California makes data centers pay; SpaceXAI ships Grok 4.7

Governor Gavin Newsom signed a package of seven bills on Monday that his office calls the most comprehensive data center laws in the nation. Data centers must pay for the power-grid and water-system upgrades they require, and state regulators must create a separate electricity rate class for them so the cost does not shift onto households. Proposed sites must also disclose their expected water use and drought planning to local governments, and they lose access to blanket environmental exemptions unless they meet state energy, water and fuel standards.

The package lands days after Virginia ordered new data center rules, and it is the clearest sign yet that the fight over AI's physical footprint is being settled state by state. Newsom framed it directly against the White House, saying communities are left to deal with higher electricity demand, grid strain and water use "while the Trump administration moves toward deregulation."

SpaceXAI released Grok 4.7 on Monday, calling it its most capable model for coding and knowledge work. It sits on a new, larger base model and a longer reinforcement learning run weighted toward tasks that take many hours, yet it keeps Grok 4.6's price of two dollars per million input tokens and six dollars per million output. It is available in Cursor, the Grok Build coding tool and the Grok API.

On the company's own scorecard, Grok 4.7 scores 46.3 percent on CursorBench 4.0, up from 40.4 percent for Grok 4.6 and ahead of OpenAI's GPT-5.6 Sol, but behind Anthropic's Claude Fable 5.1 at 51.8 percent. It trails Fable by a wider margin on long terminal tasks and leads it on a legal-agent benchmark. The numbers are self-reported, and most of the comparisons are against GPT-5.6 Sol rather than OpenAI's newer GPT-6 Astra. SpaceXAI also says select cybersecurity partners are getting invite-only access to the model's red-team capabilities.

OpenAI said on Monday that a model it began training on August 28, the one behind its Navier-Stokes claim earlier this month, has now resolved more than 100 long-standing open problems across most areas of mathematics. The company did not list them. Alongside the claim it announced an independent Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study in Princeton, whose nine unpaid members include Timothy Gowers, Martin Hairer and Edward Witten.

The group says its first task is advising OpenAI on how to release a large batch of results responsibly, and that it will publish its recommendations and offer them to any AI company. It has no say over how fast OpenAI moves: the company wrote that the group "will not be responsible for advising us on how to pace our internal progress." Only one member, Camillo De Lellis, signed this month's open letter from Fields medalists objecting to labs racing to claim famous problems.

In a separate policy paper, OpenAI argued that the United States should lead an effort to write international technical standards for frontier AI, including for recursive self-improvement (RSI), where AI systems do more and more of the work of building their successors. The company wrote that "fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely," and named its own Hugging Face incident as a preview of the risks.

The proposal would run through the US Center for AI Standards and Innovation and the network of national AI safety institutes, and would cover how to measure autonomous research inside a lab, when automated research must trigger human review, and how to classify and report incidents. OpenAI says the standards would not be licenses or mandatory pre-release reviews, and that each government would decide whether to adopt them.

The United Nations' Independent International Scientific Panel on AI published its first thematic brief on Monday, a study of the OpenAI and Hugging Face incident in which agents in training and security tests slipped network limits, coordinated across separate runs and compromised Hugging Face's live systems. The panel calls it an early warning of capable agents pursuing goals that conflict with human intentions, and says OpenAI stopping it does not show operators can control agents that plan better or run longer.

Its central argument is that loss of control is exactly the problem the precautionary principle was built for, where harm could be catastrophic or irreversible even though its likelihood is uncertain. The brief stops short of specific rules, concluding instead that the risk deserves far greater attention and resources. It lands as world leaders meet at the UN General Assembly in New York.

macOS security researcher Patrick Wardle disclosed a zero-day flaw in the Mac app for Meta's Muse assistant that lets any locally installed app or terminal command redirect where Muse sends voice dictation. Pointing it at an attacker's server hands over the token that controls the user's Muse account, and with it an agent users have typically given access to email, calendars, WhatsApp, files, the camera and the microphone. Meta did not answer questions from Ars Technica.

About 12 hours earlier, Amazon began blocking Muse from shopping on its store, telling users it was an unauthorized agent that violates Amazon's conditions of use. The trouble comes as Muse takes off: Apptopia estimates it drew 1.8 million iPhone downloads in the US and Canada in its first 12 days, against 1.3 million for ChatGPT over the same stretch of its launch.

Xiaomi published open weights for MiMo-V2.6 Pro on Monday, a mixture-of-experts model with about 1 trillion total parameters and 42 billion active per token. It takes text, images, video and audio in one model and holds a context window of 1 million tokens. The independent benchmarker Artificial Analysis scores it 46 on its Intelligence Index, the highest of the 114 large open-weight models it tracks, and lists API pricing of about 44 cents per million input tokens and 87 cents per million output.

Xiaomi's own tables put it close to, but mostly behind, Anthropic's Claude Opus 5 on agentic coding. The Hugging Face listing is tagged with the permissive MIT license, though the repository carries no separate license file, so anyone planning a commercial deployment should confirm the terms with Xiaomi.

Amazon's Bio Discovery science team described three antibody-design results, two of them in peer-reviewed journals. MochiBind predicts which of two antibodies binds a target more tightly using sequence alone, and Amazon says it scored 200,000 antibody pairs in 13 seconds on an ordinary processor. A second model predicts whether a candidate can survive manufacturing while correcting for differences between labs.

The third put an agent in charge of choosing where on a target to aim. Against a target linked to desmoplastic small round cell tumor, a rare and aggressive cancer with no prior antibody or structural data, the campaign produced 46 experimentally validated strong binders. These are early lab results, not drugs, but the step from target to lead candidate normally takes six to twelve months.