Thibault Sottiaux, Engineering Lead on OpenAI's Codex team, proposed adding batch compute support to Codex under the label slow mode for handling compute-intensive coding tasks through grouped or delayed processing

@thsottiaux Yes there are often tasks I can kick off before bed that don’t take 8 hours and I would be happy to pay less to go slow.

@thsottiaux no, don't believe it

@thsottiaux 100% yes!! Specially for long /goal tasks!

@thsottiaux No

@ah20im Tell me more Ahmed

@thsottiaux I am still getting these errors so often and there is no fix to it yet. This has been a huge time drain trying to get things sorted and rearrange my schedule. I appreciate the token resets, but they don’t give me that time back. Could you please help fix this issue?

@thsottiaux For automations yes!

I always have a bunch I run at random times in the morning.

@thsottiaux yeeeessss, but as an option , as u said slow mode , sometimes i want multiple parallel experiments on the same piece of code that don't make sense if done sequentially , example is launching many optimizations based on competing hypotheses

@thsottiaux Oh that’d be cool yeah, especially when paired with /goal

@thsottiaux We need to prioritize /goblin

@thsottiaux as a matter of fact , i don't think batch compute should be up to the user in most cases, if the model figures out there are subtasks in the user query that don't need to share context , then it would be best to run them as seperate batches then aggregate

@thsottiaux basically any branching where the accumulation of techniques doesn't make sense (still talking about perf optimization), should be batched, less context for the model and more , let's call it direction

@thsottiaux seperate sample*, example that im working with: user isn't sure which technique will work best , doing fft conv over 2 large sparse matrices or doing small kernel convs then accumulate reduce or ...

@thsottiaux hope I made sense!

@thsottiaux personally i only need a fast response from an agent when i'm attending its output in real time , but if my goal is to explore possibilities, i would rather check it later for correctness and results to choose whether to keep working on that branch or not!

@thsottiaux Yes

@thsottiaux yes my backlog is exploding

@thsottiaux I would do the opposite and only allow /superfast and introduce 1k/month and cancel everything else

@thsottiaux Fantastic idea. I would love slow mode for overnight tasks or tasks that I'm not in a hurry to finish on Codex.

@thsottiaux Would be dope for automations yes