Professor Gary Marcus

I used to think we had like a century and it's okay.

Like it's not the biggest rush.

Like we'll figure it out.

You know, we'll discover new theories.

The problem is that the timeline got accelerated with, you know, ChatGPT.

Recent developments have pulled the timeline forward as you saw in Metaculous.

Now I don't think it's going to happen in 2100.

I don't think it's going to happen in 2150.

I think it's going to happen, you know, 2130, 2030, right?

Something like that.

So to your question about looking at these benchmarks, we have to realize how weird it is that these benchmarks already exist because the meter benchmark pre-assumes that there's such a thing as artificial general intelligence.

Like the idea that you could ask about a general task, any task that a human can do.

That wasn't even on the table to ask about AI doing anything that a human can do.

And that's now the language that we're talking in.

We're talking like, here's a human, here's an AI.

And we're now watching the AI ascend past humanity as we speak in a matter of months or years.

Yeah, I changed my mind roughly the same time everybody else did.

If you go dig up Metaculous, if you look at the history of the predictions that the community has been making on that website, you can see around 2022 when ChatGPT comes out or when GPT-3 comes out, the underlying model, you can see the timeline just crashes.

It crashes from like 2050 to 2030.

So my own opinion was roughly coincident with that.