Wednesday's window is dominated by the widening gap between AI ambition and AI operations: Meta's CTO publicly called its AI reorg 'atrocious' as token costs balloon, while separate research shows senior engineers now spend more time cleaning up AI-generated code than writing their own, and IT teams admit they don't actually know who owns most of their AI agents. On the model side, Sakana shipped an autonomous research agent called Marlin, Google DeepMind published a long paper mapping a path to ASI, and LMSYS detailed next-gen speculative decoding (DFlash/Spec V2) that meaningfully speeds up inference. Anthropic disabled its Fable and Mythos models globally under a US export-control directive, a move Stratechery frames as turning safety compliance into a competitive moat. Meanwhile Factory 2.0 and Facebook's new AI Mode show the product layer racing ahead of the operational maturity underneath it, and infra pieces on GPU longevity and inference engineering are must-reads for anyone sizing AI capacity.

• Meta's CTO admitted the company's AI reorg was 'atrocious' amid runaway token spend.
• Anthropic disabled Fable and Mythos globally under a US export-control order.
• Sakana released Marlin, an autonomous AI research agent.
• LMSYS's DFlash/Spec V2 pushes speculative decoding to a new generation for faster inference.
• Senior engineers now spend more time fixing AI-generated code than shipping their own.