Show only each day's banner story

OpenAI confirms it, Anthropic and Google have been negotiating on SI safety for weeks, and backs a federal law on external model audits

SI leaders are building rules for themselves and are ready to admit outside auditors: if the industry agrees first, it will set the pace and framework rather than have them imposed from above by officials.

OpenAI's head of global policy Chris Lehane confirmed that the company has been discussing common safety measures with Anthropic and Google DeepMind for several weeks; according to TechCrunch, the three labs are preparing an industry standards body. Lehane said they "don't need" antitrust exemptions from the government for such coordination. At the same time, OpenAI for the first time backed a federal requirement for mandatory independent verification: the bipartisan FRONTIER Act from Representatives Jay Obernolte (R) and Lori Trahan (D) requires an outside organization to confirm that the largest developers' safety protocols keep "an acceptable level of catastrophic risk," and also requires them to publish model reports and report incidents. OpenAI did not endorse the bill as a whole.

Sources 3

techcrunch.comcbsnews.compolitico.com

The US government does not intend to slow SI with laws: safety remains the companies' own business, and the race for leadership continues at full speed.

White House: SI threats should be handled by the private sector, not the government

The US government does not intend to slow SI with laws: safety remains the companies' own business, and the race for leadership continues at full speed.

National Economic Council director Kevin Hassett said on CNBC that SI risks are "solvable problems, and the private sector is the right place to solve them"; for irresponsible corporate behavior, he said, ordinary law enforcement mechanisms should apply. The statement came the day after Trump called concerns about SI safety a "conspiracy." Hassett nonetheless endorsed Amodei's idea of independent observers inside companies, and Treasury Secretary Bessent, according to the outlet, is coordinating a review of Anthropic's Mythos model.

Sources 2

cnbc.comcryptobriefing.com

Another open frontier-level model from China: agentic systems capable of running entire projects are becoming available to anyone who can download the weights.

Shanghai SI lab quietly releases open 744-billion-parameter agentic model, with the report published after the release

Another open frontier-level model from China: agentic systems capable of running entire projects are becoming available to anyone who can download the weights.

Atria Dawn Preview is built on the GLM-5.2 base model (744 billion parameters, mixture-of-experts architecture), with a 256 thousand-token context and weights under the MIT license. The repository appeared on September 11 without an announcement, an FP8 version was added on the 12th, and on the 14th a technical report by more than 140 authors came out on arXiv. In a study of 769 tasks with 56 participants, people rated roughly a third of the completed tasks as impossible without SI. All results are claimed by the authors themselves; on SWE-bench Pro the model trails Claude noticeably (59.6 versus 74.7).

Sources 2

aiweekly.cogithub.com

The main supplier of SI's "brains" is openly against brakes: while people argue about slowing down, hardware will keep accelerating the whole industry.

Jensen Huang: "We don't need new laws" because SI safety is an engineering problem, not a legal one

The main supplier of SI's "brains" is openly against brakes: while people argue about slowing down, hardware will keep accelerating the whole industry.

Speaking at Salesforce's Dreamforce conference, the Nvidia chief said SI is "hardware and software" controlled by engineering methods, and that market forces and product liability laws are enough: "We don't need new laws. We don't need new rules." He urged the industry to "run as fast as possible" while making products safe. TechCrunch notes that the position matches Nvidia's financial interests and contrasts it with Satya Nadella's call for global safety standards.

Sources 1

techcrunch.com

Mutual checking is a way to make safety part of competition rather than a government barrier: rivals know best where to look for weak spots.

Musk proposes that US and Chinese labs check each other's models before release

Mutual checking is a way to make safety part of competition rather than a government barrier: rivals know best where to look for weak spots.

At the All-In Summit in Los Angeles, Elon Musk proposed a shared "test bed": xAI, OpenAI, Anthropic, Google, Meta and "three or four leading Chinese companies" would let competitors test their models before release. "Instead of grading your own homework, competitors would check it and raise the alarm if something is wrong," he said, adding that he was confident China would agree. No lab has yet confirmed participation.

