Hacker News Reader

Comment by skeeter2020 | Hacker News Reader

skeeter2020

the headline hints that there's some sort of non-obvious factor that's going to be revealled. I scrolled past countless redundant and information-sparse graph-like figures and never found it; moved on.

If you've only got a paragraph worth of information to share, say it and let us get on with our lives.

View on HN

Well I couldn't find any other thorough dataset on this topic, so in that sense this is non-obvious since it took weeks to assemble the information. And it was fun doing it using the LLM Wiki technique.

> it was fun doing it using the LLM Wiki technique

What is that?

https://gist.github.com/karpathy/442a6bf555914893e9891c11519...

Could you share the tools you used to do this? I'd love to organise my esoteric research side quests like this.

Git + claude code in yolo mode. In the first prompt, I passed it Kaparthy's gist, and had it put together a high level plan of all of the sections that needed to be written to complete a vision I provided. Essentially put together a complete wiki on everything for getting global hardware certification.

I then had it loop once an hour. It would pick the next wiki to write, research it, gather raw sources, and then synthesize the wiki for me and push. I could nudge it in between hours if I wanted.

Do you add this as a skill or permanent prompt, making it always maintain wikis in the background? Or do you direct it to make these only when you're in the project?

There's skills if you want. I didn't want to do that since I don't feel like I need the smarts to work on the LLM wiki in every coding session. I like to keep my context clean and scoped to what I'm working on.

I am running this in a long running session that spawns a subagent once an hour. So the context of the main session doesn't get out of control.

The general idea is to have the LLM maintain longer-term context/background by storing it in a format/structure that's akin to a standard Wiki. The result is (hopefully) a series of human-readable and editable documents that's developed and maintained by the agent.

There's great coverage of it at https://gist.github.com/karpathy/442a6bf555914893e9891c11519...

It's actually also now a base capability in the Hermes agent and has been really helpful for me, at least.

Reply on news.ycombinator.com