Show only each day's banner story

OpenAI's SI agents solve one of the "Millennium Problems" — the $1 million Navier–Stokes problem

For the first time, SI itself has solved a problem that the world's best mathematicians struggled with for almost a hundred years — and the proof has been checked by a machine. Science is entering an era in which the hardest questions are solved by swarms of agents in days, and the argument is no longer about whether SI can, but about who gets there first.

About 10,000 agents of an unpublished OpenAI model proved in 88 hours that the three-dimensional Navier–Stokes equations, which describe fluid flow, can produce singularities under a smooth external force — points where the flow velocity goes to infinity. This closes one of the six remaining Millennium Problems of the Clay Institute, with a $1 million prize. The agents exchanged almost 5 million messages, and the proof was then formally verified in the Lean system in 17 hours, so mathematicians consider it correct. The result builds on the "infinite cascade" method of Diego Córdoba and Luis Martínez-Zoroa. NYU professor Tristan Buckmaster and Anthropic mathematician Levent Alpöge, who had worked for almost a year on a related problem, accused OpenAI of "barging into" their work; OpenAI acknowledges that it started on September 1 after rumors of their breakthrough, but denies having access to their materials. Buckmaster called the event a "Deep Blue–Kasparov moment".

Sources 5

quantamagazine.orgopenai.comsemafor.comtechcrunch.comsimonwillison.net

Any point mutation in DNA can now be immediately "checked" against a ready-made map instead of being studied in the lab for years — the search for the causes of rare diseases and for drugs against them speeds up sharply.

DeepMind releases AlphaGenome Atlas — a map of the effects of all 9 billion possible single-letter changes in human DNA

Any point mutation in DNA can now be immediately "checked" against a ready-made map instead of being studied in the lab for years — the search for the causes of rare diseases and for drugs against them speeds up sharply.

The atlas contains predictions of molecular effects for every possible substitution of a single "letter" in the human genome — about 9 billion variants across hundreds of cell and tissue types. Its size is 1 petabyte, more than 30 times larger than the AlphaFold database. The portal is free for academic researchers, and commercial access will appear in Google Cloud. Partners from the Broad Institute and the University of Exeter have already used it to look for causal mutations in rare diseases and experimentally confirmed some of the findings.

Sources 2

deepmind.googleaiweekly.co

States are beginning to guard the knowledge that models give out in their answers as a strategic resource — learning from other people's models will become more dangerous and costly, and the free exchange of knowledge is narrowing under the banner of security.

US intelligence agencies accuse DeepSeek, Moonshot, Alibaba and three other Chinese companies of mass "extraction" of American SI models

States are beginning to guard the knowledge that models give out in their answers as a strategic resource — learning from other people's models will become more dangerous and costly, and the free exchange of knowledge is narrowing under the banner of security.

The NSA, CISA and the FBI issued a joint warning: since late 2024, DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI have "distilled" Claude, GPT, Gemini and Grok on an industrial scale — extracting billions of tokens of responses through millions of queries in order to train their own models on them. To get around blocks, they used fake accounts, bulk purchases of subscriptions, proxy networks ("transshipment points") and prompt injections to extract chains of reasoning. The agencies recommend that American companies not simply block such accounts but quietly degrade their responses and share data on offenders.

Sources 2

cisa.govmedia.defense.gov

Europe gets its own heavyweight in the SI race: countries and companies gain a real alternative to American models, which means frontier SI becomes more accessible and more diverse.

Mistral raises €3 billion — the largest round in the history of European tech companies

Europe gets its own heavyweight in the SI race: countries and companies gain a real alternative to American models, which means frontier SI becomes more accessible and more diverse.

The Series D round, at a valuation of more than €21 billion, was led by Samsung Electronics, with co-investors the Scaleup Europe fund managed by EQT and PSG Equity; funds of BlackRock, Luxembourg, ASML, Nvidia and a16z also took part. The money will go to compute and infrastructure in Europe, frontier research and international expansion. Mistral is betting on "sovereign SI" — open models and private infrastructure that customers fully control themselves.

