in our universe?

And the conclusion that he came to is that, yes, there is a physical limit to the power of computers.

The fastest serial computer with the physics that we have today that we can imagine building is a black hole.

I've spoken about this on the pod previously, a sort of desktop black hole supercomputer where you maybe fire in the inputs via X-ray or gamma ray lasers.

The new Max video.

Yeah, no, when Apple gets around to actually launching maybe a new Mac Pro, it should be a black hole, maybe.

And the output readout could be via Hawking radiation.

So we know in principle how to build a black hole-based serial computer, the ultimate serial computer.

He found that under certain constraints, the fastest parallel computer might look like a box of plasma, a so-called plasma-based computer.

So we do know in some sense how to build

the smartest possible computer at the infra level that our universe will allow us to build unless there's a lot of surprising new physics.

And then that provides, in some sense, an ultimate constraint on the level of intelligence for agents that can be built on top of it.

I also strongly suspect that at the algorithmic level, we're going to find that there is a perfect agent algorithm.

Folks who've studied AICSI, which is a theoretical approach that's mostly popular among the AI theorist community, it hasn't turned out to be very useful in practice.

In some sense, represents an information theoretically optimal AI, including an AI agent and has all sorts of nice properties like Bayesian superintelligence.

It's not very practical, but we do know, at least algorithmically, what the point of diminishing returns is for the agent algorithm level is all.

Yes, there may be a lot of room at the ceiling, as it were, at the top, but the universe does seem to impose limits.