NanoGPT speedrun sets record: 11.3% faster via architecture-only change, paper coming
yoavartzi · x · 2026-10-06
The NanoGPT speedrun project has set a (pending) record, improving over the previous best time by 11.3%—by far the largest architecture-centric improvement since December 2024, achieved without extra n-gram tricks or more CPU compute. The authors say a paper is coming soon.
More from Infra
- Reflection says Beam hit ~19% MFU with 92% goodput during its 4-week RL run — alexpolozov · 2026-10-06
- SGLang lands layer_boundary for cross-model reuse of TP/DP/CP semantics — BanghuaZ · 2026-10-06
- Scam alert: SpotGPUs.com fakes 'transaction errors' to steal crypto deposits — Equivalent_West7788 · 2026-10-06
- How a ChatGPT-like system actually works: request-to-stream architecture explained — jawadhamza · 2026-10-06
- SGLang Summit 2026 lineup: Intel CEO, Lilian Weng, Perplexity CEO headline SF event — BanghuaZ · 2026-10-06
- M5 Ultra vs 2 DGX Sparks: Local LLM Hardware Put to the Test — kaggie · 2026-10-06