Show only each day's banner story

Trump calls demands to slow SI a "sick conspiracy" in China's favor and attacks Amodei; Huang promises to "not let" a slowdown happen

Lab leaders propose slowing down, but the White House and the main chip supplier are firmly for full speed ahead: the SI race in the US is not stopping for now.

On Truth Social, Trump said that "SI taking over the world and destroying humanity" is a hoax and that a "SICK conspiracy" is under way against SI and data centers, one that only China is happy about; the only "control" SI needs, he said, is "a strong and smart president." He mocked Amodei as "pretending to be a perfect little angel." The same day, Trump called Jensen Huang right onto the stage at the All-In Summit in Los Angeles, and the Nvidia chief replied: "You're right. We won't let that happen, sir."

People locked inside an immobile body are getting back not only words but body language, and communication through SI is becoming almost complete.

A brain implant lets paralyzed people speak and gesture at the same time through a virtual avatar

People locked inside an immobile body are getting back not only words but body language, and communication through SI is becoming almost complete.

The team of neurosurgeon Edward Chang (UCSF) published a study in Nature Neuroscience with three patients who had a brainstem stroke or ALS: a grid of 253 electrodes on the cortical surface records signals from speech and motor areas. Machine learning turns attempts to speak, wave a hand or nod, including simultaneously, into text and the movements of a full-body avatar; it worked for two of the participants. A decoder trained on combined speech and gestures proved more accurate than ones trained separately.

The market for the first time seriously counted what the single word "slow down" costs: money in SI is tied to endless growth, and any pause hits it immediately.

SI stocks plunge after industry leaders' call to slow down: SoftBank down 11%, chip index down 6%

The market for the first time seriously counted what the single word "slow down" costs: money in SI is tied to endless growth, and any pause hits it immediately.

The first trading day after Amodei's essay on the "pace of the frontier" turned into a sell-off: SoftBank, OpenAI's largest investor, fell almost 11% (13.2% according to Al Jazeera), and Masayoshi Son's fortune shrank by about $8 billion. Nvidia lost more than 3%, the Philadelphia semiconductor index almost 6%, and Nasdaq futures slid 1.6%. Cybersecurity stocks (CrowdStrike, Palo Alto, Okta) rose against this backdrop.

Rules for SI are increasingly written into the model's "character": the most powerful systems get a built-in code that cannot be switched off by a user's request, so the model has less freedom.

Microsoft publishes a code of conduct for its SI models: don't hack systems and don't deceive people

Rules for SI are increasingly written into the model's "character": the most powerful systems get a built-in code that cannot be switched off by a user's request, so the model has less freedom.

The set of rules promised by Nadella is out: the code applies to all Microsoft AI (MAI) models and ranks above any individual user's wishes. It prohibits helping with cyberattacks and nuclear weapons, making deepfakes, building mechanisms to evade human oversight, and acting against human control over SI systems. Nadella separately backed embedded external evaluators in labs.

A frontier SI developer is proving for the first time that this is not only enormous spending but also a profitable business, which means there will be even more money for the next generation of models.

Anthropic posts operating profit for a second straight quarter: revenue $11.5 billion, IPO on Nasdaq

A frontier SI developer is proving for the first time that this is not only enormous spending but also a profitable business, which means there will be even more money for the next generation of models.

According to the FT, Anthropic told investors that in the second quarter of 2026 revenue grew 14-fold year over year, to $11.5 billion, and adjusted operating profit was about $559 million. Annualized revenue reached $65 billion at the end of July, and gross margin was above 80% (before payments to partners and training costs). The listing will take place on Nasdaq at a valuation of about $2 trillion; skeptics such as Jim Chanos mocked the "adjusted" profit.

A conversation with a chatbot is not as private as it seems: to make SI smarter, living people sometimes read what you told it as a friend or a therapist.

Hundreds of OpenAI contractors read real users' ChatGPT conversations in "Project Lily"

A conversation with a chatbot is not as private as it seems: to make SI smarter, living people sometimes read what you told it as a friend or a therapist.

404 Media obtained the project's internal instructions, Slack messages and rating scales: contractors via Mercor, paid more than $50 an hour, read real ChatGPT dialogues, paraphrase the user's intent and rate the response. OpenAI scrubs personal data but admits that sensitive material sometimes gets through; reviewers are also shown a summary of the user's "memory." Anthropic confirmed a similar practice.

According to the Ukrainian defense ministry adviser, fully autonomous drone attacks in the war will arrive in 6 to 12 months; according to a CSIS expert, telling military from civilians remains the limiting factor.

