Hacker Newsnew | past | comments | ask | show | jobs | submit | maille's commentslogin

Concrete cost data from today: rebuilding SQLite in Rust via planner/worker split cut agent spend 22x


tldr:

GPT-5.5 Codex model exhibits a clustering phenomenon in which reasoning_output_tokens cluster at fixed values spaced 518 apart.

These stuck responses at fixed thresholds are strongly correlated with errors in complex tasks.

Observed phenomenon is specific to GPT-5.5; it is much less prevalent in GPT-5.4 and almost absent in GPT-5.2 and 5.3


Intersting piece that was posted here and there in HN's comments but could deserve more visibility. Many points make sense


Looks a lot like Codex!


What are your use cases?


Wondering too


Does it work on windows? And how do you then remote in?


Bad at coding, but would it be good at code review?


Good compared to what? Nothing? Probably better.


How can I learn that clever prompting?


Try to pack as much clear work into your prompt as you can so you don't go back and forth.


Do hacks like “read prompt.md, and follow its instructions. When you’re done, read it again and follow its instructions.” And then you have some background process appending to the file to keep it warm and you just keep writing there?


You could do that. I was just trying to say that if you make your original prompt complete enough, and you have well-defined success criteria, you can tell it to keep going until they are met.


Agreed - my experience mirrors this.

> "Fix the following compile errors" -> one shot try and stops.

> "Fix the following compile errors. When done, test your work and continue iterating until build passes without error" -> same cost but it gets the job done.


There is a limit on how much copilot can do in one request, pretty generous but after some time vscode will say "this request is taking very long, do you want to continue" and that would count as a seperate request


> but after some time vscode will say "this request is taking very long, do you want to continue" and that would count as a seperate request

I don't think that's true. In VS Code, that's also configurable via the chat.agent.maxRequests setting.

There was absurd latency in the Copilot Opus 4.6 model on 1st and 2nd April which led to lots of my requests timing out with nothing to show though.


> chat.agent.maxRequests

"Maximum number of requests that copilot can make using agents"

I don't get how this setting is relevant?


Let's say I want a free, local or free-tier-llm, simple solution to search information mostly from my emails and a little bit from text, doc and pdf files. Are there any tool I should try to have ollamma or gemini able to reply with my own knowledge base?


https://onyx.app/

This could be useful.


If you are still ready comments, have you heard of this? https://docs.kiln.tech/docs/documents-and-search-rag


Are you using it? I will definitely give it a shot, any pointers to online resources will be appreciated


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: