cuddlyogre
· Level 12
random
Pro tip. If you have Codex working on something in batches, make sure you tell it to work on the largest batch size it can, or else it will split your 3000 step project into 1500 batches of 2.
Comments
> '(...) or else it will split your 3000 step project into 1500 batches of 2.'.
Sounds like a good optimization.
/s
So... you've learned that lesson the hard way, I'm guessing?
Yup. I was running a ghidra database through Codex and having it document everything and it decided to only tackle two functions at a time. Even when I told it to do more at once, I'm not convinced it actually did much more that 3 or 4.
Can it do that for concurrency maybe? What's the downside?
Edit: ah, I thought you were talking about batching AI calls server side for generating training data or so. In your scenario sounds like much overhead. But normally running smth batched in batched of 6 like 6*500 concurrently sound like very good idea. But many times stuff like that is blocked. I wrote many things that open router couldn't handle so ending up at openai, grok or DeepSeek. DeepSeek allows s conncurrency of 1200. That's a flex.
I worked so much with DeepSeek, it's a blast in so many ways.
I haven't tried concurrency mainly because of the usage limits. I imagine if I had unlimited tokens, the goal I had that ran for 5 days would have taken maybe a few hours.
You know, if you use Sonnet with your x5 or whatever, you have a very powerful model that you can use very concurrently. Don't only use the last model. It's many times not worth it. Especially not fabel.
So lately, since opus 5 and stuff, i noticed that thing not listening to me anymore. Now, I use Muse and the muse-cli and you know what? It uses the same claude.md files and THEY OWRK. Goddammit. It is really the fucking new claude models that fucking fuck suck.
@cuddlyogre if you have some money left, try the meta muse subscription for 13 or 5. 5 was 50 calls per 5hrs and too limited for me because that model is FAST. it's very convenient to work with and the quality is really flagship quality. Would say better than deepseek and close to anthropic. But to be fair, it did not fuck up hard yet. I use it for really complex stuff at the moment. Every change needs so much checks, that this model is way more useful than anthropic because of its speed .