MUD Extends Muon Optimizer to Full Transformer Architecture.

Shared entities / Same source domain / What happened next

Adam isn't gauge-equivariant, so it can't inherit gradient descent's low-rank bias. 43-44% held-out error gap.

Both cover Adam, Muon; reported by the same outlet (arxiv.org); picks up the Adam thread on 2026-08-06.

Shared entity: Muon / Same source domain / What happened next / Downstream implication

Muown implicitly performs angular step-size decay during pretraining.

Both cover Muon; reported by the same outlet (arxiv.org); picks up the Muon thread on 2026-06-23.

Shared entity: Faster / Shared topic / What happened next

Developers Feel 24% Faster with AI. They're Actually 19% Slower.

Both cover Faster; overlapping topics (compute, faster); picks up the Faster thread on 2026-04-27.

Adam uses Anthropic / Same source domain

Vercel's seven-agent "software factory" now writes a third of merged PRs on the AI SDK

Linked by a graph relationship (Adam uses Anthropic); reported by the same outlet (arxiv.org).

Simulator collapse: RL against one LLM user-simulator doesn't transfer.

Linked by a graph relationship (Adam uses Anthropic); reported by the same outlet (arxiv.org).

Static analysis catches malicious agent skills at 0.93 AUC and misses shell-command host destruction 100% of the time.

Linked by a graph relationship (Adam uses Anthropic); reported by the same outlet (arxiv.org).

Personalized agent skills barely beat having no skill at all

Linked by a graph relationship (Adam uses Anthropic); reported by the same outlet (arxiv.org).

Every frontier lab reused one encryption key per model family, so you could decrypt hidden reasoning traces through weaker siblings.

Linked by a graph relationship (Adam uses Anthropic); reported by the same outlet (arxiv.org).