Pete Huang

AI has always been a moving target of a definition.

I mean, at some point, even the autopilot systems on planes, that was called AI.

Even the programs that play chess against you on a computer, those were called AI and

All of this AI up until the last few years is very narrow.

Like that AI program that helps you fly a plane cannot also play chess.

They're all super specialized to what they were trained to do.

By whatever description you can think of, GPT-4 is probably somewhere in the range of a high schooler when it comes to reasoning capability, even if it's regularly beating human averages at complicated tasks like medical school exams and things like that.

As we continue, the big debate is about two general levels of AI capability.

One is called artificial general intelligence or AGI.

The second is called artificial super intelligence or ASI.

AGI is an AI that matches humans.

Anything a human can do, AGI can also do.

ASI is an AI that exceeds humans.

It can do things that humans can't.

Now, this is like Skynet from the Terminator series.

This is an uber powerful system that can take over anything if it really wanted to.

So let's talk about AGI, ASI, and what Leopold Aschenbrenner is seeing in his perspective.

Getting from GPT-2 to GPT-4, from a preschooler to a high schooler was simply a matter of training more.

The models didn't really change all that much.

You just jammed more training and more data in there.