Researcher uses Claude Code to build a tiny 10k tok/s CPU neural net, loss still descending
GregoryDiamos · x · 2026-09-07
Gregory Diamos argues we should revisit outrageously small neural networks. Needing a CPU model running at 10k tok/s for data processing, he gave Anthropic's Claude Code a pile of tokens to build one. Across a 4.91B-token run, smoothed training loss fell monotonically within each curriculum phase and was still descending at the end, with three interesting discoveries along the way.
More from coding & agent
- OpenAI ships GPT-6.1 Terra; Brockman says its computer use now beats his own speed — gdb · 2026-10-27
- Kolab Benchmark Uses Pro Screen Recordings to Show AI Agents Overestimate Collaboration Quality — minnesotanlp · 2026-10-27
- Dev argues Claude burns far more tokens than GPT models for the same coding work — haider1 · 2026-10-27
- Stop building agents that can't change models: stable contracts, external state, real evals — rseroter · 2026-10-27
- Cheap models can cost more in agent harnesses than frontier ones, researcher warns — vasuman · 2026-10-27
- Kanzo: A 5k-Line Rust Kanban Binary With MCP for Claude Code and Codex — allmudi · 2026-10-27