alex000kim

No, Uzbeks themselves use “non” when using Latin alphabet. Fun fact: in the past Uzbek language has been written in arabic, cyrillic, and, since ~2018, in latin.

Source: grew up there

same experience last year, existing without having to stop was especially surprising

That directory is huge already! I guess the index.md helps the agent find what it needs, but even the markdown file is very long - this would consume a ton of tokens.

Also I wonder who/what decides what papers go in there.

In the blog post, the agent is allowed to do its own search.

yup, as the blog says

> The full setup works with any project that has a benchmark and test suite.

so having a clear and measurable verification step is key. Meaning you can't simply give an AI agent a vague goal e.g. "improve the quality of the codebase" because it's too general.

I am sure this would works well in general. There is a challenge wrt to how to make them communicate effectively to e.g. 1) avoid duplicative work and 2) allow them to combine/overlay each others' findings to yield even better results

sounds similar to "LLM Knowledge Bases" https://xcancel.com/karpathy/status/2039805659525644595

technically you're correct, but look at the prompt https://github.com/alex000kim/claude-code/blob/main/src/util...

it's written to _actively_ avoid any signs of AI generated code when "in a PUBLIC/OPEN-SOURCE repository".

Also, it's not about you. Undercover mode only activates for Anthropic employees (it's gated on USER_TYPE === 'ant', which is a build-time flag baked into internal builds).

Oh right, I just saw https://news.ycombinator.com/item?id=47582220 will update the post with this link

Author here. I've seen the docs you linked to: Slurm uses "gang scheduling" to mean something specific (timesliced oversubscription where jobs alternate on shared resources).

I'm using the term in its broader CS sense: all-or-nothing co-scheduling of related processes across multiple processors [1]. This is the definition used across the K8s ecosystem e.g. Volcano [2], Kueue [3], and its Coscheduling plugin all define gang scheduling as "all or nothing" allocation.

I still stand by the origianl claim:

Slurm allocates multi-node jobs atomically, while vanilla K8s doesn't. its default scheduler places pods as resources become available, leading to partial allocations and deadlocks for distributed training. It's just a terminology clash. Thanks for the comment anyway.

[1] https://en.wikipedia.org/wiki/Gang_scheduling [2] https://volcano.sh/en/docs/plugins/ [3] https://www.coreweave.com/blog/kueue-a-kubernetes-native-sys...

This was so clearly LLM-generated that I couldn't get through the whole thing.

I created this PR to make it easier for folks to train and serve it on any cloud (or their own K8s): https://github.com/karpathy/nanochat/pull/18

Kind of. Fastai abstracts a lot more of Pytorch then what keras does to TF.