@steipete Is it? I was looking at ccusage data, I feel claude does a lot more re-read, but they should be priced much lower. FYI, there are 2 claude, 1 codex accounts (used to full to my best effort)
https://t.co/JaSRi8Xmzx
4 pre-trained models in the gallery if you just want to play, or train your own from scratch in 5-15 min and watch the loss curve fall. Curious what you train it on.
downstream of @karpathy 's nanoGPT in spirit. cc @greggman if you find the WebGPU side interesting
A few things that made this actually work: Flash Attention 2 in WGSL, weights stored packed-fp16 with fp32 accumulators (gated by a numerics check on your actual GPU), and a SharedArrayBuffer + worker setup so the UI stays responsive while training.
Asked why it happened:
107GB of logs were written in ~3 minutes (10:55 to 10:58 today). Multiple subagents (ae85830b839d, ac1de4986f09, a139e167925a) were all writing to the same output file simultaneously, and they were doing things like curling websites
(Scraping career pages for your resume tailor).
The bug is that Claude Code's task output system was appending every progress tick and every message from all subagents into one shared file at an absurd rate — writing ~600MB/second of JSONL logs. This is a Claude Code bug, not something
You caused.
It all happened today around 10:55 UTC (likely when you last used the resume-tailor project).
I had too many subs in my mail. And Gmail does not have a super convenient way. So I build a small tool to show me some basic analytics and help me unsub easily. It just needs reach access and doesn't store anything, feel free to give a try if you have this problem too.
Releasing a small tool to arrange your GitHub stars into lists and tags. Because GitHub lists hardly work for me, I have ~1k stars.
You can even share your lists and possibly review other repositories. Let me know what features would you like to have in this