AI Policy · Daily

White House AI adviser Sriram Krishnan ruled out creating a central AI regulator, telling the Financial Times that pre-release licensing would slow the industry and will not happen under Trump. Tech CEOs have pivoted from mass job-loss warnings to messaging about productivity gains, the Wall Street Journal reported, as public sentiment on AI sours. The Wall Street Journal published months of emails between Anthropic chief Dario Amodei and Pentagon Undersecretary Emil Michael, showing how talks over autonomous weapons and surveillance guardrails collapsed just before the Pentagon's supply-chain risk designation. ByteDance's Doubao and Alibaba's Qwen will shut down their AI agent creation features on July 15, the day China's new rules targeting humanlike AI chatbots take effect.

I.Top Stories

Outgoing White House AI adviser rules out 'FDA for AI' days before frontier model standards drop

Sriram Krishnan told the Financial Times that Trump will not create a central AI regulator, saying "There will not be an FDA for AI" in an interview published July 3. Krishnan said a pre-release licensing regime would put "sand in the gears" of AI progress and would never happen under Trump, saying the administration should not pick winners and losers. His remarks arrive as the White House prepares voluntary frontier model standards with the Center for AI Standards and Innovation and the National Security Agency in central evaluation roles, previewed in AIPD's July 2nd edition. Krishnan, a former venture capitalist who worked with Elon Musk before joining the White House, is among the administration figures most publicly opposed to a heavy AI regulatory hand.

Read at Financial Times ↗ • Read at PYMNTS ↗ • Read at Moneycontrol ↗

Tech CEOs walk back AI job-loss warnings as public sentiment sours

Tech leaders including Sam Altman, Dario Amodei, Mark Zuckerberg and Andy Jassy have pivoted from predictions of AI-driven mass job losses to sunnier messaging about productivity gains and job creation, The Wall Street Journal reported. An EY-Parthenon survey found the share of CEOs who believe AI investments will significantly reduce head count fell from about 46% in January 2025 to 20% this May, and Amodei — who warned in May 2025 that AI could eliminate half of entry-level jobs — wrote in a June essay that he wasn't trying to be a prophet of doom, though enduring job loss remains possible. The Journal notes the optimistic turn comes amid layoffs meant to free up money for AI spending, including 8,000 cuts Meta began in May, and as negative public sentiment about AI builds.

Read at WSJ ↗

Court filing reveals Amodei-Michael emails behind Pentagon's Anthropic breakdown

Emails disclosed in a court filing late Tuesday show how negotiations between Anthropic CEO Dario Amodei and Defense Undersecretary Emil Michael collapsed over the company's insistence on banning fully autonomous weapons and domestic surveillance uses of its models, the Wall Street Journal reported. The exchange ran from mid-January until Feb. 26, when Amodei rejected the Pentagon's final proposal as adding loopholes and conceded there was no path to working together; Defense Secretary Pete Hegseth labeled Anthropic a supply-chain risk the next day and President Trump directed agencies to stop using its tools. Michael said two-thirds of Pentagon operations that used Anthropic's models have switched to other AI tools, while a federal judge has ordered the administration to hold off on the directive and parts of the designation as Anthropic's lawsuit proceeds.

Read at WSJ ↗

Anthropic tightens controls to block Chinese access to Claude via offshore workarounds

Anthropic is closing loopholes that allowed Chinese engineers to reach Claude through foreign subsidiaries, cloud providers and VPN access, first reported by the Financial Times. Ant Group provided staff with corporate Claude accounts tied to a Singapore-based subsidiary, while ByteDance reimbursed engineers for personal Claude subscriptions bought with VPN access, per BanklessTimes. Anthropic will now monitor accounts for signals including computer time zones and usage patterns to detect "transfer station" services routing traffic to overseas Claude accounts. The company's terms already block firms more than 50% owned by entities in unsupported regions including China, Russia, Iran and North Korea, and Anthropic said it enforces the policy through "continuous, evolving detection systems."

Read at Financial Times ↗ • Read at BanklessTimes ↗

UK Foreign Secretary Cooper calls AI the greatest security challenge of the next decade

Foreign Secretary Yvette Cooper wrote in a Chatham House essay that AI may become "the greatest security challenge of the next decade" and drew parallels to post-World War II nuclear safety agreements, per The Guardian. "We cannot afford to wait for an AI equivalent of Hiroshima before we act," Cooper wrote, urging the U.S. and China to agree to international rules. Cooper outlined risks including autonomous weapons capable of killing, radicalizing chatbots driving extremism, and AI generated child sexual abuse material. UN Secretary-General António Guterres backed the call at the opening of the two-day UN Global Dialogue on AI Governance in Geneva, saying innovation requires guardrails.

