[empath75]

> But it's becoming clear that within a few more years, computers will be far better at absolutely everything than human beings could ever be.

Comparative advantage. Even if that's true, AI can't possibly do _everything_. China is better at manufacturing pretty much anything than most countries on earth, but that doesn't mean China is the only country in the world that does manufacturing.

[Philpax]

> AI can't possibly do _everything_

Why not? There's the human bias of wanting to consume things created by humans - that's fine, I'm not questioning that - but objectively, if we get to human-threshold AGI and continue scaling, there's no reason why it couldn't do everything, and better.

[seadan83]

Why not - IMO you perhaps underestimate human complexity. There was a guardian article where researchers created a map of a mouse's brain, 1 cubic millimeter. Contains 45km worth of neurons and billions of synapses. IMO the AGI crowd are suffering expert beginner syndrome.

[Philpax]

Humans are one solution to the problem of intelligence, but they are not the only solution, nor are they the most efficient. Today's LLMs are capable of outperforming your average human in a variety (not all, obviously!) of fields, despite being of wholly different origin and complexity.

[seadan83]

I don't think I agree. I'm trying to point out the 'expert-beginner' problem. We don't realize how much is involved in human intelligence. To the extent we think it is easy, that AGI will be here in a couple years. It's the same reason that in software "90% done is 90% left to go." We are way under-estimating what is involved with human intelligence.

An analogy I think is like crypto problems that would require 1 billion years to compute. Even if we find a way to get that 100x more efficient, we're still not coming up with a solution anywhere near in our lifetimes.

> Today's LLMs are capable of outperforming your average human in a variety (not all, obviously!) of fields

My impression is many of those are benchmarks that are chosen by companies to look good for VCs. For example, the video showing off Devin was almost completely faked (time gaps were cut out, tasks were actually simpler and more tailor made than they were implied to be).

Something I was trying to convey to a non-technical stake holder is that some tasks are stupid easy for humans, but insanely hard for computers - and vice versa. A big trick was therefore to delegate some things to humans and some things to computers. For example, computers are excellent at recollection and numerical computations - while humans can taste salt easily and tell you when something is too salty or undersalted trivially. In my opinion, AGI is an attempt to have computers do those things that are trivial for humans, but insanely tough for humans. There is a long, long way to go; getting that first 50% is the easy part, the last 50% (particularly the last 30% and the last 5%) IMO is several hundreds (if not thousands) of __magnitudes__ harder.