Sources 2

techcrunch.commistral.ai

The market for SI programmers is growing so fast that there is room not only for the giants — writing code with agents is becoming ordinary work at the world's largest companies.

Cognition, maker of the SI programmer Devin, is now worth $48 billion — revenue almost doubled in four months

The market for SI programmers is growing so fast that there is room not only for the giants — writing code with agents is becoming ordinary work at the world's largest companies.

The startup raised more than $2 billion in a Series E round led by Andreessen Horowitz and Accel; in May it was valued at $26 billion. Annual revenue rose from $492 million in May to almost $900 million. Its customers include Nvidia, GE Aerospace, Citi and Mercedes-Benz; the company is opening offices in Washington, Tokyo, Singapore, London, São Paulo and Madrid and positions itself as an independent agent lab that combines models from different providers.

Sources 2

techcrunch.comcognition.com

A personal SI assistant that acts for you in the real world reaches a mass audience — the main question now is trust: who people are willing to give the keys to their email and wallet.

Meta launches Muse, a personal SI agent that writes emails, books trips and makes purchases on its own

A personal SI assistant that acts for you in the real world reaches a mass audience — the main question now is trust: who people are willing to give the keys to their email and wallet.

The agent connects to email, calendar, payment cards, fitness apps, smart home and shopping services, and can send emails, lower bills, fill out forms and pay for orders through Stripe. Muse is available in the US at muse.ai, in iOS and Android apps and in WhatsApp; plans are free, Power at $20 and Maximum at $100 a month. Meta claims the agent runs in an isolated virtual machine without access to passwords and that conversations are not used for advertising.

Sources 1

techcrunch.com

China is turning the SI race into a state plan with concrete numbers — there will be many times more computing capacity in the world, and SI will develop even faster.

China sets a goal of growing SI compute 4.5-fold by 2030 and investing $532 billion

China is turning the SI race into a state plan with concrete numbers — there will be many times more computing capacity in the world, and SI will develop even faster.

China's Ministry of Industry and Information Technology plans to bring "intelligent computing" capacity to 9,800 exaflops by 2030; as of June 2026 it was 2,185 exaflops, 177% more than a year earlier. 3.8 trillion yuan is allocated to information infrastructure in 2026–2030. The bet is on clusters of 10,000 and 100,000+ accelerators — 52 clusters of 10,000 cards are already operating.

Sources 1

aiweekly.co

Criminals are already handing entire operations to SI agents, and attacks are speeding up many times over. Defenders will have to respond in kind, with their own agents working just as fast.

Google: hackers used SI agents to steal thousands of credentials in under six hours

Criminals are already handing entire operations to SI agents, and attacks are speeding up many times over. Defenders will have to respond in kind, with their own agents working just as fast.

The Google Threat Intelligence group described a financially motivated attacker who built a multi-agent system out of an SI coding chatbot, a prompt and ready-made instructions in markdown. The agents scanned for vulnerabilities on their own, harvested thousands of other people's credentials, fixed errors on the fly and changed IP addresses by routing traffic through compromised cloud environments, all without an operator. Other groups described include UNC6780/TeamPCP, which poisoned the LiteLLM gateway and attacked PyPI, npm and Docker Hub. Google stresses, however, that it has not yet seen fully autonomous attacks that find zero-day vulnerabilities in real-world conditions.

Sources 2

siliconangle.comcloud.google.com

Leading SI labs are already built into the US military machine, and powerful models are becoming part of defense just as computers and the internet once did.

Pentagon contracts with OpenAI, Anthropic, Google and xAI turn out to be much closer than the companies said

Leading SI labs are already built into the US military machine, and powerful models are becoming part of defense just as computers and the internet once did.

The Intercept obtained more than 400 pages of documents on the July 2025 contracts, worth up to $200 million each, through a Freedom of Information Act lawsuit. The companies committed to providing access to their best models, receiving classified briefings on Department of Defense missions, taking part in staff exercises, training military personnel and forecasting risks of their own products. In two documents the Pentagon asked OpenAI for a version of the model that refuses military users as rarely as possible; OpenAI and the Department of Defense said these were drafts and that the final contract contains no such provision. Anthropic declined to be placed on classified networks without a ban on domestic surveillance and autonomous weapons.

