A judge blocks the Pentagon's Anthropic blacklist

A federal judge ruled that the Pentagon's supply-chain-risk designation was unlawful. "The empty invocation of national security is not a blank check to punish and retaliate against government critics," Judge Rita Lin wrote. The government is expected to appeal, and Anthropic is contesting a second designation in the D.C. Circuit.

Anthropic proposes a Model Hardware Standard

Anthropic

Anthropic says the Model Hardware Standard gives Claude a unified interface to lab equipment. Its demonstrations show Claude configuring a laser-scanning microscope, generating a live-sample tracking script, and operating within predefined physical limits.

Gemini Co-Scientist moves from hypotheses to execution

arXiv · Google Research authors

The authors report a Co-Scientist system connected to a semi-automated chemical vapor deposition reactor, tailored growth recipes to lab constraints, and matched unpublished wet-lab measurements in a biology task. Their MXene-like material still needs atomic-structure confirmation.

Changing the harness changes coding-agent results

arXiv · Sydney Lewis

With model weights and tasks held fixed, a treatment that shortened older tool output and responded to stalled work raised complete SWE-bench Verified solutions from 43 to 72 under tight context. The paper argues that evaluations need to identify the model and harness together.

SKILL.state replaces transcripts with mutable execution state

arXiv · Sanket Badhe, Priyanka Tiwari, and Jonghyun Chung

The runtime sends the immutable skill specification, current structured state, and latest observation, then discards intermediate reasoning after a validated state update. Across several datasets, models, and environments, the authors report higher accuracy and lower cumulative token use.

PILOT adds live supervisor steering

arXiv · Yang Xiao and collaborators

A separate supervisor can redirect or abort a running worker and turn lessons from the run into reusable skills and memory. Across two frozen backbones, the authors report fewer output tokens in the self-improvement tests. Terminal-Bench 2.0 scores rose by up to 9.8 percentage points.

Nvidia's chip profits finance its next customers

Axios

Following yesterday's reported Hugging Face deal, Axios puts Nvidia's wider reach at more than $750 billion of investments, financing deals, and partnerships, citing PitchBook. Nvidia now supplies, finances, partners with, and competes against companies across the same AI ecosystem.

EU AI Act transparency rules reach users

Axios

The August 2 deadline put chatbot and AI-content disclosures into effect. Anthropic plans text watermarking and a detection API; Google and Meta are developing transparency tools, OpenAI is publishing training-data summaries and provenance signals, and Microsoft has changed governance and risk management.

South Korea funds premium domestic AI as a public utility

The Wall Street Journal via Techmeme

South Korea plans to work with KT, SK Telecom, and Kakao to provide premium AI tools to the public at no charge, including unlimited tokens. The program is intended to steer use toward domestic chatbots and reduce reliance on US or Chinese services.

Tencent enters open models at 770 billion parameters

Bloomberg via Techmeme

Hy4 Preview is a 770-billion-parameter open model with a one-million-token context window. Tencent says it outperforms models from Z.AI and Moonshot in internal tests; the cited report includes no independent benchmark results.