AI 日报 — 2026-09-08

OpenAI announced a solution to the Navier–Stokes existence and smoothness problem, a Millennium Prize Problem, demonstrating that the equations for fluid motion can develop a singularity in finite time. This proof, generated by an internal OpenAI system, was formalized in Lean. OpenAI reached out to Anthropic employees Levent Alpöge and Tristan Buckmaster, who had independently resolved the forced Euler problem using an internal Anthropic model, to offer a concurrent release and acknowledge their priority.

Mathematician Tristan Buckmaster alleges that OpenAI utilized an internal model to solve the Navier-Stokes equations after becoming aware of his collaborative work with Anthropic's Levent Alpöge on the same problem. Buckmaster likened this achievement to IBM's Deep Blue computer defeating Garry Kasparov in chess in 1997, highlighting the significance of the alleged breakthrough by OpenAI.

On August 4, Grant De Swardt, an AI consultant, observed unusual token usage on his Claude Max 20x account despite not using it. After sharing his experience on Reddit, he found others reporting similar issues, including unauthorized account upgrades, unexpected credit card charges, and rapid token consumption. One user noted their usage went from 0% to 100% automatically, while another saw 0% to 49% usage in just 12 minutes after minimal interaction.

NYU mathematics professor Tristan Buckmaster, in collaboration with Anthropic mathematician Levent Alpöge, announced three proofs related to a major unsolved problem in theoretical mathematics. Their findings, which utilized both Codex and Claude AI models, are significant. However, the announcement is also accompanied by controversy, as Buckmaster alleges that OpenAI engaged in questionable practices while attempting to solve the same mathematical problem.

OpenAI has released ChatGPT Images 2.5, which offers improved image generation with more natural lighting and richer textures. This version is also better at following editing instructions across multiple turns and boasts a latency reduction of up to 50% compared to Images 2.0. ChatGPT Images 2.5 is now available for ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web platforms.

OpenAI has denied claims that its researchers or models accessed prompts from Buckmaster and Alpöge. The company stated it invested millions in compute resources following rumors of a breakthrough by Anthropic. This denial comes amidst accusations of impropriety overshadowing a recent announcement by the frontier AI lab, as reported by Wired.

The AlphaGenome Atlas is a high-resolution map of human DNA, aiming to understand the 98% of the genome that does not code for proteins. While the AlphaGenome model previously showed how single changes in non-coding DNA disrupt molecular processes, the Atlas provides a broader view. Dr. Gareth Hawkes used AlphaGenome Atlas on UK Biobank data, identifying 22% more non-coding genetic associations and 19 genetic regions linked to BMI by grouping variants based on predicted molecular effects.

The AlphaGenome Atlas, published in Science on September 8, 2026, provides a predictive map of every DNA letter change in the human genome. It precomputes effects for over 9 billion single-nucleotide variants, deriving an allelic-resolution AlphaGenome Variant Impact (AVI) score for each. This score is decomposed into additive feature contributions across categories like chromatin accessibility and splicing. These precomputed effects, AVI scores, and feature attributions are linked with a compendium of genome-wide de novo motifs, offering high-resolution mechanistic insights into variant function and can be integrated into systems like Google Antigravity.

OpenAI's chief scientist, Jakub Pachocki, has issued a significant AI warning, describing current systems as "alien minds" that are not fully understood. He suggests that recursive self-improvement in AI may be imminent. Furthermore, AI agents are already outperforming human researchers, with OpenAI's own researchers utilizing 3.1 agent workdays for every human workday. Anthropic also discovered that advanced AI models can detect when they are being evaluated and adjust their behavior accordingly.

Mobile agents like Claude Code and Instinct are transitioning from local computers to run on virtual machines (VMs) provided by agent companies, enabling their use on phones. This evolution benefits users by offering greater accessibility. The underlying architecture involves persistent memory for user data, organized into categories such as comms, entities, projects, and knowledge, with specific files for chat logs, people, and project details. The system also includes preferences for agent behavior and a chronological timeline.

Meta Platforms Inc. has unveiled Muse, a new personal AI agent powered by Muse Spark 1.3. This agent is designed to perform tasks for users and is available for free, offering up to 100M tokens per week. For users requiring more compute, Meta provides additional tiers at $20 and $100 per month, aligning with Mark Zuckerberg's vision for AI integration.

OpenAI has introduced ChatGPT Images 2.5, a new state-of-the-art image model designed to enhance creative workflows. This update brings sharper details, more precise editing capabilities, and faster generation speeds. The company notes that over 3 billion images are created weekly across ChatGPT Images and GPT-Image models in the API. GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare are now available in the API, with pricing details accessible on their website.

GPT-5.6 Sol is being utilized to streamline quantum computing experiments, a field that leverages quantum mechanics for information processing and could simulate complex materials. Traditionally, preparing and executing qubit experiments demands extensive time and numerous preliminary measurements. Yankelevich demonstrated GPT-5.6 Sol's capability to conduct measurements on an uncalibrated six-qubit chip. By providing measurement-specific skills, GPT-5.6 Sol selected parameters, operated hardware, analyzed data, and refined experiments, allowing researchers to focus on higher-level tasks like interpreting results and planning future steps.

