Speculation: 2T Training Run Points to V4.1 Pro, as Critics Lament Scaling Is Still Alchemy

teortaxesTex · x · 2026-09-21

teortaxesTex speculates (unconfirmed) that a reported 2T-parameter training run corresponds to V4.1 Pro, with roughly 540GB of engram memory if it follows V4.1's design — questioning whether engram memory saturates or turns fragile at scale. He expects roughly Fable 5-level performance and laments that scaling remains alchemical: Meta ran 405B active params on 16K H100s two years ago, yet rushing this scale today can still waste months of compute.

Original post →

More from Infra

Infra channel →