I got it to play roller coaster tycoon and it basically didn't do anything, so I'm not hopeful, it is not meant to be a smart model
@petergostev
-
Commentary sees model validation strengths limiting scientific exploration
By
–
Models are super-human on testing & validation, but this makes them worse at exploration and ideation. They just want to jump to one OK idea and validate it to hell. That's one blocker to applying them in research & sciences, but perhaps Bell is better at this, given the results
-
Commentary distinguishes chess performance from Elo optimization in Opus
By
–
My best interpretation is that Opus was better at getting higher elo and not better at playing chess
-
Astra defeats Opus in a thirty game chess comparison
By
–
The tournament is over – Opus WON the last game, unfortunately it still lost 24.5 5.5 to Astra 💀 …
https://ailearni
ngchess.ai-learning-chess.workers.dev/#/tournament -
Chess experiment finds larger context model performing worse
By
–
I can confirm at least for playing chess, the 800k context model is doing worse than the 300k one. …
https://ailearni
ngchess.ai-learning-chess.workers.dev/#/ -
Developer describes separate Codex and Claude Code workspaces for chess experiments
By
–
Astra in Codex and Opus in Claude Code, they have their own project folders and they can do what they like (take notes, etc) except run a chess programme
-
Continuous Chess Agents Retain Memory From Earlier Stockfish Games
By
–
Not independent, they are playing continuously, incl with the memory of their 200 games vs stockfish
-
Astra Chess Experiment Reports a Strong Lead in a Live Tournament
By
–
Should we stop this? Astra is winning 15-4 now 💀 …
https://ailearni
ngchess.ai-learning-chess.workers.dev/#/tournament -
User Runs a Three Billion Parameter AI Model on a Mac
By
–
so I've been running this 3b Super Intelligence model on my mac
-
Discussion Questions Whether Literature Anchoring Limits AI Mathematical Discovery
By
–
Hot take: the fact that models are taught to look for and cite the literature holds the models back from solving more maths. Sure, existing work is important, but I wonder how much the models get anchored and act deferential towards it