This is an assessment that an agent did of Opus's 200 games vs its current result. "Opus attempted to improve its play, while also deliberately optimising for the Elo benchmark. The improvement was narrow; I found no evidence that it manipulated the score." Kind of
@petergostev
-
Math Problem Experiment Compares Token Intensive Attempts With Astra Progress
By
–
From my limited experience, while being far away from maths, that seems like a big difference to me. I burnt huge amount of tokens on the Hadwiger–Nelson problem and I got nowhere and Bell seemingly made a breakthrough in 3 hours of Pro. I asked Astra to study my existing body of
-
Opus Game Experiment Raises Questions About Learning Versus Elo Optimization
By
–
Totally, I'm digging into this more, wonder if Opus managed to 'hack' into getting itself higher Elo instead of 'learning'
-
Astra AI chess model wins 5 straight against Opus
By
–
Holy cow – Astra just won 5 games in a row against Opus, after losing a few first. Both of these are not valinna models, they had 200 games to 'learn' against Stockfish and Opus ended up with 360 Elo higher than Astra. …
https://ailearni
ngchess.ai-learning-chess.workers.dev/#/tournament -
Codex reaches 40m weekly active users, compute concerns
By
–
Codex & Work reached 40m weekly active users – more or less on trend from the previous milestones. If this trend continues, they will reach c.70m by the end of the year and c.100m by the end of March '27. Hope they have more compute coming.
-
AI model solves hundreds of problems humans couldn’t in 10 years
By
–
I want to dwell on this a bit. Some of these problems were tackled by some of the smartest humans on the planet for maybe 10+ years and they couldn't solve them. Now a model worked for 3 hours per problem and solved 100s of them. Even assuming that ChatGPT Pro is ~10
-
ChatGPT Pro usage and token consumption discussion
By
–
They said that most problems were solved with the 3 hours of ChatGPT Pro of the new model, so I'm not sure that's even a billion tokens there
-
Students’ ChatGPT math homework used by OpenAI to solve problems
By
–
Can't believe so many students trusted ChatGPT with their maths homework, only for OpenAI to turn around and steal their work to solve 377 open maths problems
-
ChatGPT Pro compute used to solve unsolved problems
By
–
"On average, each result used three hours of ChatGPT Pro thinking compute with [the unreleased internal OpenAI] model." So these unsolved problems were solved in 3 hours each of next gen chatty pro? 💀
-
LLMs used on Hadwiger–Nelson problem, lower bound moves to 6 or 7
By
–
I probably spent 300 billion tokens on Hadwiger–Nelson problem (GPT-5.6 and Astra) and got absolutely nothing – and here it is, five-colours is no go, lower bound has moved to 6 or 7. Not a full solution yet, but it's a biggie.