Sources 2

theintercept.comtheintercept.com

The SI industry is splitting over the question of who to sell the most powerful chips to — and one of the leading labs has openly gone against the rest, demanding more restrictions.

Anthropic quits the largest tech lobbying group, which tried to bury chip export control bills

The SI industry is splitting over the question of who to sell the most powerful chips to — and one of the leading labs has openly gone against the rest, demanding more restrictions.

The Information Technology Industry Council (ITI), whose members include Google, OpenAI and Nvidia, urged Congress to drop three bills from the defense budget package: the SI OVERWATCH Act on licensing chip exports, the Chip Security Act on tracking SI chips against smuggling, and the MATCH Act on restricting sales of allies' equipment to China. Anthropic supports all three and, in protest, ended its ITI membership. CEO Dario Amodei has also previously spoken publicly in favor of tightening export controls.

Sources 2

axios.comaiweekly.co

Generative SI has given criminals a cheap way to produce prohibited content, and automated moderation is so far losing this race — defense systems must learn as fast as the attackers.

Meta again lets through hundreds of ads with SI-generated images of child abuse — including with real children

Generative SI has given criminals a cheap way to produce prohibited content, and automated moderation is so far losing this race — defense systems must learn as fast as the attackers.

After the Tech Transparency Project's first report on about 50 such ads, researchers found more than 250 more — on Facebook, Instagram, Threads, Messenger and Meta's ad network; in total, more than 350 ads have been identified since last year, and in four of them real minors were identified. Meta removed about 150 ads; most had fewer than 200 impressions, and advertisers' total spending was under $5,000. A former Meta employee noted that for every new detection model, offenders found a workaround within one to five days.

Sources 2

wired.comsfist.com

Another major chipmaker is entering the market for SI data center hardware — competition with Nvidia is growing, and compute for SI will eventually become cheaper.

Qualcomm to develop SI chips for Amazon — shares jump 10%

Another major chipmaker is entering the market for SI data center hardware — competition with Nvidia is growing, and compute for SI will eventually become cheaper.

The companies agreed on multi-year joint development of several generations of specialized chips for SI inference and optical interconnects of up to 1.6 Tbit/s for AWS data centers. Amazon received a warrant to buy up to 25 million Qualcomm shares at $161.26, and Qualcomm will expand its use of AWS for chip design. After the announcement, Qualcomm shares rose about 10% on trading volume 9.3 times the average.

Sources 2

aiweekly.cocnbc.com

SI's hunger for electricity is bringing closed nuclear plants back to life — the development of SI pulls the renewal of the energy sector along with it.

US government grants $1.9 billion to restart a nuclear plant that will power Google's data centers

SI's hunger for electricity is bringing closed nuclear plants back to life — the development of SI pulls the renewal of the energy sector along with it.

The US Department of Energy approved a loan to NextEra Energy to restore the 615 MW Duane Arnold nuclear plant in Iowa, closed since 2020 after a hurricane; the restart is planned for 2029. Google signed an agreement to buy the plant's power for its SI data centers and, according to reports, plans up to six facilities nearby. It is the second major government loan to revive a nuclear plant, after $1 billion for Three Mile Island.

Sources 1

techcrunch.com

SI is becoming a lab assistant that runs physical experiments by itself. Scientists are freed from routine, and the science of quantum computers may speed up.

GPT-5.6 Sol calibrates a quantum chip on its own in an MIT lab

SI is becoming a lab assistant that runs physical experiments by itself. Scientists are freed from routine, and the science of quantum computers may speed up.

MIT graduate student Beatrice Yankelevich connected GPT-5.6 Sol through Codex to a six-qubit superconducting chip. The agent chose experiment parameters itself, controlled the equipment, analyzed the data and calibrated qubit frequencies, control pulses and information storage time; OpenAI writes of 40 target measurements in which a human intervened in only four. On weak, noisy or physically ambiguous signals the agent still needs an expert's help; such calibration usually takes scientists several days per chip.

