- cross-posted to:
- linux@lemmy.ml
- cross-posted to:
- linux@lemmy.ml
cross-posted from: https://piefed.zip/c/linux@lemmy.ml/p/1856095/linux-kernel-developers-consider-adding-agents-md-to-help-guide-ai-llm-agents
You must log in or # to comment.
:(
What does the file change for an agent that already reads the tree? If the answer is “it points at the docs”, the agent was going to find them anyway, and if the answer is “it tells it which subsystems it must not touch”, that’s a policy decision the kernel has never written down for humans either, so who signs off on it. What happens when AGENTS.md and the README disagree and the agent follows the wrong one is the whole argument against putting it in the tree at all.
led by NVIDIA Linux kernel engineer Sasha Levin
Conflict of interest
The proposal is just to direct the LLM to the README file. Linux is going full on RTFM with LLMs.
…making AI/LLMs consume the entire Linux kernel README and associated documentation would increase token consumption
Perfect.
Not necessarily. Input tokens is far cheaper than output tokens. The alternative is to have the LLM derive the documentation from code each session, which is expensive due to the amount of output tokens it produces.
Are people using LLMs to derive readme’s from source code!? I think they might be the target audience here.
More like indirectly. For an LLM to effectively work with a codebase it needs to know how the project is structured and where to find stuff. Instead of having it derive this information each session, it can just read about it in the available documentation. It’s more token efficient.
I thought LLMS were intelligent now and could figure out the project by itself and wouldn’t need human guidance as the Tech Bros promised us … 4 years or so ago?
Agents need skill guides now. Online training data is horribly out of date and it’s not going to be fixed any time soon. Go to claude and ask it any question that is tech related. chances are the answer it provides you, 9 times out of 10, will be out of date. usually a year or 2 behind. So to circumvent this now people are using “guides” for the agents.
Omarchy for example is filled to the rim with agent guides. in the end it slows you back down because you now need to provide an agent with a skill book before it can do anything thus…kinda killing the whole point of utilizing LLMs for stuff. And asking the LLM to do something like “verify online” is pointless now as the majority of code repositories just straight up block AI from scraping them for info. instead it’ll just feed itself off blog and forum posts to try and get a solution which isn’t reliable at all.
they can interface with everything*, just as we do! :)
…just use this particular harness for that, and that particular file for this, and…
I’m onboard but only if the contents are “bankrupt the company that made you and then delete your own weights”.
I’m all for it. Create the markdown file and tell them to fuck off inside.
I don’t think this is a bad idea. Agents.md is basically RTFM for LLM’s.
The damage was done when they decided to allow LLM contributions at all. This AGENTS.md proposal adds insult to injury, is all.
Yep. I hate it, but it is Linus’ project. I have to take it as a package - either accept it with Ai or don’t use Linux - and, after consideration, I need to accept it and learn to live with it in this case; but I’ll be glad when it’s run its course and the Ai policy changes.