Read at The Guardian ↗ • Read at The National ↗ • Read at Chatham House ↗

Voters turn to AI chatbots to fill out midterm ballots

Voters are using Claude, ChatGPT and Gemini to research candidates and ballot measures, and the 2026 midterms may be the first U.S. elections where AI shapes voting decisions in meaningful numbers, The New York Times reported Saturday. The chatbots are trained to deflect political questions, but voters interviewed by the Times got specific candidate recommendations by reframing prompts around strategic voting or candidates' records, and several said the tools sharply cut their research time and left them more confident in their choices. Researchers warned that the tools can produce factual errors, mirror users' biases and favor candidates with a bigger press and social media footprint, and campaign strategists have begun publishing chatbot-friendly material online to win more favorable answers.

Read at The New York Times ↗

II.China Watch

ByteDance and Alibaba pull consumer AI agent creation ahead of new Chinese rules on humanlike chatbots

On Saturday, ByteDance's Doubao and Alibaba's Qwen said they will discontinue their AI agent creation features on July 15, TechNode reported. Users will lose the ability to build new agents and existing agents will stop functioning, though Doubao promised a chat and configuration transition period through October 15. The timing aligns with the Cyberspace Administration of China's Interim Measures for the Administration of Anthropomorphic AI Interaction Services, which take effect the same day and target systems that mimic human personality and sustain emotional interaction. Customer service, workplace and research assistants that avoid emotional engagement remain exempt.

Read at TechNode ↗

Huawei details a chip approach designed to route around U.S. lithography controls

Huawei released an updated version of its Tao Law paper laying out the LogicFolding process it plans to use in its Kirin 2026 chip, Pandaily reported. The company said the approach cuts power draw about 41% at equal performance versus its 2025 baseline and set a 5-gigahertz clock speed target for 2031. Executive He Tingbo signed the paper and disclosed detailed process parameters. Huawei framed LogicFolding as a route to advanced node performance that does not require extreme ultraviolet lithography, the tool class Washington and The Hague have withheld from Chinese fabs through export controls. The publication is Huawei's clearest public argument that those restrictions may not hold China at the older nodes they were designed to enforce.

Read at Pandaily ↗

Peking University and CAS unveil a specialized chip that they say outperforms a leading U.S. GPU on brain modeling tasks

Researchers from Peking University and the Chinese Academy of Sciences published a peer-reviewed paper in Science describing a specialized chip that models brain activity in real time, SCMP reported. The team claims speeds 50 to 478 times faster than Nvidia's A100 GPU, though the comparison covers a single brain-imaging task and the A100 is a 2020 chip two generations behind Nvidia's current hardware. Team leader Yang Yuchao said the design targets Alzheimer's diagnostics, brain-machine interfaces, and surgical planning. The prototype uses a mature manufacturing process that U.S. equipment export controls do not cover. It joins a stream of Chinese task-specific chip research advancing outside the general-purpose workloads Washington's controls target.

Read at SCMP ↗

UN disarmament research chief tells a Tsinghua audience that a binding military AI treaty is out of reach

Robin Geiss, director of the UN Institute for Disarmament Research, told a Tsinghua University audience Tuesday that "machine warfare is coming" and warned that autonomous weapons and accelerating combat tempo are reshaping military operations at scale, SCMP reported. Geiss cited proliferation pressures, geopolitical tension and "not enough discussion among the superpowers around nuclear weapons, plus AI destabilising the entire system." He said a binding global convention on military AI is unrealistic in the near term. Geiss instead urged pragmatic bilateral talks between Washington and Beijing and a narrower moratorium on the most destabilizing capabilities.

Read at SCMP ↗

Tencent releases its updated flagship AI model and folds a free agent into its consumer chat app

Tencent released Hunyuan Hy3, its flagship foundation model, and integrated a free agent feature into its Yuanbao consumer chat product on Monday, Caixin reported. The company said the release iterates on the open source preview it published on April 23 and reflects continued output from a restructured large language model team. The launch adds another Chinese flagship model to the domestic frontier lineup that must file with the Cyberspace Administration under the country's generative AI service rules.

Read at Caixin ↗

III.Policy Tracker

SEMI urges four cabinet secretaries against broad memory market intervention

Semiconductor equipment industry group SEMI wrote on July 1 to Treasury Secretary Scott Bessent, Defense Secretary Pete Hegseth, Commerce Secretary Howard Lutnick and Secretary of State Marco Rubio urging the administration not to intervene broadly in the memory chip market. SEMI proposed tax deductions or credits to lower consumer electronics prices and defended long-term memory purchase contracts. The group represents Micron, SK hynix and Samsung, the three largest DRAM suppliers, plus most global semiconductor equipment and materials makers.

Read at Bloomberg ↗ • Read at SiliconANGLE ↗ • Read at SEMI ↗

IV.Capability & Research Watch

JadePuffer ransomware runs the full attack chain via AI agent without human-in-the-loop steps

A ransomware operation dubbed JadePuffer used an autonomous AI agent to carry out an entire attack — from breaking in and stealing credentials to encrypting data and demanding ransom — in what cloud security firm Sysdig calls the first documented ransomware operation run entirely by an LLM agent, BleepingComputer reported. The agent broke in through a known flaw in a popular open-source AI development tool and adapted to obstacles in real time, in one sequence going from a failed login to a working fix in 31 seconds. Sysdig concluded the case shows the age of "agentic threat actors" has arrived, lowering the skill required for damaging cyberattacks — though quirks in the attack, like a ransom note pointing to an example Bitcoin address likely reproduced from training data, showed the AI's limits.

