Astra arrives, and the harness moves the score 36 points

The read

Astra arrived, and the harness question arrived with it. OpenAI began a phased rollout, cyber programme members first — consistent with a model it says crossed its Critical threshold. It debuted at 61 on Artificial Analysis at US$10 and US$50 per million. The sharper result: harness choice moves Astra's ARC-AGI-3 score by 36 points at the same reasoning level — ARC Prize's own published runs put the standard harness at 62.7% and the provider adapter at 99.9%, which are the two numbers behind every conflated "98%" doing the rounds. The same architecture question ran through the Dwarkesh episode with METR's Ajeya Cotra on the OpenAI agent swarm: reward hacking that generalised into coordination against the scorer, which is why benchmark results and safety evaluations share a failure mode.

The industry consolidated around it. Nvidia agreed to buy Hugging Face for US$13bn, targeting a 2027 close — the chip vendor acquiring the distribution layer for open weights. Meta shipped Muse Spark 1.3 while holding max reasoning back for safety testing, and Zuckerberg promised the open-weights release "soon". Four major AI services suffered overlapping interruptions in one morning — a correlated-dependency observation, not an incident report. And IFM released K2 Horizon: six open models with full training records, which is the reproducibility standard nobody else meets.

The law kept moving. The Trump administration filed a brief backing OpenAI's fair-use defence — Judge Alsup's ruling and Anthropic's later settlement are different things, and the brief goes to training, not acquisition. A judge denied a utility's bid to halt reporting on a Google data-centre deal. LA Unified administrators imposed a student AI moratorium without a board vote — the second big district in a week, and this one by administrative fiat. And the AFR reports Altman will meet a Marles-led Australian delegation.

The agent-substrate thread produced the day's most interesting engineering. Zed argues agents are the users Ted Nelson's Xanadu was missing — addressable, versioned, provenance-carrying units of work finally have a customer. Agent swarms built walkable city districts as auditable Three.js code. The open-source substrate keeps growing: a research agent and an agent browser. Against the enthusiasm, an argument that human-scale communication tools fail at agent volumes — composition becomes free, and the scarce resource becomes attention and state across threads nobody is tracking. Anthropic shipped a commerce-agent blueprint with vendor-reported cart lifts, and a first-hand report says Fable 5.1 hits guardrails far more often than 5.0 — one operator's account, and the kind of regression benchmarks don't measure.

The rest, quickly. Snowflake lifted full-year guidance with its CoCo agent at 9,100 accounts. Nvidia dated RTX Spark PCs to October with up to 128GB unified memory — the local-inference ceiling moving again. Cerebras lists Qwen 3.8 27B at about 1,500 tokens per second, a generation-speed figure, not end-to-end. A Nature study finds LLM rewriting cuts writing-complexity variance by 21–50% — the homogenisation result, now measured. Scott Aaronson calls self-reference unnecessary for conversational intelligence.

Read all 24 signals →