Anthropic says it is expanding its AI for Science program to 20 scientific foundations, institutions and research groups, including the US National Science Foundation. Anthropic says partners get free Claude credits to analyze data, design experiments and build shared tools, aiming to produce open datasets and software.
The Department of Justice is questioning Nvidia about how its $12.9 billion acquisition of Hugging Face, an open-source platform for hosting and testing AI models, could affect competitors. Hugging Face sells infrastructure and collaboration tools, including to chip companies. Nvidia, which announced the deal in September, declined to comment.
DoorDash reports only several hundred orders from third-party AI agents such as Muse, Instinct and Fo, CEO Tony Xu said on the company's third-quarter earnings call, against 3 million daily deliveries. Xu said ordering remains overwhelmingly in DoorDash's own app, though the company sees future advertising potential inside agents.
Google's Suncatcher prototype space data center is online after launching on a SpaceX Falcon 9. It carries four Tensor Processing Units but can run one for about 20 minutes at a time due to power limits; engineers prompted DeepMind's Gemma 4 to write a space haiku. Google says it operates nominally, with two satellites planned for 2027.
Runware plans a factory to assemble its Sonic Inference Pods on an assembly line, CEO Flaviu Radulescu says, aiming for 30 megawatts of capacity per day by 2028 as operators seek faster AI deployment. The company says it has two factories and targets 48-hour installation. It has raised over $60 million and plans a Series B.
Meta has moved its LLM-generated Catching JiTTests from research prototype into production infrastructure, generating failing tests that expose bugs in code diffs before they land. Meta says 589 engineers engaged with the system at a 60% true-positive rate, and it caught 74 serious privacy bugs in the first half of 2026, cutting privacy SEVs by 20%.
Factory released Droid 1.0, an agent harness that now runs tool calls through sandboxed JavaScript programs, cutting model requests on long tasks. Factory says Droid scores 57.6% on Terminal-Bench 4 with GPT-6.1 Sol, against 54.0% for Codex, at 12% lower cost. The company claims up to 20% more efficient use than models' native harnesses.
Google DeepMind researchers Alex Imas, Andrew Koh and Shane Legg argue that AI's economic impact hinges on coordination, not solo competence. Individual leaderboards miss the shift, they say, mapping four regimes from minimal AGI to ASI. They cite OpenAI's Noam Brown, who said coordinating agent swarms productively is very difficult.
Raindrop researchers report that only 5% of 2,160 simulated outage trials produced a leader whose instructions changed teammates' decisions. Agents running Claude Sonnet 5.5 or GPT-6.1 Sol were about 1.7 times as likely to lead as Gemini 3.1 Pro. Naming an agent first in the system prompt tripled its odds of issuing a plan.
Researcher vigji reconstructs Meta's Muse Spark Open charter by piecing together fragments the model quotes in its chain of thought, drawing on 884 transcripts from 260 prompts via OpenRouter. The reconstructed text lists corrigibility and duties to humanity as hard constraints and casts Muse as owing users fiduciary duties. Meta says it will publish the specification.