Jeet Ganatra — labwork-in-progress experiments and questionshttps://jeetganatra.me/fine-tuning nanochat on mathhttps://jeetganatra.me/lab#finetune-nanochat-math/https://jeetganatra.me/lab#finetune-nanochat-math/does GRPO actually help on AIME problems?Wed, 22 Apr 2026 00:00:00 GMTrunningrlmathagentsagentic eval harnesshttps://jeetganatra.me/lab#agentic-eval-harness/https://jeetganatra.me/lab#agentic-eval-harness/LLM-as-judge, but with rubrics.Mon, 20 Apr 2026 00:00:00 GMTrunningevalsagentsspeculative decoding from scratchhttps://jeetganatra.me/lab#speculative-decoding/https://jeetganatra.me/lab#speculative-decoding/how much latency can a draft model save?Sat, 18 Apr 2026 00:00:00 GMTrunninginferencelatencyreading: attention is all you need (5th time)https://jeetganatra.me/lab#reading-attention-is-all-you-need/https://jeetganatra.me/lab#reading-attention-is-all-you-need/still finding new things in the same paper.Mon, 30 Mar 2026 00:00:00 GMTpausedtransformerspaperslong-context experimentshttps://jeetganatra.me/lab#long-context-experiments/https://jeetganatra.me/lab#long-context-experiments/what breaks past 100k tokens?Tue, 10 Mar 2026 00:00:00 GMTpausedlong-contextinferenceloss curve forensicshttps://jeetganatra.me/lab#loss-curve-forensics/https://jeetganatra.me/lab#loss-curve-forensics/why did training spike at step 11k?Sat, 28 Feb 2026 00:00:00 GMTshippedtrainingdebugging