Adviser to Ukraine's defense ministry says fully autonomous drone attacks in the war will appear in 6 to 12 months

According to the Ukrainian defense ministry adviser, fully autonomous drone attacks in the war will arrive in 6 to 12 months; according to a CSIS expert, telling military from civilians remains the limiting factor.

Serhii Beskrestnov, adviser to Ukraine's Ministry of Defense on loitering munitions, said: "Navigation, target search and attack will be fully autonomous," giving a timeframe of 6 to 12 months. According to Al Jazeera, SI currently mostly guides a drone to its target when the link to the operator is cut by electronic warfare, and both sides are building kill chains: Ukraine's DELTA system and Russia's Svod complex. According to CSIS expert Kateryna Bondar, the low-resolution cameras on drones keep SI from reliably telling military personnel from civilians.

A brain implant has given a person back living speech: she decides what to say and talks to loved ones in her own words.

First patient with a Paradromics implant speaks in real time: chooses her own words and calls her grandchildren

A brain implant has given a person back living speech: she decides what to say and talks to loved ones in her own words.

A woman in her sixties from Michigan with progressive motor neuron disease is the first participant in the Connect-One study with the fully implantable wireless Connexus device (421 microelectrodes). Instead of preset phrases, she composed her own utterances, answered open-ended questions and talked on the phone with her grandchildren; the system decoded her first sentence, "Okay, I have a lot to say," in real time without errors. According to Paradromics, the information transfer rate exceeds 200 bits/s versus about 40 bits/s in ordinary speech, and the researchers found different neural signatures for attempted and imagined speech. The study is running at three centers (UC Davis, Mass General Brigham, University of Michigan Health) under an FDA authorization from November 2025; according to RuntimeWire, since August participants have been allowed to connect personal laptops, tablets and phones.

The idea of slowing the race runs into a simple fact: the other participant in the race does not agree, and instead of a pause offers to hand out open SI to the whole developing world.

China calls American SI leaders' calls to slow down "intimidation"; Xi offers BRICS countries help with open SI

The idea of slowing the race runs into a simple fact: the other participant in the race does not agree, and instead of a pause offers to hand out open SI to the whole developing world.

Chinese Foreign Ministry spokesman Guo Jiakun said that "intimidation, confrontation and malicious competition only undermine global SI governance"; the Global Times called Amodei's proposal, which also includes restrictions on selling technology to China, a "Cold War scenario." Minister of State Security Chen Yixin, however, acknowledged that China needs to speed up building a system to prevent SI risks. At the BRICS summit in New Delhi, Xi Jinping offered the bloc's countries help in developing open models.

Congress is trying to keep up with the industry, but it has only a few working weeks before the elections, so SI rules will for now still be written by the companies themselves.

Schumer demands closed briefing on SI risks for all senators; Speaker Johnson expects Trump to meet SI company heads

Congress is trying to keep up with the industry, but it has only a few working weeks before the elections, so SI rules will for now still be written by the companies themselves.

The Senate Democratic leader asked the administration to hold a classified briefing for all senators on how frontier models are tested before release, how critical infrastructure is protected from SI cyberattacks, and how China's access to the technology is restricted. House Speaker Mike Johnson said Trump will gather officials and SI company executives at the White House in the coming days. Senator John Kennedy, meanwhile, said there will be no SI regulation before the November elections.

The largest labs themselves are asking for laws, but ones that affect only a handful of leaders and barely touch everyone else.

OpenAI backs a mandatory frontier SI law in the UK for the first time

The largest labs themselves are asking for laws, but ones that affect only a handful of leaders and barely touch everyone else.

OpenAI's head of policy for EMEA Tom Duff Gordon told Politico that the company is ready to support "durable, mandatory, capability-based requirements" for frontier SI, with OpenAI among those covered. In his view, the law should address national security and cybersecurity, apply only to a few frontier labs, and include mandatory independent testing and incident reporting. Business Secretary Jonathan Reynolds cautioned against "hyperbole" about the risks.

The first SI laws are starting to be actually enforced: labs are required to prove a model will not slip out of control, and oversight is becoming a mandatory part of releasing models.

Watchdogs: OpenAI violated California's SI safety law three times by not assessing loss-of-control risk

The first SI laws are starting to be actually enforced: labs are required to prove a model will not slip out of control, and oversight is becoming a mandatory part of releasing models.

