Mathematicians back Claude assisted Riemann proof, US Senate passes grid cost bill, and more - Daily AI Brief #589
In this edition, covering 19 to 21 November:
- Independent mathematicians, including MIT's Andrew Sutherland, say the proof of the Riemann Hypothesis that Anthropic mathematician Levent Alpöge found with Claude is almost certainly correct; one final check is pending.
- The US Senate passed a bill by 62 to 33 that would make data centers larger than 20 megawatts cover more of the grid costs linked to their power use.
- California Attorney General Rob Bonta issued an investigative subpoena to Nvidia over whether advanced chips or confidential information reached China.
From our lists
An AI Brief list tracks AI tools for medical coding and billing.

AI medical coding and billing tools
Riemann Hypothesis
Anthropic mathematician Levent Alpöge says he and Claude found a proof of the Riemann Hypothesis and formalized it in Lean. "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis," he wrote. Independent mathematicians, including MIT's Andrew Sutherland, recompiled the proof without issues and say it is almost certainly correct, Scientific American reports; one final comparator check is pending. The same day, Stanford's Ravi Vakil posted a separate, human written proof of the Generalized Riemann Hypothesis by Alessandro Masullo. Riemann proof
Alpöge and New York University professor Tristan Buckmaster also say they produced an AI aided proof that closes a loophole in OpenAI's September Navier Stokes blowup result, under more realistic conditions. They have not yet released a verification certificate. Navier Stokes
Nvidia and chips
California Attorney General Rob Bonta issued an investigative subpoena to Nvidia over whether advanced chips or confidential information reached China, and whether the company misled investors. The probe follows reports about former CEO Jensen Huang's affair with an alleged Chinese intelligence asset. Subpoena
ByteDance's data center plans in Malaysia face delays. Banks froze a $480 million loan for AI chips in Malaysia after new US export rules covering Malaysia and Singapore. Malaysia
Malaysia's YTL Power denies reports that its Johor data centres helped Alibaba access Nvidia chips through Megaspeed International and Speedmatrix. The company says its facilities serve global customers under Malaysian law and US trade controls. YTL Power
Nscale brought its first Nvidia Vera Rubin NVL72 rack online, a Dell PowerEdge XE9812 system, and plans a worldwide rollout in the coming months. Nscale
Agents at work
Apple may be developing an AI agent for the Mac that could use apps, browse websites, run commands and write and debug code, according to a job listing. The project is at an early stage, and Apple is also adding controls that make it harder for third party agents to gain full access to Macs. Apple
Google Labs introduced Stitch Loop, a long running agent that connects to codebases, metrics, logs and user feedback, writes code in cloud sandboxes and opens GitHub pull requests for review. A global waitlist is open. Stitch Loop
Superhuman is rolling out a more personalized version of its Go assistant across Mail and Docs, with background tasks, routines and 20 connected agents. The company says Go has more than 1 million active users. Superhuman Go
Danaher and Microsoft are building a platform on Azure for AI agents that control scientific instruments and interpret results in laboratories. More agent enabled instruments are planned for 2027. Danaher
Signal blocks screenshots on iOS 27 and detects AI assistants trying to access screens on supported Android devices. It is asking Apple, Google and Microsoft for clearer controls and documentation for AI data access. Signal
Anthropic says Claude writes and reviews about 90% of the code produced by the Claude Code team. Deterministic checks run before model based reviews, and engineers remain accountable for shipped code; the company did not disclose reviewer error rates. Claude Code
OpenAI
Oracle says it has secured all required financing for its AI infrastructure contract with OpenAI, including vendor financing. Oracle says its Abilene campus is nearly complete and that Nvidia Vera Rubin rack deliveries remain on track for the coming weeks. Oracle
The U.S. Army plans to work with OpenAI in the first phase of Project Gatekeeper, a system to speed up security checks for commercial software vendors. Army CIO Gabe Chiulli said AI model integration is planned for the next quarter. Project Gatekeeper
OpenAI says updated safeguards against unauthorized distillation cut the average success rate for extracting protected reasoning from GPT 6.2 Astra to 0.2 percent, compared with 54.0 percent for GPT 6 Astra under its original safeguards. Distillation
OpenAI introduced MedChemBench, a medicinal chemistry benchmark of 99 problems from 117 research papers and patents. Models did better at structure reconstruction and scientific reasoning than at compound selection and numerical prediction. MedChemBench
