Debasish Pattanayak

Assistant Professor, CSE @IITIOfficial. PhD, BTech @IITGuwahati. Postdoc @UniLUISS @UQO @Carleton_U @uottawa Distributed Algorithms.

[Debasish Pattanayak @drdebmath]

codex rate limits feel like decreasing every time I use it.

[Debasish Pattanayak @drdebmath]

seems like claude has higher token generation speed (even with max effort) after the outage from saturday.

[Debasish Pattanayak @drdebmath]

this impression is not complete. at some point i need to test them both on the same task to compare. however i feel that is a waste of tokens since i don't have much.

[Debasish Pattanayak @drdebmath]

opus 4.8 does more work with the 5 hour limit compared to codex 5.5 xhigh

[Debasish Pattanayak @drdebmath]

More negative rewards lead to more precise outcomes.

[Debasish Pattanayak @drdebmath]

There is a bottleneck in RLVR that can only be accomplished by significant amount of comprehensive experimentation covering all parameter ranges.

[Debasish Pattanayak @drdebmath]

@petergyang codex always copies claude.md to create agents.md but the reverse is not true.

[Debasish Pattanayak @drdebmath]

Asking codex to fill word forms for claiming reimbursement. Filling the word forms work properly, but asking it to convert the word forms to pdf is where codex botches it.

[Debasish Pattanayak @drdebmath]

@TomasBabej @peterrhague I tried and it said something else, then i just ask claude to commit and push. Can’t bother to clean up all the files.

[Debasish Pattanayak @drdebmath]

@Conor_D_Dart I mean even after sacrificing sleep, i will get 5 blocks and each 5 hour block uses max 10 percent of weekly limit. No point in sacrificing sleep imo.

[Debasish Pattanayak @drdebmath]

Since claude reset limit, and its going to reset again soon, I am experimenting with /goal. Seems like it doesn’t do much differently than what it does normally.

[Debasish Pattanayak @drdebmath]

@teodorio Seems like a terminal behaviour from claude.

[Debasish Pattanayak @drdebmath]

@ClaudeDevs No wonder
https://t.co/dRQ6OXvpov

[Debasish Pattanayak @drdebmath]

Hate when this happens. Claude code on low effort does things fine but with xhigh effort it does something else and fails in weird loops. And then stops halfway without completing anything and burning through the limit very very fast. Need to think what to do as token poor.

[Debasish Pattanayak @drdebmath]

@jiangyan @Claude I know this error

https://t.co/dRQ6OXvpov

[Debasish Pattanayak @drdebmath]

Hate when this happens. Claude code on low effort does things fine but with xhigh effort it does something else and fails in weird loops. And then stops halfway without completing anything and burning through the limit very very fast. Need to think what to do as token poor.

[Debasish Pattanayak @drdebmath]

@claude rate limits are reset now.

[Debasish Pattanayak @drdebmath]

@jiangyan @Claude they filed for IPO, so that's something i guess. I was expecting a ratelimit reset for fundraise though

[Debasish Pattanayak @drdebmath]

@nvidia Also please bring back 256GB ram mac studio! @Apple

[Debasish Pattanayak @drdebmath]

Why is there no 256GB DGX Spark? @nvidia

[Debasish Pattanayak @drdebmath]

@scaling01 This optimization is what is worth doing. The output should be minimal with most effectiveness.