Gary Marcus

Rank 17 of 47
|
Score 70
4/10/2024, 5:07:50 PM
X
In reply to:
Shiven Sinha
@sinha_shiven
路
541d
Our work might be of interest to @rao2z @GaryMarcus @wtgowers @ead @ChombaBupe @vr4300 @mmitchell_ai @thtrieu_ @Yuhu_ai_ @lmthang
Shiven Sinha
@sinha_shiven
路
541d
This was a project with amazing co-authors @AmyPrb
@ponguru @druidoftheclaw @MatthiasBethge.

Thanks to the wonderful people at @bethgelab and @iiit_hyderabad for all the help and feedback!
Shiven Sinha
@sinha_shiven
路
541d
6/ We hope that this paves the way for future innovations in automated reasoning.

Overall, IMO Geometry felt like a chess endgame, fixed with little wiggle room. The real fun lies in identifying and teaching machines to solve the Geometry middlegames!
Shiven Sinha
@sinha_shiven
路
541d
5/ 馃弲 Remarkably, Wu's method solved 2 problems that even AlphaGeometry couldn't. Combining their strengths, we set a new standard, solving 27/30 IMO problems! For the first time, an AI method has outperformed IMO gold medalists.
Shiven Sinha
@sinha_shiven
路
541d
4/ This synergy not only places us 4 problems shy of AlphaGeometry's record but also marks the birth of the first fully symbolic approach that can rival an IMO silver medalist at Geometry. 馃
Shiven Sinha
@sinha_shiven
路
541d
3/ 馃 Algebraic methods like Wu's, and synthetic methods like deductive databases (DD+AR) also solve complementary problems. We find this surprisingly simple method solves 21/30 problems on a simple CPU laptop within 3 minutes.

The statement is a casual acknowledgment within a technical and academic discussion about AI and automated reasoning in geometry. It does not address broader societal issues or engage in public discourse.