NVIDIA Cosmos 3 Edge: A 4B World Model for Edge Robotics

NVIDIA releases a 4-billion-parameter open world model that runs on edge devices — Jetson Thor, RTX GPUs — delivering real-time reasoning and 32 actions per inference at 15 Hz. It ranks #1 on VANTAGE-Bench for vision analytics and sets state-of-the-art for robot policy learning in its size class.

Impact: Medium HuggingFace

Learn more →

Patch Policy: Efficient Embodied Control via Dense Visual Representations

A collaboration including Yann LeCun shows that pretrained dense ViT features — long underutilized in robot learning — can drive efficient manipulation policies without compressing observations into a single token. The approach preserves fine-grained spatial detail and outperforms scratch-trained vision backbones.

Impact: Medium arXiv

Learn more →

Automated Discovery Has No Universally Superior Harness

A sweeping evaluation of autonomous discovery systems like OpenEvolve and TTT-Discover finds that no single configuration of archives, parent selection, and budget allocation dominates. The composite recipes sold as general-purpose harnesses are just one design choice among many — and swapping any component changes the outcome.

Impact: Low arXiv

Learn more →

It’s Not What You Say, It’s How You Say It: LLMs and Expressions of Belief

When users tell an LLM what they believe, should the model accept it as true or push back? This study shows that how a belief is framed — presupposition, assertion, or hedging — dramatically changes whether models comply or resist, revealing a blind spot in alignment evaluations that only test direct statements.

Impact: Low arXiv

Learn more →

 

MCP Goes Stateless: AI’s Key Interoperability Protocol Levels Up

The Model Context Protocol — the plumbing that lets AI models reach into calendars, databases, and internal tools — is dropping its stateful session-ID model next week in favor of a stateless design. The change lets servers scale behind load balancers without sticky sessions, removing a major deployment headache for the millions of MCP servers now in production.

Impact: High TechCrunch

Learn more →

Reverse-Engineering Is Cheap Now

Simon Willison observes a quiet revolution: coding agents have dropped the cost of reverse-engineering home-device APIs to near zero, changing the ROI calculus. When a failed integration costs minutes instead of days, throwaway automation becomes rational — and the psychological barrier to maintaining brittle, undocumented code vanishes.

Impact: Medium Simon Willison

Learn more →

Who’s Afraid of Chinese Models? Ben Thompson’s Distillation Proposal

Ben Thompson argues the US should legalize training data collection as fair use AND bar companies from forbidding distillation via terms of service — leveling the playing field between American labs that trained on copyrighted data and Chinese labs that distill American models. The piece also links Alibaba’s Qwen 3.8 Max open-weights release to Xi Jinping’s July 18 speech urging open-source AI collaboration.

Impact: High Simon Willison

Learn more →

 

Anthropic’s $1.5B Copyright Settlement Approved — Largest in US History

A federal judge gave final approval to Anthropic’s $1.5 billion class-action settlement, paying $3,000 per work across an estimated 500,000 books. The judge ruled training on copyrighted text is fair use — a landmark for the AI industry — but penalized Anthropic for downloading books from pirate sites. Many authors don’t consider it a win: the practice was legalized; the piracy was fined.

Impact: High TechCrunch

Learn more →

Trump’s AI Czar Resigns — Third Departure in Six Months

Chris Fall, director of the Center for AI Standards and Innovation (CAISI), has resigned after just three months on the job. He follows Collin Burns (pushed out in April over Anthropic ties) and David Sacks (stepped down in March). CAISI — the primary US agency for AI technical standards and cybersecurity risk assessment — has now had three leaders in half a year.

Impact: High TechCrunch

Learn more →

Trump’s AI World at War: Advisors Trade Insults Over Kimi K3

MIT Tech Review reports that Kimi K3 has split Trump’s AI orbit into factions. David Sacks called Anthropic’s models “lobotomized” and “woke”; a Pentagon official called OpenAI’s head of strategic futures a “supreme village idiot.” OpenAI’s Dean Ball argued the US government should manufacture regulatory FUD around open-weight models. The infighting underscores a real economic threat: free Chinese models erode the case for paying American labs.

Impact: High MIT Tech Review

Learn more →

Google’s ‘Frozen v2’ Chip Targets 6–10x Gemini Efficiency Gains

Alphabet is designing a new server chip internally dubbed “Frozen v2,” targeting 6–10x efficiency improvements over existing TPUs by tokens-per-watt. Slated for 2028, the chip underscores the industry’s push to reduce dependence on Nvidia and address the compute-cost anxiety now dampening AI stock euphoria. Google neither confirmed nor denied the report.

Impact: Medium TechCrunch

Learn more →

AI Invents Hiring Biases Humans Don’t Have

Princeton and University of Chicago researchers ran ChatGPT, Claude, and Gemini through a simulated hiring game with fictional ethnic groups. All candidates were equally likely to succeed — yet the models segregated groups into job niches 65% more aggressively than human participants, with OpenAI’s o3 scoring near the maximum on the segregation scale. LLMs don’t just inherit stereotypes from training data; they manufacture new ones from experience.

Impact: Medium MIT Tech Review

Learn more →

The governance vacuum is the story. In a single week, the US government’s AI safety agency lost its third director in six months — CAISI can’t keep a leader, while Trump’s former AI advisors trade “lobotomized” and “supreme village idiot” insults on X over whether to ban a free Chinese model. Meanwhile a federal judge approved the largest copyright settlement in American history — $1.5 billion to authors — while simultaneously ruling that the practice being compensated for is perfectly legal. The authors get $3,000 per book; the AI industry gets a permanent green light. And as AI models invent hiring stereotypes humans don’t have, the institutions meant to measure and mitigate these harms are the ones collapsing. The technology is accelerating; the guardrails are resigning.