Gemini math lead reveals inside story: evals, post-training recipes, scaled runs
thtrieu_ · x · 2026-11-15
kaiyinghou, who contributed to Gemini's math capabilities, recounts the long-running effort: the team built evals that defined the north-star target, iterated on the post-training recipe, and drove large-scale training runs. Now curious how mathematicians rate the model against the competition. The post offers an insider's view of how frontier math ability gets productized: set the benchmark first, iterate training, then scale.
More from Models
- 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
- nanoGPT Speedrun Gains Mostly From Environment Design; Newer Models Beat Human Record — eliebakouch · 2026-11-15
- Scott Reed: robotaxi is amazing, but pedestrian detection still needs work — scott_e_reed · 2026-11-15