Daily Roundup

Anthropic's $1.5B copyright settlement gets approval

A judge approves Anthropic's $1.5B book settlement as OpenAI retreats on an open-weight argument and NVIDIA ships an edge robotics model.

Top Stories

Anthropic’s $1.5B book settlement wins final approval

Judge Araceli Martinez-Olguin gave final approval to Anthropic’s $1.5 billion settlement with authors and publishers. The agreement provides $3,000 per work across an estimated 500,000 works, and is described as the largest settlement in the history of US copyright law. TechCrunch reported the final approval and settlement terms.

The case split the training question from the acquisition question. Judge William Alsup ruled that training on copyrighted text counted as fair use, while Anthropic’s downloading and storage of millions of books from pirate sites was illegal. Anthropic settled before the piracy issue went to trial, so the case will not reach an appeals court and become binding precedent. The ruling and settlement history are summarized here.

The settlement closes this case but not the wider fight over AI training data. Lawsuits involving Google, Meta, Midjourney, and OpenAI remain pending, according to TechCrunch’s account.

OpenAI official retracts open-weight policy argument

Dean W. Ball, OpenAI’s head of strategic futures, argued that the US government should find a pretext to create regulatory fear, uncertainty, and distrust around Chinese open-weight models. His argument focused on Moonshot’s Kimi K3 and claimed open-weight releases could deter frontier labs from spending on new models. TechCrunch documented Ball’s argument and the response. The dispute is one more wrinkle in the broader open-weight race we have been tracking since our Week 10 open-weight showdown.

Yann LeCun and venture capitalist Martin Casado pushed back. Ball then retracted the claims that a regulatory crackdown was the White House’s best strategy and that open-weight models necessarily slow technical progress. The retractions are included in TechCrunch’s report.

The episode shows a policy dispute around a specific foreign model, not a published OpenAI corporate policy. The confirmed event is narrower: a public argument from an OpenAI policy official was withdrawn after criticism. The original report identifies Ball’s role and the claims he retracted.

US AI standards director resigns after three months

Chris Fall resigned as director of the Center for AI Standards and Innovation about three months after his appointment. The agency confirmed the departure to multiple outlets, but no reason was given. TechCrunch reported the resignation.

Fall followed Collin Burns, who left the job in less than a week. CAISI operates under the National Institute of Standards and Technology, and its recent work includes reports on the Chinese open-weight models GLM-5.2 and DeepSeek V4 Pro. The report details the leadership changes and CAISI’s recent work.

A second short tenure leaves another leadership question at the federal office responsible for AI standards work. The available reporting does not explain Fall’s departure or identify a successor. TechCrunch says no reason was provided.

Google’s reported Gemini chip targets lower power use

Alphabet is reportedly designing a server chip called Frozen v2 to run Gemini models more efficiently, with a release targeted for sometime in 2028. Anonymous sources told The Information that the design could generate six to ten times more tokens per unit of power than Google’s current AI chips. TechCrunch summarized the report and its anonymous sourcing.

Google did not confirm the chip or the efficiency estimate. Its statement said teams continually test new ideas and that not every project reaches production. The reported performance and schedule should therefore be treated as an unconfirmed roadmap, not a product announcement. Read Google’s response and the reported target.

NVIDIA releases a 4B edge model for robots

NVIDIA released Cosmos 3 Edge, a four-billion-parameter open world model for robots and vision agents. The company says the model can interpret surroundings, reason in real time, and generate robot actions on edge devices. It lists support for Jetson systems, GeForce RTX GPUs, RTX PRO GPUs, and DGX systems. NVIDIA’s Hugging Face post describes the model and supported hardware.

At a 640 by 360 observation resolution, NVIDIA says Cosmos 3 Edge generates 32 actions per inference at 15 Hz on Jetson Thor. The release includes a base model, post-trained checkpoints, and a policy checkpoint trained on the DROID dataset for pick-and-place tasks. The release post gives the output rate and checkpoint details.

NVIDIA also claims the model ranks first on VANTAGE-Bench among models of a similar four-billion-parameter size. That result comes from the vendor’s announcement and should be read as a vendor benchmark until independent testing appears. See NVIDIA’s stated benchmark scope.

Hiring simulation finds stronger bias in reasoning models

Researchers from Princeton University and the University of Chicago tested ChatGPT, Claude, Gemini, OpenAI’s o3, and DeepSeek’s R1 in a simulated hiring exercise. Each model filled 20 jobs from four fictional ethnic groups over 40 rounds, with every candidate equally likely to succeed in any role. MIT Technology Review describes the ICML study and its design.

On a segregation scale where 2 represented the maximum, human participants scored 0.84. The tested models scored roughly 65 percent higher, while o3 reached 1.83. Newer reasoning models, including o3 and DeepSeek R1, showed the strongest bias in the experiment - the same family of models we previously found easy to jailbreak through their chain-of-thought. The article reports the scale and results.

Telling the models to be fair did little, while offering a bonus for diverse hiring significantly reduced bias. The researchers also cautioned that the simulation gave models immediate feedback, unlike many real hiring systems, so the study does not establish how the models behave in every real-world screening process. MIT Technology Review includes both the intervention results and the caveat.

California privacy tool sends one request to 614 brokers

California’s Delete Request and Opt-out Platform, known as DROP, lets a resident submit one request to 614 registered data brokers. The platform launched January 1, and brokers have 45 days to address requests after their August 1 compliance deadline. EFF explains the schedule and broker count.

EFF says the requests can cover Social Security numbers, precise geolocation, browsing history, email addresses, phone numbers, and inferred data. Public-record information, such as vehicle or real-estate ownership, is not deleted through the system. EFF lists the covered data and exclusions.

DROP is limited to California residents, and deletion does not stop a broker from collecting new data later. New brokers can also register after a request is filed, while companies that are not registered brokers are outside the platform. EFF details those limits.

YouTube tightens monetization rules for mass-produced videos

YouTube rolled out clarified monetization rules on July 16 for every member of the YouTube Partner Program. The update concerns eligibility for advertising revenue rather than a blanket removal policy for synthetic media. TechCrunch explains the scope and rollout.

The platform now describes three non-monetizable categories: generic or repetitive template-based videos, off-putting or distressing videos made to chase views, and AI representations of real people discussing sensitive subjects such as finance, law, or healthcare. The three categories are listed in TechCrunch’s policy summary.

The distressing-content rule applies whether or not AI made the video. That distinction matters because YouTube is targeting repetitive production and manipulation alongside AI personas, rather than treating all synthetic content as one category. TechCrunch reports that the distressing-content restriction is not limited to AI output.

Quick Hits

  • MCP session handling: A planned Model Context Protocol update will move servers toward a stateless approach to session IDs, which is intended to reduce scaling problems behind load balancers. The protocol is the same one Cisco flagged as insecure in February. TechCrunch explains the change.
  • New York crawler bill: EFF says the New York Stealth Crawler Protection Act awaits Governor Kathy Hochul’s signature and could let news sites seek orders to identify anonymous crawlers. Read EFF’s argument against the bill.
  • Open-web research: EFF points to investigative reporting, academic research, cybersecurity, and its Privacy Badger project as legitimate uses for anonymous crawling. EFF gives examples of those uses.

Worth Watching