LLMs are like an athletic child - they can do some basic moves, but in general they're uncoordinated.

Occasionally they do something that surprises you. If you put them inside a search loop and spend huge sums of money you'll get some nuggets of value, but for most hard problems the cost isn't worth it.

And they work best with a lot of steering.

From that POV they're exactly like human culture. Most scientific papers are trash, a few are gamechangers. Same with the arts, etc.

But we've gone from 'Can write sentences' to 'Can do frontier math problems' in maybe five years.

We absolutely do have AGI now. It's not very good AGI, but it's probably better than a lot of people were expecting.

Being better than humans at everything isn't AGI, it's ASI. A lot of the comments here seem to be lamenting that we don't have that yet, which is an interesting data point about how far the goal posts have moved.

Edit: also it makes things like this possible. So...

https://x.com/alifarhat79/status/2118478127202713778

if you search papers with it, and from there search related papers and have it do literature reviews, it comes up with very useful white areas uncovered, and from there problem statements and experimental designs to uncover the white spots.

it is nothing but combining known knowledge that begets new.