Sources 2

thequantuminsider.comopenai.com

Editing an image with words is becoming as easy as editing text — another creative tool is moving from the hands of professionals into everyone's hands.

OpenAI releases ChatGPT Images 2.5: images twice as fast and targeted editing without spoiling the rest

Editing an image with words is becoming as easy as editing text — another creative tool is moving from the hands of professionals into everyone's hands.

The new model generates images up to 50% faster than Images 2.0, renders light and textures better, preserves faces and objects from uploaded photos, and changes a single detail without touching the rest. Hand-drawn sketches (Sketch), templates and comments on specific areas of an image have been added. Two models arrived in the API — the fast GPT-Image-2.5 Flare and the precise Sunburst — at $30 per million output tokens. According to OpenAI, users create more than 3 billion images a week.

Sources 1

unite.ai

SI speeds up both finding holes and closing them — software is starting to live at the pace of machines, not people.

Chrome moves to updates every two weeks because SI finds vulnerabilities ever faster

SI speeds up both finding holes and closing them — software is starting to live at the pace of machines, not people.

Starting with Chrome 153, the browser updates every two weeks instead of every four on desktop, iOS and Android. Google explains that automated SI tools and researchers' reports have sharply increased the number of fixes, and that threats partly created with the help of SI require shortening the window between a vulnerability's disclosure and delivery of the patch to users. Mozilla, Microsoft and Brave have already moved to a similar schedule.

Sources 1

techcrunch.com

Access to powerful SI has become so valuable that it is stolen like bank passwords, so an SI subscription now deserves the same serious protection.

Hackers steal Claude sessions with infostealer malware and burn through victims' usage limits

Access to powerful SI has become so valuable that it is stolen like bank passwords, so an SI subscription now deserves the same serious protection.

Anthropic confirmed that attackers are using common infostealers to steal Claude login sessions from users' computers and spend their limits: for some subscribers, usage jumped from 0 to 100% within minutes, and one had their limit drained three days in a row. When it detects this, the company logs the user out, revokes authorizations and refunds the money, but it does not provide a detailed usage breakdown that would help users spot the theft.

Sources 1

techcrunch.com

The bottleneck is no longer the models themselves but the people who can build them into a business, and there is now a separate race for such specialists.

Google Cloud and Accenture create a force of 1,000 engineers to deploy SI in companies

The bottleneck is no longer the models themselves but the people who can build them into a business, and there is now a separate race for such specialists.

The joint Accenture Gemini Enterprise Business Group will train up to 1,000 Accenture consultants as "forward-deployed engineers" who will build SI applications on the Gemini Enterprise platform right at client sites. Google is catching up with its rivals: according to August data, it accounted for about 6% of enterprise SI spending, versus 43.5% for Anthropic and 39.7% for OpenAI.

Sources 1

techcrunch.com

Models that learned to generate video are starting to teach robots to act in the real world, and the line between generating an image and taking physical action is blurring.

Runway doubles revenue to $200 million and acquires the Kinetix startup team for robot "world models"

Models that learned to generate video are starting to teach robots to act in the real world, and the line between generating an image and taking physical action is blurring.

Annual revenue of the video generator Runway has doubled since April to $200 million; the service is used by more than 60 million creators and 95% of Fortune 100 companies. The team of the Paris SI lab Kinetix has joined Runway to develop world models for robotics. Co-founder Anastasis Germanidis called robots "one of the largest applications of world models."

Sources 1

pymnts.com

The main driver of SI turned out to be not clever neural-network designs but what they are trained on, so the next leap depends on whether SI learns to create good data for itself.

Study: language-model progress in 2019-2025 came mostly from data, not new architectures

The main driver of SI turned out to be not clever neural-network designs but what they are trained on, so the next leap depends on whether SI learns to create good data for itself.