OpenAI is expanding ChatGPT Health in the United States from iPhone, iPad and the web to its Mac app, Android, ChatGPT Sites and Dots. Conversations that reference connected Health data may be used to improve its models when the model improvement setting is on, OpenAI says. ChatGPT Health
OpenAI says its researchers now use an average of 5.7 coding agent workdays per researcher per day, and that safety and alignment work nearly doubled during 2026. Research report
OpenAI built Performance Factory, an internal system powered by Codex that investigates performance regressions in ChatGPT. It tracked more than 1,700 investigations and found the causes of over 600 in just over three months. Performance Factory
Chile's Comptroller General's Office gave the Ministry of Science 10 business days to provide information about its agreement with OpenAI, including its legality and data protection. The request does not determine whether the agreement is legal. Chile
Data centers and energy
The US Senate passed the bipartisan American Affordability and Jobs Act of 2026 by 62 to 33. The bill would direct FERC to require data centers larger than 20 megawatts to cover more of the grid infrastructure costs linked to their electricity use. It now goes to the House of Representatives. Senate bill
A New Jersey judge let a class action proceed against DataOne over noise from its 300 megawatt Vineland data center, which is being built for Nebius Group's AI infrastructure under its $17 billion deal with Microsoft. DataOne says some noise may come from temporary construction. Vineland noise
President Donald Trump plans tax penalties for US technology companies that build AI data centers outside the United States. He says the penalty will be large but has not provided details. Tax penalties
An MIT study estimates that air pollution linked to US data center electricity demand caused about 650 premature deaths in 2023. The health effects were highest in several Midwestern states, while much of the demand is in Virginia and Texas. MIT study
A study by George Washington University and American University of 5,291 US data centers found that majority Black neighborhoods had 2.2 times the odds of hosting one compared with white neighborhoods in the same state, after accounting for income, land costs and other factors. Racial gaps
ECB President Christine Lagarde said European public funds, including the EU budget, should take initial losses on chip plants and data centers to attract private investment. Lagarde
Chicago Fed President Austan Goolsbee said rapid AI investment and data center construction could strain the US economy's capacity and increase inflationary pressure. Goolsbee
Tesla unveiled Powercore, a platform that combines a solid state transformer, switchgear and inverter, delivers up to 7.4 MW and outputs 800VDC directly to servers. Tesla is developing it with Nvidia and expects operations to begin in 2027. Powercore
Policy
Representatives Nathaniel Moran, George Whitesides and Don Beyer formed the House AI Safety and Innovation Caucus, the first congressional caucus focused on AI safety and security risks. House caucus
Senator Chris Van Hollen is making AI regulation central to his challenge to Senate Democratic leader Chuck Schumer, with proposals for mandatory safeguards and liability for harms. Democratic senators vote on 1 December. Van Hollen
The U.S. Government Accountability Office reports that generative and agentic AI can increase the speed, scale and precision of cyberattacks while improving threat detection, and outlines 16 policy options. GAO report
The UK Government says it will keep AI regulation under review and rejects immediate commitments to a new AI Bill or an independent AI regulator. UK response
Australia and China will hold talks on AI after Prime Minister Anthony Albanese met Chinese Premier Li Qiang in Beijing. Australia and China
A Maricopa County judicial officer dismissed a felony aggravated assault charge in a Gilbert, Arizona, case after police used AI to help draft the probable cause statement. AI police reports
The U.S. Conference of Catholic Bishops formed an AI task force led by Bishop Oscar Cantú to lobby for clear regulations, monitor job losses and promote AI literacy. Catholic bishops
Banks and security
The Bank of England, the PRA and the FCA published nonbinding practices for financial firms to manage AI related cyber risks, covering governance, human oversight and AI for cyber defence. UK regulators
The UK Financial Conduct Authority launched the AI Spotlight Observatory, an open platform with AI use cases from more than 200 organisations across 150 financial services activities. FCA
The Monetary Authority of Singapore outlined safeguards for AI agents that handle payments and financial decisions, including a trusted agent access protocol and an Agentic Money Sandbox. MAS
Beazley Security says ransomware incidents rose 38% from the second to the third quarter, and its report details an intrusion carried out largely by AI agents. Beazley
