AI Brief, November 8: METR on Anthropic, Microsoft kills AI Frontiers, Google buys Crusoe site

METR says Anthropic's models now improve twice as fast as a year ago, Microsoft Research's AI Frontiers group ends, and Google takes over Crusoe's Wyoming site.

The day's lead is METR's report on Anthropic, which The Verge says found the company's models improving roughly twice as quickly as a year ago, with most of the acceleration attributed to AI itself. Elsewhere the news is about where research and compute are going: Microsoft Research's AI Frontiers group has wound down, Google took over a stalled Crusoe data center project, and new reporting questions who is cleared to buy Nvidia chips in China.

METR report: Anthropic's models improving twice as fast as a year ago

The Verge reports on a METR assessment published Thursday, following an 11-week investigation inside Anthropic, that found benchmark scores suggesting "Anthropic has made roughly two years of historical progress over the last year." METR said it is confident the acceleration was not simply from scaling training compute, and that the majority of the speedup as of October "is due to AI," with employees self-reporting being around 12 times as productive using Anthropic's most capable internal model. METR also said Anthropic finished training three unreleased models in the past six months, codenamed Model 2, Model 3 and Model 4, and that Anthropic redacted a majority of the evidence METR found most informative from the ~100-page full report. Cofounder Sam McCandlish said the company made the "right call" on redactions, citing commercial sensitivity and the risk of handing rivals a roadmap, and plans to publish its own report on the pace of progress in the coming weeks.

Why it matters: It is the clearest public data point yet on how much AI-assisted development is compounding inside a frontier lab, which bears on how fast the models you build on will change.

Sources: The Verge AI

Microsoft Research's AI Frontiers group is over, says its departing leader

In a blog post, a founding leader of Microsoft Research's AI Frontiers sub-organization wrote that the roughly 50-person group, created in 2023, has ended and that he and every other founding AIF leadership team member have left. The post says AIF produced the Autogen orchestration framework, the Fara computer use agent, and fundamental work including the Next Latent objective, Dion3 and the Full Bandwidth Transformer, and that the group was reorganized into Microsoft AI in early September. The author describes the ex-Inflection leadership of Microsoft AI as focused on scaling and "aggressively not research," a philosophy he calls incompatible with AIF's work, and says there will be no AIF internships in the coming year.

Why it matters: A first-hand account of a large lab trading fundamental research for scaling, and a signal about where the next round of architecture and agent-orchestration work will come from.

Sources: hunch.net / Hacker News discussion

Crusoe sells unfinished Wyoming data center project to Google

TechCrunch reports Crusoe confirmed Friday it has sold an unfinished southeastern Wyoming data center project to Google, which had earlier declined to become a customer for it and reportedly raised concerns about cost and construction timeline under Crusoe's management. The project was announced last year with Blackstone-backed Tallgrass and was planned to scale to 10 gigawatts of power capacity; Google did not disclose what it paid but announced plans to develop the campus. Crusoe continues to work with Google on another data center near Amarillo, Texas, and is building its Abilene, Texas campus in phases for Oracle and OpenAI plus a separate 900-megawatt campus for Microsoft, with its first building expected online in mid-2027.

Why it matters: Hyperscalers absorbing stalled third-party builds changes who controls new capacity and when it lands, which shapes GPU availability for everyone renting compute.

Sources: TechCrunch AI

The Verge: US approved Nvidia H200 sales to a Chinese firm with an empty office

Citing The New York Times, The Verge reports that New Matrix, a company created 17 months ago, is on the list of Chinese firms approved to handle Nvidia's H200 AI chips despite offices the Times described as "a hollow gray shell with wires dangling from the ceiling where lights should be." The Times found the firm absent from its building's company directory and unknown to the industrial park's property management, and reported that its owners have ties to a Chinese private equity firm blacklisted by the US government since 2024. Per the report this is a pattern rather than an isolated lapse, following the Trump administration's relaxation of Biden-era restrictions after Nvidia agreed to give the government a cut of sales.

Why it matters: Weak vetting invites tighter rules later, and export-control whiplash is a direct input to GPU supply and lead times.

Sources: The Verge AI

VFlash decodes visual plans in one pass, reporting ~25x faster 3D motion forecasting

A new paper introduces VFlash, which replaces token-by-token autoregressive decoding of visual plans (points, bounding boxes, trajectories) with Flash Primitive Decoding: a lightweight drafter predicts a whole primitive's coordinates in one forward pass, with a layout that adapts to the primitive's type, dimensionality and length. The authors train on annotated primitives and accept structurally valid predictions without requiring agreement with an autoregressive model's exact token sequence. They report lower prediction errors than autoregressive decoding across the evaluated 2D and 3D robot trajectory tasks, preliminary single-rollout measurements of roughly 25x faster generation for 3D point motion forecasting, and 70.2% average success across four MolmoSpaces task suites for a world action model conditioned on VFlash plans, versus 62.6% for the baseline.

Why it matters: For robotics and driving stacks, treating spatial output as structured primitives instead of token streams makes richer plans affordable at inference time.

Sources: Hugging Face Daily Papers

Harness-S rescues failed RL rollouts using sibling agent harnesses

Harness-S trains agent policies by building a fixed pool of specialized "sibling" harnesses via Directional Harness Evolution, then invoking that pool only when all initial attempts under the base harness fail, replacing the failed rollout group with fresh trajectories from selected auxiliary harnesses. An Adaptive Harness Router picks Top-k variants using Harness Behavior Profiles and voting across multiple LLMs, with policy-aware profile re-anchoring as the policy learns. The authors report that across AppWorld, SWE-Gym and DBBench the trained policy substantially outperforms policy-only training when returned to the original base harness, and that reusing the router for Rank-1 inference adds gains at the same executor rollout count.

Why it matters: It offers a concrete way to get value from scaffold variants without shipping a different harness in production, since the behaviors land in the policy itself.

Sources: Hugging Face Daily Papers

How this brief is made: each morning, stories are picked from AI lab announcements, analysts, Hacker News, Hugging Face Daily Papers and the tech press, then summarized with AI assistance from the linked sources. Every item links to where its facts came from.