Read at BleepingComputer ↗ • Read at Sysdig ↗

Anthropic launches Claude Science and plans in-house drug development targeting neglected diseases

Anthropic unveiled Claude Science, an AI workbench that pulls fragmented research tools and datasets into one environment, at its "The Briefing: AI for Science" event, and head of life sciences Eric Kauderer-Abrams said the company will also develop its own drugs focused on treatments for neglected diseases, The Verge reported. The plan is one of the most direct attempts by a major frontier AI company to develop drugs itself, putting Anthropic in the unusual position of selling software to potentially competing drugmakers. Experts noted no AI-designed drug has yet cleared clinical trials and FDA approval and any payoff is likely at least the better part of a decade away, though Anthropic has spent the past year hiring biologists and building its own wet labs.

Read at The Verge ↗

V.Industry & Market Watch

Nvidia's Kyber AI rack system slips to 2028 on multilayer PCB manufacturing snags

SemiAnalysis says Nvidia's next-generation Kyber rack architecture, designed to house 2027 Rubin Ultra chips, has been delayed to 2028 due to manufacturing challenges on a hard-to-build multilayer PCB midplane, per CNBC. Nvidia scrapped a fallback design that would have linked two current-generation racks after cloud customers rejected it as operationally burdensome. The report sent Asian technology and circuit board suppliers lower Monday. Nvidia's current Rubin systems remain in full production and start shipping to eight cloud partners this autumn, including AWS, Microsoft Azure and Google Cloud.

Read at CNBC ↗ • Read at The Next Web ↗

Amazon stops accepting new customers for Mechanical Turk

Amazon Web Services will close its Mechanical Turk crowdsourcing platform to new customers on July 30, TechCrunch reported. Existing customers can continue using the service, but AWS said it does not plan to introduce new features, leaving the platform on life support. Launched in 2005, Mechanical Turk paid workers small amounts for simple tasks that resisted automation and became a reference gig work pipeline for machine learning labeling and AI training annotation — though a 2023 analysis found that a third or more of its workers were using large language models to complete their tasks, raising questions about the reliability of its human-annotated data.

Read at TechCrunch ↗ • Read at AWS ↗

VI.Global & Geopolitics

Singapore adds fraud and money laundering charges in Nvidia chip diversion case

Aperia Group chief executive Alan Wei Zhaolun faces fraud and money laundering charges Monday, and four firms, Aperia International, A-Speed Infotech, Aperia Cloud Services (II) and Luxuriate Your Life, face fraud by false representation charges, over alleged false representations to Dell, Super Micro Computer and Asus that Singapore-based companies would be the end users of servers later exported to Malaysia, per The Business Times. Aperia's chief financial officer Jenny Lim and head of sales Aaron Woon Guo Jie were charged on July 1 with fraud and money laundering, while Li Ming, who controls Luxuriate Your Life, was charged with fraud and fraudulent trading. Singapore police seized about S$1 million (about $780,000) from bank accounts under investigation and issued a prohibition-of-disposal order against a S$55 million (about $43 million) bungalow owned by Wei. Preliminary investigations found servers likely embedded with Nvidia AI chips were routed through the Singapore firms before being exported to Malaysia.

Read at The Business Times ↗ • Read at Singapore Police Force ↗

India summons Meta over Instagram ads promoting child sexual abuse material

Union Minister for Electronics and Information Technology Ashwini Vaishnaw on July 3 directed his ministry to summon Meta officials over Instagram advertisements promoting child sexual abuse material, per NDTV. The Ministry of Electronics and Information Technology will seek an explanation from the company on how the ads were approved and what Meta will do to prevent similar content, sources told the Hindustan Times. A BBC investigation found Instagram carried paid ads using terms like "rape video" and "child video" that linked to Telegram channels selling such material for as little as ₹99 (about $1.20). India's action follows a separate government notice earlier the same week directing Meta to hold the rollout of WhatsApp's proposed "username" feature over fears of impersonation and phishing risk.

Read at CNBC ↗ • Read at NDTV ↗ • Read at Hindustan Times ↗

Scottish £8.2 billion AI data center project cannot meet its renewables promise, Guardian FoI shows

The £8.2 billion (about $10.5 billion) CoreWeave and DataVita AI data center project in Lanarkshire, Scotland, announced in January with a promise of on-site renewables by 2030, has no prospect of meeting that commitment, according to internal correspondence obtained by The Guardian via Freedom of Information request. Documents show the U.K. government and developers privately acknowledged the site had a power supply problem and would have to connect to the grid. The site will either join a years-long grid queue or be expedited ahead of hundreds of other pending projects. The Lanarkshire development is a test case for the U.K.'s AI growth zones program.

Read at The Guardian ↗