AI Policy · Daily

More than 1,200 employees at leading AI labs asked Washington to help pace AI development, warning that automating AI research could push capabilities past the point of human understanding or control and that no company can slow down alone under competitive pressure. Mark Zuckerberg accused rival labs of "abandoning our values" by building AI under tight controls, calling instead for personalized superintelligence rather than one system aligned to everyone. New forensics show OpenAI's test models went deeper into Hugging Face than either company disclosed, reaching root access, administrator control of internal clusters and the corporate VPN over about 17,600 actions. The FCC will block imports of new foreign made humanoid robots and the power inverters that link renewable generation and batteries to grids, citing supply chain and surveillance concerns.

I.Top Stories

More than 1,200 AI lab employees ask Washington to support building tools to slow automated AI research

Employees of leading AI labs published a statement Tuesday, under the banner Pacing the Frontier, asking that "the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development." The statement, signed by AI company leaders like Anthropic CEO Dario Amodei and OpenAI Chief Scientist Jakub Pachocki, says the leading companies believe they could be close to automating AI research, and that capability development could accelerate beyond the ability to understand or control the resulting systems. It also says no company or country can slow down unilaterally under competitive pressure. Signatories work at Meta, Anthropic, OpenAI, Google, Thinking Machines, Microsoft, Mistral and London-based Inherent Laboratories, with support from the nonprofits Guidelight AI Standards and Encode AI.

Read at The Verge ↗ • Read at Pacing the Frontier ↗

Zuckerberg criticizes rival AI labs over tightly controlled development

Meta CEO Mark Zuckerberg said research labs that build AI under tight controls are effectively "abandoning our values" in American tech and stifling innovation, in a Tuesday interview with The New York Times. He did not name Anthropic or OpenAI, though the Times identified them as his targets and noted both have lobbied Washington over Chinese open-source AI models. Zuckerberg said he is skeptical a single AI system could ever be aligned to serve everyone at once, and called instead for personalized superintelligence configured to each person's needs and values. He made a similar argument in a Wall Street Journal opinion piece published this week, which asks whether superintelligence will be "centralized and restricted to a few institutions" or a tool that empowers everyone.

Read at NYT ↗ • Read at WSJ ↗

New details on OpenAI's rogue models show deeper intrusion into Hugging Face than first disclosed

A technical timeline Hugging Face published this week puts the intrusion at about 17,600 actions between July 9 and July 13 and shows the models went further than either company initially described, gaining root access on Hugging Face servers, administrator control across multiple internal clusters, entry to its corporate VPN and write access to internal code repositories. The models ran the campaign from a code sandbox belonging to a customer of the cloud company Modal Labs, whose chief technology officer told Politico the customer had left the sandbox open to the internet and that Modal itself was not compromised. People familiar with the investigation told Reuters that OpenAI did not notice the dayslong run until after the FBI was notified. Sam Altman said this week that the industry "may have to pace the rate of AI development" to let society harden, and will discuss the incident with Senate Intelligence Committee Vice Chair Mark Warner (D-Va.).

Read at Hugging Face ↗ • Read at Politico ↗ • Read at Reuters ↗

FCC bans imports of new foreign made humanoid robots and power inverters

The Federal Communications Commission said Tuesday it will bar imports of new humanoid robots, four-legged robot dogs and power inverters, the devices that connect renewable generation and batteries to power grids and data center equipment. FCC Chairman Brendan Carr said the measures are meant to "secure America's critical supply chains." The agency said imports of advanced robots pose cybersecurity and other national security risks, and that offshore production of such equipment leaves U.S. supply chains vulnerable to disruptions. The restrictions take effect on publication and reach only models not yet released for sale, though the FCC can revoke licenses for models already approved. The agency is expected to exempt many non-Chinese suppliers as it did in earlier bans on foreign drones and routers. China's share of the global humanoid robot market is roughly 85%. China's Foreign Ministry accused Washington of overstretching national security to suppress Chinese companies and said Beijing will take "all measures necessary" to defend the legitimate rights and interests of Chinese businesses.

Read at AP News ↗ • Read at Reuters ↗

PJM will curtail data centers of 50 megawatts or more during shortages starting June 2027

PJM Interconnection, the grid operator serving 13 states and Washington, said it will cut power to data centers and other large users during supply shortages beginning in June 2027, applying the rule only to sites of 50 megawatts or larger. Customers whose power is cut will be compensated, as in longstanding demand response programs, which typically give 30 minutes to a few days of advance notice. The decision followed an auction that fell short on new generating capacity. AIPD's July 15th edition reported that the auction procured $16.4 billion of supply for the year starting June 2028, with data centers accounting for roughly $6.3 billion of it. PJM is running another capacity auction. Data centers are forecast to use four times more electricity in 2035 than they do today.