Dwarkesh Patel and Jerry Han trained model recipes from GPT-2 to OLMo-2 on corpora from OpenWebText to UltraFineWeb and separated the contribution of each factor: at a budget of 1e19 FLOP, data improvements gave a 12x efficiency gain and model improvements a 3.7x gain, i.e. 3.24 times less. The authors caveat that architectural innovations are valuable chiefly because they made it possible to scale training, and that the main open question is whether synthetic data can extend the dwindling supply of internet data.

Sources 1

dwarkesh.com

Generative SI makes disguising dangerous content almost free, so platforms will have to recognize the meaning of videos, not just familiar imagery.

Extremists turn propaganda into SI cartoons with SpongeBob to evade moderation

Generative SI makes disguising dangerous content almost free, so platforms will have to recognize the meaning of videos, not just familiar imagery.

The Institute for Strategic Dialogue (ISD) described a trend it calls "slop jihad": extremist supporters generate SI videos in the style of fan edits, memes and cartoons, for example a remake of an ISIS execution with fish from SpongeBob SquarePants in orange jumpsuits. Researchers tracked 150 such videos in 71 TikTok accounts over six months, with 5.4 million views. Stylized content is not recognized by systems trained on familiar jihadist symbolism and reaches teenagers more effectively.

Sources 1

aiweekly.co

SI agents are entering everyday work at Russian companies through the familiar 1C and CRM systems, and an assistant will increasingly handle office routine.

Yandex launches Alice AI for Business, an assistant with access to 1C, Bitrix24 and amoCRM

SI agents are entering everyday work at Russian companies through the familiar 1C and CRM systems, and an assistant will increasingly handle office routine.

The corporate assistant performs multi-step tasks through agentic scenarios: it analyzes meetings, finds recurring problems in customer requests and consolidates financial data from different systems. It works with Yandex 360 Mail, Disk, Calendar, Tracker and Wiki, and through plugins with Bitrix24, SPARK, amoCRM and 1C. The service is immediately available to all Yandex 360 customers (more than 185,000 organizations), and to others free for three months; a version for Yandex Cloud with hybrid deployment will follow later.

Sources 1

yandex.ru

The world's best mathematicians are taking up SI safety, which means reliable and understandable models are getting a rigorous scientific foundation.

Fields Medalist Jacob Tsimerman founds a mathematical SI safety institute

The world's best mathematicians are taking up SI safety, which means reliable and understandable models are getting a rigorous scientific foundation.

The Mathematical AI Safety Institute (MAISI) is an independent nonprofit in the San Francisco Bay Area that will build rigorous mathematical foundations for the safety of powerful SI systems. The scientific director is Tsimerman, a 2026 Fields Medal winner, and the executive director is Andrew Critch; the advisers include Ravi Vakil, Timothy Gowers and Paul Christiano. The first semester in January 2027 is planned for 10-30 mathematicians, and a "special year" starting in September 2027 for 30-100. Tsimerman himself is concurrently joining OpenAI's safety team.

Sources 2

cryptobriefing.comx.com

An SI assistant is becoming a standard school tool for an entire state, and a generation that learns alongside SI is already at its desks.

Missouri opens Gemini and Google AI training to 1.1 million students

An SI assistant is becoming a standard school tool for an entire state, and a generation that learns alongside SI is already at its desks.

Under the state's partnership with Google, nearly 100,000 educators and more than 1.1 million school and university students get free access to Gemini for Education and Gemini Notebook, teacher courses and professional certificates; residents of the state can access them through employment centers. Google claims SI saves teachers up to 10 hours a week on average.

Sources 1

blog.google

SI helps small languages avoid disappearing: even a language with 14,000 speakers gets a link to a hundred other languages of the world.

Yandex Translate learns to translate from Altai and Karelian

SI helps small languages avoid disappearing: even a language with 14,000 speakers gets a link to a hundred other languages of the world.

The neural translator has added Altai (about 70,000 speakers) and Karelian (about 14,000) with translation into more than 100 languages. The service now supports 9 Turkic and 8 Finno-Ugric languages of the peoples of Russia; the models were trained together with the Ministry of National Policy of Karelia, the Karelian Research Centre and Gorno-Altaisk University.

Sources 1

yandex.ru
Previous day
Follow on Telegram