nanoGPT Speedrun Gains Mostly From Environment Design; Newer Models Beat Human Record
eliebakouch · x · 2026-11-15
The author shares an experimental observation: most nanoGPT speedrun gains likely come from environment definition—tuning hyperparameters so the model is actually incentivized in that direction. Newer models (e.g., Opus 5.5) already beat the human record on this task, with a blog post promised soon. He further speculates—unverified—that current internal Anthropic models (or non-nerfed base models) are trained on environments rewarding scalable methods and may already match or exceed human researchers at such tasks.
More from Models
- Gemini math lead reveals inside story: evals, post-training recipes, scaled runs — thtrieu_ · 2026-11-15
- Blogger saves $53 in 43 hours after switching from Claude Code to Gemini 4 Carbon — mazzaTalk · 2026-11-15
- SatBench: 63% of lab-cited AI benchmarks already saturated, and solving is speeding up — emollick · 2026-11-15
- LFM2.5-24B-A2B praised for local agent use: 37 tok/s on an AMD 780M iGPU — jeremyckahn · 2026-11-15
- Will Depue's Blackwell 'forensic evidence' claim on DeepSeek hits snag: V4.1-Flash not cracked — teortaxesTex · 2026-11-15
- Scott Reed: robotaxi is amazing, but pedestrian detection still needs work — scott_e_reed · 2026-11-15