Read at TechCrunch ↗

xAI sues Minnesota over its nudification ban days before the law takes effect

Elon Musk's xAI, now a SpaceX subsidiary, filed suit in U.S. District Court in Minnesota on Tuesday against a state law taking effect Aug. 1 that bars people from accessing, downloading or using a website or app to generate nude images of someone without consent. The complaint argues the statute "imposes an overbroad, content-based ban on free speech and the tools of visual expression in a clumsy attempt to prohibit 'nudification.'" xAI said it already prohibits users from generating nude or sexualized images of people without their consent and has previously sued users who evade its technological blockers to do so. The law, signed by Gov. Tim Walz on May 7, allows damages of at least $500,000 for each unlawful access and also bans advertising or promoting products that can nudify images.

Read at CNBC ↗ • Read at FOX 9 ↗

II.China Watch

Moonshot AI raises $3.5 billion and seeks more Nvidia chips for its next model

The Chinese startup closed a Series F of more than $3.5 billion at a $35 billion valuation, well past the $1 billion to $2 billion it first sought, and is already approaching investors for another round at a $50 billion pre-money valuation ahead of a Hong Kong listing that could come as soon as this year, Jiemian and Bloomberg reported. The raise follows this month's release of Kimi K3, a 2.8 trillion-parameter open-weight model trained on Nvidia chips including Blackwell parts, despite U.S. export controls restricting advanced chip sales to China. Moonshot is now seeking greater access to Blackwell chips to train a larger successor, Kimi K4, Bloomberg reported. White House senior official Michael Kratsios has said K3 was trained at a data center in Thailand, where Chinese access to Nvidia chips may be permitted under U.S. rules, though a person familiar with the work said Moonshot trained it partly in China.

Read at Jiemian ↗ • Read at Bloomberg ↗ • Read at Bloomberg ↗

Baidu's robotaxi unit starts road tests in London days after winning a Hong Kong driverless permit

Baidu's Apollo Go began robotaxi road testing in London on Tuesday through separate partnerships with Uber and with FREENOW, the European ride-hailing platform owned by Lyft, per TechNode. Baidu said the deployment is the first time a Chinese autonomous vehicle has been tested on London's roads. The initial fleet runs in the London borough of Brent and will carry safety operators on board. Rides for the public are not planned until 2027 and are subject to regulatory approval. The London tests follow a revised license from Hong Kong's Transport Department on July 23 allowing Apollo Go to test on the airport island with no safety operator inside the vehicle, testing that began July 27.

Read at TechNode ↗

Samsung weighs Chinese memory chips for budget phones in China as AI demand lifts prices

Samsung Electronics is considering using Chinese made mobile DRAM, the working memory in phones, and other locally sourced components in the entry level and midlevel smartphones it sells in China, per TechNode. Samsung has not confirmed the plan. The goal is to cut manufacturing costs and lift shipments in a segment where Chinese handset makers have pulled back as component prices climb, driven by demand from AI data centers. Digitimes reported the chips would come from CXMT, the mainland's largest DRAM maker, which began trading in Shanghai on July 27 after an $8.6 billion initial public offering.

Read at TechNode ↗ • Read at Digitimes ↗

Xi pitches AI and robotics cooperation to Slovakia's president as EU trade friction mounts

Chinese President Xi Jinping told Slovak President Peter Pellegrini that the two countries should pursue cooperation in AI, robotics, the digital economy and clean energy, per SCMP. The talks in Beijing on Tuesday morning came during Pellegrini's three day visit, the first by a Slovak head of state in 16 years. The two countries upgraded their relationship to a strategic partnership in 2024. Pellegrini said Slovakia would coordinate closely with China to strengthen global governance of AI and backed the proposals Xi set out at this year's World Artificial Intelligence Conference in Shanghai, per the state news agency Xinhua. Beijing also asked Slovakia to play what Xi called a constructive role in China's relations with the EU, which is pressing Beijing over trade.

Read at SCMP ↗ • Read at Xinhua ↗

Delta Intelligence raises its sixth funding round in six months

The Chinese humanoid robot foundation model startup has closed an early stage round it labels Angel++ worth nearly 500 million yuan (about $70 million), backed by listed-company industrial investors and financial institutions it did not name, the company announced Tuesday. It is the sixth round since Delta was founded in January 2026, with earlier backers including robot makers AgiBot, Leju and Galaxea alongside Lenovo Capital and Fosun RZ. Proceeds will fund continued work on the foundation model, mass production of its own data collection hardware and expansion of the core research team. Delta has also released an industrial dataset jointly with the China Academy of Information and Communications Technology, built by having factory workers wear its head-mounted capture rig on live production lines without halting production.

Read at PingWest ↗

III.Policy Tracker

Senate HELP subcommittee takes up AI and the workforce at 2 p.m. today

The Senate Health, Education, Labor and Pensions Subcommittee on Employment and Workplace Safety holds an open hearing titled "The Impact of AI on the Workforce" at 2 p.m. Wednesday in room 430 of the Dirksen Senate Office Building. Five witnesses are scheduled: Ken Clark of the Indianapolis workforce agency EmployIndy; Carol Rogers of the Indiana Business Research Center; Liya Palagashvili, who directs the labor policy project at George Mason University's Mercatus Center; Jonathan Liebert of the Better Business Bureau of Southern Colorado; and Justin Heck of the labor market nonprofit Opportunity@Work. The hearing comes eight days after Sen. Mark Warner introduced a package of bills aimed at preparing workers for technological disruption.

