You Jiacheng(@YouJiacheng):> The table was already significantly larger than the model, so the bigram table was already far too large to ever scale to frontier training. 5× active embedding is somewhat acceptable (ds v4.1 flash use ~12×), but 500× is next level. we need new infra.

2026.09.29 06:04

> The table was already significantly larger than the model, so the bigram table was already far too large to ever scale to frontier training. 5× active embedding is somewhat acceptable (ds v4.1 flash use ~12×), but 500× is next level. we need new infra.