AI2's Ricursive blog looks AI-written and misnames its own model, sparking criticism

code_star · x · 2026-11-16

@codestar flags a "wtf" paragraph in AI2's blog for its Ripple/Ricursive model: a model apparently read the release blog without web search and made wild assumptions about AI2 and model training, while the blog itself misspells Ricursive as Recursive throughout. The blog claims up to 16x compute efficiency gains at 4B scale via a series of "model laddering" experiments (StepFun scaling laws, pre+post norm, sigmoid gated attention, 2-stream mHC rejected at 4B, global NoPE, value residuals). The author argues laddering is not a novel invention — it's how all models have always been developed.

Original post →

More from Fun

Fun channel →