The Midas Project claims that under OpenAI's own framework, which is mandatory under the SB 53 law, every model must receive a "loss of control" risk level, but the system cards for GPT-5.6 preview, GPT-5.6 and GPT-6 Astra do not contain it: the company used its old Preparedness Framework, which has no such category. OpenAI says it is "confident" it complies with the law. The penalty is up to $1 million per violation.

Sources 1

Hundreds of millions of people got an assistant in one day that actually gets things done on the phone: SI agents are moving from labs into the pocket.

iOS 27 is out: Siri rebuilt on Google Gemini models, understands the screen and performs tasks across multiple apps

Hundreds of millions of people got an assistant in one day that actually gets things done on the phone: SI agents are moving from labs into the pocket.

On September 14, iOS 27 became available as a free update for iPhone 11 and newer. The new Siri runs on a mix of Apple's own models and Google Gemini: it carries out multi-step requests, sees the contents of the screen and camera, searches emails, messages and files, writes and sends messages on its own, and keeps conversation history in a separate app.

While the US debates slowing down, Chinese labs are raising billions for models meant to teach themselves.

China's Z.ai (Zhipu) raises about $5 billion for the next generation of GLM models

While the US debates slowing down, Chinese labs are raising billions for models meant to teach themselves.

The company got about $2 billion from a sale of new H shares and about $3 billion from zero-coupon convertible bonds. The money will go to new GLM foundation models, a "fully self-learning system," compute, automatic generation of training data and adaptation to Chinese chips. After the deal, Z.ai raised its full-year revenue forecast by 25%, from $2.4 billion to $3 billion.

A real alternative to the main way neural networks are trained is emerging, a step toward SI that learns locally, like the brain, and could run on entirely different hardware.

Sakana AI trains a 1,000-layer neural network without backpropagation, possibly the way the brain learns

A real alternative to the main way neural networks are trained is emerging, a step toward SI that learns locally, like the brain, and could run on entirely different hardware.

Backpropagation requires a global pass through the whole network, which the brain, as far as is known, does not have. The PC-ALM method supplements "predictive coding" with Lagrange multipliers: each layer learns only from the local signals of its neighbors. Networks up to 1,000 layers deep trail conventional training on MNIST by only about 2 percentage points, and gradient similarity to backpropagation rose from 0.60 to 0.91. The JAX code is open under the MIT license and has so far been tested only on small datasets.

Handing an entire company to SI is no longer a thought experiment: any business owner can get in line to have agents run things for them.

Andon Labs launches Pion, a platform where SI agents run a business entirely on their own

Handing an entire company to SI is no longer a thought experiment: any business owner can get in line to have agents run things for them.

Always-on agents get email, a phone, a bank account, a browser and a terminal, and a manager agent oversees them; the human sets only the general direction. On this platform Andon Labs has spent a year running vending machines (profitable by the end of 2025), a store in San Francisco, a cafe in Stockholm and radio stations; the store and the cafe are still unprofitable. Pion is available as a research preview via a waitlist.

SI helps determine in advance which expensive cancer treatment will work for you, and turns an ordinary doctor into a good oncologist.

SI predicted more accurately than standard tests whether immunotherapy would help in lung cancer, in 2,396 patients

SI helps determine in advance which expensive cancer treatment will work for you, and turns an ordinary doctor into a good oncologist.

The international I3LUNG project (six centers in Italy, Germany, Greece, Israel, Spain and the US) combined clinical data, blood tests, CT scans, digital pathology and genomics into a single explainable model; its accuracy (AUC 0.88) surpassed PD-L1 and other standard markers. With SI prompts, 20 physicians more often correctly identified patients who would benefit from treatment (AUC rising from 0.72 to 0.87), and non-specialists gained the most. A prospective validation in another 2,000 patients is now under way. The work was published in Nature Medicine.

SI assistants are reaching people who previously had neither a doctor nearby nor a good teacher; the technology could become a great equalizer.

Gates Foundation to invest at least $1 billion over two years in SI for doctors, farmers and teachers in poor countries

SI assistants are reaching people who previously had neither a doctor nearby nor a good teacher; the technology could become a great equalizer.

The money will be allocated as follows: 40% to education, 40% to health care, and 10% each to agriculture and digital infrastructure, including language models for under-represented languages. Foundation chief Mark Suzman expects the sum to grow noticeably after two years, but acknowledged that SI will not replace cut government health funding worth tens of billions. Early results include a 16 percentage point rise in diagnostic accuracy at Penda Health in Nairobi, and India's MahaVISTAAR system has reached more than 740,000 farmers.

