John Langford weighs in on FBT note: fixed-point convergence and quantized caching
JohnCLangford · x · 2026-11-22
Vowpal Wabbit creator John Langford comments on xidulu's FBT technical note, which addresses fixed-point iteration convergence, speculative decoding via 3-pass verification with pipelining and rollback, and quantized hidden-state caching to cut inference memory. Langford asks why not cache everything in 16 bits and suggests amplifier-model sensitivity may come from pass 1 not being conditioned for pass 2; overall he finds it great work.
More from Infra
- Fewer than 10 companies can safely move tenants across multi-tenant GPU clusters, argues compute-market skeptic — AccBalanced · 2026-11-22
- Tens of thousands of compute nodes power this multi-agent swarm architecture — hwchase17 · 2026-11-22
- Teknium shares Hermes telemetry: 92.4% of tokens are cache reads — Teknium · 2026-11-22
- Musk demos real-time AI video generation on half of one AI4 TRIP accelerator — jamesdouma · 2026-11-22
- Sentdex asks for ~150 free H100s right after Martin Shkreli posted wanting to buy them — Sentdex · 2026-11-22
- ECB finds $4.5 trillion in hidden AI debt, 287% above 2025 tally — Actual__Wizard · 2026-11-22