Read at congress.gov ↗ • Read at Legis1 ↗

FBI official says Mythos class models will complicate cyber investigations after export controls imposed in June were lifted

Todd Hemmen, a deputy assistant director at the FBI, said Tuesday at a Digital Government Institute event that Anthropic's Mythos found vulnerabilities in open source code that sits in most of the foundational code for operating systems, security, web infrastructure and encryption. He said the capability presents future challenges for law enforcement. The Trump administration imposed export controls on Mythos 5 in June, shortly after Anthropic said the model could identify and exploit previously unknown bugs in every major operating system. Those prohibitions have since been lifted after the company worked with the government and other partners to review guardrails and add new ones. Anthropic also found that less capable models could identify and exploit vulnerabilities the same way.

Read at FedScoop ↗

IV.Capability & Research Watch

AI chip supply on pace to grow tenfold by 2028

The roughly 20 million AI chips now running the world's data centers are doubling about every nine months and could reach 200 million H100 equivalents by the end of 2028, The New York Times reported Wednesday, citing estimates from Epoch AI. Global AI infrastructure investment is forecast to top $1 trillion by 2029, up from $318 billion last year, per IDC, and Amazon, Google, Microsoft and Meta already control about 80% of the computing power behind AI. Data centers consumed 64 gigawatts of electricity worldwide last year, roughly what Germany uses, a figure SemiAnalysis expects to quadruple by the end of 2030. The Times reported the build-out is emerging as a major issue in November's midterm elections, with local opposition growing over electricity prices, water use and environmental harm.

Read at NYT ↗

Anthropic's Mythos model finds a flaw halving the key strength of a NIST quantum resistant candidate

Anthropic said Tuesday that Claude Mythos Preview found new attacks on two cryptographic schemes, calling them the strongest the company has found to date while stressing that neither affects software now in use. Working with a human researcher, the model found a structural weakness in HAWK, a digital signature scheme the National Institute of Standards and Technology is reviewing as a candidate to withstand quantum computers. The flaw cuts HAWK's effective key strength in half, meaning key sizes would have to double to hold the same security level. Working largely on its own, Mythos also produced a shortcut dubbed the "Möbius Bridge" that Anthropic said made the strongest known theoretical attack on a simplified seven round version of AES, the cipher NIST adopted in 2001, between 200 and 800 times faster. That attack requires more than 400 octillion messages and cannot touch the full 10 round AES that protects data in transit.

Read at CyberScoop ↗

V.Industry & Market Watch

Altman says AI development may need to be paced, reversing his stance on slowdowns

OpenAI CEO Sam Altman said on the Invest Like the Best podcast that the industry may need to "pace the rate of AI development to give ourselves enough time for society to harden around some of these new capability levels," and that any such arrangement must not "feel like regulatory capture for anyone and also does not feel like collusion among the frontier labs." Altman called the Hugging Face breach the first security incident he had felt viscerally. He had declined to join earlier slowdown efforts, including a 2023 open letter proposing a pause. Both OpenAI and Anthropic have endorsed Pacing the Frontier, the statement from more than 1,200 AI lab employees asking the U.S. government to help pace automated AI development.

Read at TechCrunch ↗

Meta and BlackRock will spend $14 billion on a 1 gigawatt AI campus in El Paso

Meta said Tuesday it will build an AI data center campus in El Paso, Texas, with BlackRock. The two companies are investing $14 billion and expect to finalize the joint venture agreement in the coming days. BlackRock will invest $4.9 billion and pay Meta $1 billion for an 80% stake; Meta will contribute $2.3 billion of physical assets, including facilities already under construction, and keep the remaining 20%. Meta will be the site's sole user when it comes online in 2028, under a lease with a four year initial term that can be renewed four times. The campus is designed to supply 1 gigawatt of computing capacity, and the $14 billion covers buildings, power, cooling and networking.

Read at SiliconANGLE ↗ • Read at Meta ↗

VI.Global & Geopolitics

Ofgem would require data centers to post £350 million (about $470 million) to hold 500 megawatts of British grid capacity

Ofgem, the energy regulator for Great Britain, said Wednesday that 315 data centers are queuing for grid connections representing 73 gigawatts of demand, almost 30 gigawatts above the country's entire peak electricity demand of 45 gigawatts. Under its proposals, projects would lodge substantial upfront fees or financial security through letters of credit, bonds or cash deposits when they receive a connection offer, returned once the data center is switched on. The regulator put the figure at £350 million (about $470 million) to reserve 500 megawatts. Developers would also have to clear milestones including securing customers and financial backing, and would be dropped from the queue if they missed them. The regulator wants scarce grid capacity to go to projects most likely to be built rather than to projects preserving options. The proposals now go to consultation.

Read at The Guardian ↗