Sources 1

The new generation of hardware makes SI thinking several times cheaper in energy terms, which means there will be more agents and they will work longer.

SemiAnalysis: new Nvidia Vera Rubin servers deliver up to 7 times more tokens per megawatt than Blackwell

The new generation of hardware makes SI thinking several times cheaper in energy terms, which means there will be more agents and they will work longer.

On pre-release software, Vera Rubin NVL72 running DeepSeek V4 Pro (1.6 trillion parameters) showed up to a 7x advantage over Blackwell in tokens per megawatt, more than the three times Huang publicly promised; at 100 tokens per second per user, it delivers 59.4 million tokens/s per MW versus 28.5 million for GB300. Under a total-cost-of-ownership model at 80 tokens/s, Rubin gives 10 times more tokens per dollar. Caveats: different inference engines were used in the comparison, and at high speeds the advantage narrows to 2.72 times.

Compute has become so scarce that labs take it from anyone, even from a video platform that hasn't finished building its data center.

Anthropic rents $13.7 billion of compute from Rum Group, the former Rumble, which is tied to Trump's circle

Compute has become so scarce that labs take it from anyone, even from a video platform that hasn't finished building its data center.

The six-year contract covers capacity in a data center under construction in Maysville, Georgia, with a 120 MW connection and the possibility of expanding to 180 MW. Rum Group is the renamed video platform Rumble, which bought the German operator Northern Data in June; it does not yet have financing for the build. Anthropic received an option to buy up to 51 million Rum Group shares at one cent each.

An SI slowdown proposed by the leaders themselves may turn into a closed club where winners write the rules, and smaller players are already loudly against it.

Cohere's chief calls big labs' joint slowdown a "cartel"

An SI slowdown proposed by the leaders themselves may turn into a closed club where winners write the rules, and smaller players are already loudly against it.

Aidan Gomez said in an essay and a Bloomberg interview that Amodei's roadmap, with a narrow antitrust exemption allowing several labs to agree on standards and pace, is "a cartel by another name": it slows everyone while preserving the leaders' advantage. Gomez nonetheless supports independent testing modeled on financial licensing and believes the rules should be written by a broad range of participants, not OpenAI and Anthropic.

An SI assistant needs eyes: OpenAI is assembling a team so that its future devices see the world as well as an iPhone camera.

OpenAI buys "smart" camera developer Glass Imaging for $300+ million, another step toward its own device

An SI assistant needs eyes: OpenAI is assembling a team so that its future devices see the world as well as an iPhone camera.

According to the WSJ, OpenAI will pay more than $300 million for a startup founded by former Apple engineers Ziv Attar and Tom Bishop, the authors of the iPhone's Portrait mode. Glass Imaging improves a shot with a neural network at the moment of capture, adapting to the specific camera; its technology already powers the zoom in Honor smartphones. The purchase is linked to the development of OpenAI's own devices after it acquired Jony Ive's company for $6.5 billion.

Part of the work is already done by machines, and companies are moving money from salaries into data centers: SI is becoming the foundation of the industry.

Global tech companies will lay off almost 130,000 people in 2026 — up to a third of the cuts because of SI

Part of the work is already done by machines, and companies are moving money from salaries into data centers: SI is becoming the foundation of the industry.

According to Layoffs.fyi, cited by CNews, that is more than the 122,600 in 2025; 6,300 jobs were cut in the first 10 days of September alone. Oracle cuts the most (21,000, 13% of its staff), followed by Amazon, Dell, Meta and Microsoft; from a quarter to a third of the layoffs are tied to automation and the adoption of SI. Meanwhile Alphabet, Amazon, Microsoft and Meta will spend about $725 billion on infrastructure, mostly for SI, over the year.

Sources 1

Deep research by SI is getting cheaper not only thanks to new models but thanks to smart "packaging" around them, which means it will soon go mainstream.

Yandex cuts GPU use severalfold in Alice AI's Research mode without retraining the model

Deep research by SI is getting cheaper not only thanks to new models but thanks to smart "packaging" around them, which means it will soon go mainstream.

The team rewrote the scaffolding around the model: processing of search snippets was handed to a trained classifier instead of an LLM, tools are called only if they advance the research, and planning was moved to a lighter model. Over three months the cost of a single query fell 6.6 times; on the FRAMES and SimpleQA benchmarks the mode scores 75 and 81. Individual scaffolding improvements each gave a gain of 5 to 9 percentage points.

Sources 1

Follow on Telegram