Breakthrough on famous open math question by AI

[Dryas]

https://x.com/OpenAI/status/2057176201782075690

Yea, so AGI seems quite likely.

[Leo Gura]

You can't believe any such news stories.

We have no idea what human involvement was part of this.

[Dryas]

2 minutes ago, Leo Gura said:

You can't believe any such news stories.

We have no idea what human involvement was part of this.

You can see the researchers testifying in the video of the tweet though. Seems like almost impossible to coordinate a lie like that.

[Jodistrict]

I did some investigation to see if it was using brute force or systematic search of possibilites, and this was not the case.  It was a general system, not tuned to mathematics, and came up with the solution independently.

"Instead, a general-purpose OpenAI reasoning model achieved this by independently bridging two completely separate fields of mathematics, finding a hidden conceptual shortcut that 80 years of human mathematicians had entirely missed."

[Leo Gura]

I'm not saying it's an outright lie. The distortions are more subtle than that.

There are many ways to massage and frame what happened in this case.

And of course AI can find unseen connections between different fields of data.

What is the connection between Hitler and a kangaroo? I'm sure AI can come up with something that no human has thought about before there.

If you put all the world's data into a box, that box can find interesting new ways that data interconnects with itself.

Edited 4 hours ago by Leo Gura

[Carl-Richard]

AGI but can't screw open a Coca-Cola.

[zurew]

Maybe, but be careful, because previously there was a case where people initially thought AI did the same with other Erdos problems, and it turned out that it just managed to find the solutions on the internet.

https://x.com/BorisMPower/status/1979328512923439455

[Carl-Richard]

@zurew The AI couldn't even find that out