Schools and consumers
Starting next year, Google will disable Search for schools that opt out of its Gemini powered features. School districts in New York, Los Angeles and Chicago are considering other search services. Google and schools
A Korea Consumer Agency survey found that 56 of 110 people who stopped or changed medication based on generative AI advice had side effects. Medication advice
The Dutch Data Protection Authority advises consumers not to buy AI glasses after testing six models and finding incorrect medication and emergency advice, weak privacy protections and security flaws. AI glasses
Canva introduced free controls that let Free and Pro users turn off generative AI features. Canva
Coca-Cola launched its US holiday campaign with human made advertising, after its recent generative AI holiday ads divided consumers. Coca-Cola
Reviewers criticized reported generative AI use in The Further Mis-Adventures of Cliff Booth, directed by David Fincher. The film reaches cinemas on 25 November. Cliff Booth
Target says customers buy at about twice the rate when they use its beta Ask Target shopping assistant. Target
Jobs
October data from the Stanford Digital Economy Lab and ADP Research show employment in highly AI exposed US occupations fell 0.8% year over year, and 4.2% among workers aged 22 to 25. ADP data
Ubisoft is cutting 166 jobs at its Montreal and Quebec City studios and opening Motherbrain, a subsidiary focused on AI for gameplay. Ubisoft
Microsoft CEO Satya Nadella said AI could increase cybersecurity risks and lead to more outages, and discussed voluntary exit offers for about 60,000 employees. Nadella
Research
METR's November evaluation of public models, including Opus 5.5 and GPT-6 Astra, finds that time horizons on its tasks double roughly every 3.6 months. METR says current evidence cannot rule out drastic AI R&D acceleration in the next few years. METR
An analysis co-authored by Stanford's Owen Queen retested 50 widely cited LLM behaviours on 20 models and found nearly half had largely disappeared after about 16 months, while sycophancy and self-preference grew stronger. Model habits
Thinking Machines Lab released Inkling 2 Mini, an open weight multimodal model with 26.4 billion total and 3.4 billion active parameters that runs on compatible laptops and desktops. Inkling 2 Mini
Researchers from the University of Illinois Urbana-Champaign and Nanyang Technological University analysed 22,096 uncensored language models on five hosting platforms and found many are shared without safety warnings. Uncensored models
Taylor & Francis retracted a paper by five academics at Northeast Forestry University in China after finding hidden prompts that told AI systems to give it a positive review. Retraction
In a Fortune test, the startup Samaya beat the Wall Street analyst consensus on earnings forecasts for 40 of 50 companies reporting third quarter results. Samaya
China
Tencent detailed its Hy4 model, a Mixture of Experts design with 770 billion total language parameters, 49 billion active per token and a context window of up to 1,048,576 tokens. Hy4
Pragmatik Labs, the Shanghai startup founded by former Alibaba AI chief Junyang Lin, raised tens of millions of dollars at a $5 billion valuation, Bloomberg reported citing unnamed sources. It was valued at $2 billion in August. Pragmatik Labs
A US House report says Cursor, which SpaceX acquired in August, is considering phasing out Chinese open weight models because of geopolitical risk. Cursor built its Composer 2 coding model on Moonshot AI's Kimi K2.5. Cursor
Kuaishou reported third quarter revenue of RMB31.6 billion and a net loss of RMB959 million, and introduced its Kling 4.0 video generation model. Kuaishou
Deals and companies
Isomorphic Labs and MSD will use AI models and MSD's data to find small molecule candidates for an oncology target, with upfront, milestone and royalty payments for Isomorphic Labs. Isomorphic Labs
John Deere acquired Orchard Robotics, which uses vehicle mounted cameras and AI software to analyze individual plants, trees and vines. The terms were not disclosed. John Deere
Mistral AI opened its Asia Pacific headquarters in Singapore and plans 200 regional staff by 2027. Mistral AI
BDO will roll out Legora across the legal practices of its member firms after pilots in Spain and Norway. BDO
Auterion and Ukrainian drone maker SkyFall plan a joint venture to manufacture drones in Germany for Ukraine, starting with Shrike FPV drones. Auterion
Helsing is scaling back its underwater drone work in Plymouth after it missed out on the Royal Navy's £90 million Atlantic Net deal, which went to Ocean Infinity Solutions. Helsing
We hope you enjoyed this article
Consider subscribing to one of our newsletters like Daily AI Brief.
Also, consider following us on social media:
Daily AI Brief
Daily report covering major AI developments and industry news, with both top stories and complete market updates