Claude Code with 4.8 was given the same prompt, and the same goal, but kept insisting I was wrong and the goal is not achievable.

My number 1 reason for preferring Codex the past 2 weeks. So frustrating.

@DimitrisPapail The weekly Dimitris-tease-post that will result in a banger blog down the line

@DimitrisPapail user: "you did great job. next step - 25kB target. go ahead, you have 10 hours till I wake up"

@xeophon

@DimitrisPapail Similar experience here. I tried opus 4.8 on the symbolic gsm8k solver you proposed. Even after I told it others had reached 15% and pointed it to your original tweet, it kept insisting the goal was impossible and that 15% must have been an LLM predicting the operations.

@vorushin i literally asked it to go to 21kb lol

@DimitrisPapail which Claude effort level did you use?

@DimitrisPapail Seeing a lot of people upset with 4.8..is this a sign of model collapse? Or is it impossible to tell because anthropic could be throttling compute? But why would they do that right after a release of best available model ??

@DimitrisPapail Interested.

@DimitrisPapail Trying to win the Hutter Prize :P

@DimitrisPapail What happened with the small target? You wanted it to be as close as possible to 50kB, and not smaller?

@DimitrisPapail i like the concept already

@DimitrisPapail This is cool, would be happy to chat, I have done something similar sounding with 4D Gaussian splats.

@DimitrisPapail would love to lear more about your setup !

@DimitrisPapail something new from Codex again?

we love it

@DimitrisPapail love when the model has the courtesy to show you exactly where you went wrong

@DimitrisPapail i'm sorry Dave. i'm afraid i can't do that.

@DimitrisPapail What’s this graph about?

@DimitrisPapail I got similar experience! Or it claimed to solve something it never did.

@DimitrisPapail 2 weeks?? More like 6 months