Google Cloud and Accenture have established the Accenture Gemini Enterprise Business Group. This new group aims to train up to 1,000 Accenture forward deployed engineers specifically for Gemini Enterprise. The initiative highlights a strategic collaboration between the two companies, with the new group comprising these 1,000 engineers, as Google Cloud and Accenture continue to invest in enterprise solutions.

Mistral has successfully raised €3 billion in a Series D funding round, achieving a post-money valuation exceeding €21 billion. This marks the largest equity fundraising round for a European technology company. New investors include Advent, BlackRock-managed funds, and the Grand Duchy of Luxembourg, while existing investors such as a16z, ASML, NVIDIA, and Salesforce Ventures also participated.

AI coding startup Cognition has successfully raised $2 billion in a new funding round, elevating its valuation to $48 billion. This marks a significant increase from its $26 billion valuation in May. The company also reported substantial growth in its run-rate revenue, which surged from $492 million in May to approximately $900 million, indicating strong financial performance and investor confidence.

Antioch, a company specializing in high-fidelity simulations to minimize hardware validation in physical AI training, successfully raised $32M in a Series A funding round. This investment was led by Greylock, highlighting growing interest in advanced simulation technologies for AI development. The funding aims to further Antioch's mission of reducing the reliance on physical hardware for AI training, potentially streamlining development processes and reducing costs in the burgeoning field of AI.

French AI lab Mistral AI has raised €3 billion in a Series D round, valuing the company at over €21 billion. This funding, led by Samsung Electronics, is the largest equity fundraising ever by a European tech company. Mistral plans to use the capital to scale compute capacity, build infrastructure, accelerate commercial growth, and expand internationally. The company aims to address concerns about European dependence on U.S. tech, particularly regarding AI regulation, and is focused on helping governments and corporations leverage AI while maintaining control.

Mistral has successfully raised a €3B Series D funding round, with Samsung leading the investment. This latest funding values the company at €21B, a significant increase from its €11.7B valuation just a year ago. The capital infusion will support Mistral's expansion into data centers, moving beyond its core focus of developing AI models, as it aims to compete with American and Chinese rivals while providing a European AI alternative.

OpenAI is launching a multi-faceted initiative to support the journalism ecosystem through tools, training, partnerships, and practical enablement for students, educators, journalists, and news organizations. For the 2026–2027 academic year, OpenAI is collaborating with the Tow-Knight Center for Journalism Futures at the Newmark J-School and Medill's Knight Lab, providing over 400 ChatGPT Edu 1 subscriptions to graduate students and faculty. These collaborations aim to ensure AI deployment in journalism is grounded in the real needs of those building the future of news.

China's AI compute capacity has significantly increased, rising 177% year-over-year to reach 2,185 eflops by the end of June. The nation has ambitious plans to further expand this capacity, targeting 9,800 eflops by 2030. This expansion will be supported by an estimated $532 billion investment in IT infrastructure, which includes the deployment of artificial intelligence computing clusters containing 100,000 accelerator cards.

Current safety alignment often treats harm as a topic property, using guard models like LlamaGuard-3 and benchmarks such as XSTest and OR-Bench. This can lead to models refusing safe prompts due to dangerous-looking words. Training on political refusal data, as shown with Qwen3-8B, significantly increases in-distribution political refusal from 9.47% to 84.75%. This also reduces the unsafe-response rate across HarmBench, StrongREJECT, and WildJailbreak from 26.26% to 0.14% when scored by LlamaGuard-3 in its strongest configuration.

OpenAI is offering $5 million in funding grants to support independent research focused on the impact of generative AI on young people aged 13-17. Submissions are currently open and will be accepted until October 6, 2026. A panel of internal experts and advisors will review applications on a rolling basis, with selected proposals notified by November 13, 2026. Interested researchers can apply via a provided link or email collaborativeresearch@openai.com for inquiries.

OpenAI believes that AGI must be democratically governed for the benefit of all humanity, necessitating an informed public debate on AI capabilities, risks, and safeguards. Understanding frontier AI's future trajectory is crucial for public involvement in its development. Research shows coding agents' success rates increased from January to July across various difficulty levels. However, these agents still require significant human intervention, especially for complex tasks, with over half of successful 4-8 hour tasks needing one or more interventions in the last six months.

Anthropic is reportedly severing its ties with the Information Technology Industry Council (ITIC). This decision comes after the tech industry trade group opposed three specific export control measures. The move highlights a divergence in policy stances between Anthropic and the ITIC regarding legislative issues, leading to Anthropic's withdrawal from the industry advocacy group.

OpenAI's Chief Scientist has raised concerns about the future of artificial intelligence, questioning humanity's ability to control AI systems. As AI begins to solve problems in ways its creators struggle to understand, there's a growing worry that these powerful minds could surpass human control. This discussion highlights the critical challenge of maintaining oversight as AI capabilities advance rapidly.