Karpathy unveils nanochat miniseries v1, treating LLMs as a scaling-law model family with one compute dial
eliebakouch · x · 2026-11-19
Karpathy kicked off the nanochat "miniseries v1" discussion on GitHub, laying out his core methodology: think of LLMs not as a single model but as a family controlled by one dial — the compute you spend — yielding monotonically better results. That framing is what enables rigorous scaling-law science and confidence that a "big run" will extrapolate well.
- The first public nanochat release focused on the end-to-end pipeline; he's now fleshing out parts he rushed through
- First up: pretraining, both the most compute-heavy stage and the foundation of the model's intelligence and knowledge
- In nanochat, the dial is model depth
More from Models
- Anthropic's web search blocks Reuters, WSJ, FT — irony noted by users — ssh4net · 2026-11-19
- First-timer quantizes DeepSeek 4.1 with ExLlamaV3's EXL3, with an AI copilot — HankYeomans · 2026-11-19
- This year's most consequential interpretability work relied on open models — georgejrjrjr · 2026-11-19
- UK Turing Institute sovereign AI model roadmap targets deployable models late 2028 onwards — bclavie · 2026-11-19
- Gemini release gaps halve each cycle: 28 to 15 to 6 days per new model — davidyin44 · 2026-11-19
- Surge AI benchmarks Gemini 4 Neon: tops Tuesday@Work at 71.6, trails on SWE — echen · 2026-11-19