Media Zone | 2026-07-15

Meta's physics-olympiad flex and New York's data-center freeze on the same day. The capability-vs-permission gap was the story.

Today's signal

• Dominant story: Meta AI scored 30/30 on Asian Physics Olympiad, tying top 3 humans (model undisclosed).
• Pattern: capability soaring while infrastructure permission tightens (NY data-center freeze).
• Counter-signal: eliebakouch and others asked which Meta model, no answer given.
• Quiet area: empty Reddit, no research-paper social debate.

Routing, KV cache, compression, GPU

Open models and deployment

• NVIDIA Nemotron: open models enterprises and nations "can trust, control, customize."
• Kilo Sandbox: auto-approve everything, the OS decides what agents touch, not the model.

LLMs, agents, safety

Meta's physics olympiad score

• 30/30 on Asian Physics Olympiad theoretical exam, tying top 3 student contestants.
• Meta did not disclose the model (Muse Spark 1.1 or internal).
• eliebakouch flagged the disclosure gap; caps what can be concluded.

Grok 4.5 coding

• Ranks #2 on FrontierSWE, behind only Fable 5, ahead of Opus 4.8 / GPT-5.5 / GLM-5.2.
• Second coding data point in a week; harness still undisclosed.

Industry and business

• New York froze new AI data centers statewide, sharpest US infrastructure-side move.
• DOGE built an AI to gut federal housing rules.
• Bill Yuchen Lin: RL on long-horizon SWE tasks showing strong cross-benchmark generalization.