Sources 2

cnbc.comtechstartups.com

Voices from inside the labs about risks are getting louder, but the argument is no longer about whether SI is powerful; it is about how fast it will become even more powerful.

Former DeepMind safety researcher: "SI could kill us all, and time is running out"

Voices from inside the labs about risks are getting louder, but the argument is no longer about whether SI is powerful; it is about how fast it will become even more powerful.

Bilal Chughtai, who worked on the AGI safety team at Google DeepMind and left in July, publicly explained his departure two months later: "I sincerely believe that SI is capable of killing us all and that we may not have enough time to avoid it." According to him, model capabilities are growing faster than the understanding of how to control them; he called for coordination to stop the companies' "manic race" and for a pace "that society can cope with." The post coincided with industry debate about slowing down after Amodei's essay.

Sources 2

thenextweb.combloomberg.com

Europe is again the first to try to put SI interlocutors within limits; for children, access to smart assistants will become noticeably narrower.

EU Kids Act draft leaks: social media from age 15, SI companions for children will not be able to switch on by themselves

Europe is again the first to try to put SI interlocutors within limits; for children, access to smart assistants will become noticeably narrower.

According to the European Commission's leaked draft, an independent social media account will be available from age 15, with a restricted mode under parental control for ages 13-14. The law also covers SI chatbots and companions, "virtual tools that may give minors advice on mental health and development": they are subject to the same age rules and the "safety by default" principle, and design that causes emotional dependence is restricted. Companies themselves, not regulators, must prove compliance; violations.

Sources 1

euronews.com

Talking to SI by voice in any of almost a hundred languages is becoming commonplace, and the language barrier to knowledge keeps falling.

Google releases Gemini 3.8 Live voice models: 97 languages with switching in the middle of a conversation

Talking to SI by voice in any of almost a hundred languages is becoming commonplace, and the language barrier to knowledge keeps falling.

There are two models: Gemini 3.8 Live, cheap and scalable, grounded in the camera image, and Gemini 3.8 Live Extended Thinking, for complex multi-step tasks. Extended Thinking scored 82.6 on Artificial Analysis's speech-to-speech quality index and 97.7% on Big Bench Audio. The models are available to developers in the Gemini API and AI Studio, and to users in Search Live, Gemini Live, Workspace, Gmail and Keep; all generated audio is watermarked with SynthID.

Sources 1

blog.google

SI is no longer just reading scientific papers and is starting to learn directly from experiments: labs that improve themselves will speed up the discovery of new materials.

Periodic Labs embeds a trillion-parameter model directly into physical labs, and it beats Fable 5.1 and GPT-6 Astra at analyzing X-ray patterns

SI is no longer just reading scientific papers and is starting to learn directly from experiments: labs that improve themselves will speed up the discovery of new materials.

The Periodic Neon model (1 trillion parameters) was fine-tuned and run through reinforcement learning on fresh data from the company's own materials science labs, closing the "experiment, model, experiment" loop. According to Periodic Labs, in X-ray diffraction analysis, a task that used to require hours of expert work, Neon outperforms Claude Fable 5.1 and GPT-6 Astra. The company calls this a first step toward autonomous discovery in materials science.

Sources 1

periodic.com

A swarm of SI agents already does in a day what would take humans years: software development is getting cheaper by orders of magnitude.

1,393 Claude agents rewrite a million-line codebase in 19 hours for about 1% of the cost of a human team

A swarm of SI agents already does in a day what would take humans years: software development is getting cheaper by orders of magnitude.

Nous Research launched 1,393 subagents on Claude Fable 5.1 (up to 218 at once) to refactor its Hermes project. The volume of non-test Python code shrank by 34.4%, from 1,063,826 to 698,365 lines; files longer than 5,000 lines went from 37 to 6, and functions longer than 300 lines from 192 to 2. The main run cost about $19.3 thousand (about $25 thousand with follow-up work), whereas the work of a small human team is estimated at $150 thousand to $1.8 million and from 2 months to 2 years.

Sources 1

nousresearch.com

The justice system, one of the most conservative spheres, is also moving to SI: the court's routine work will soon be done by neural networks, and the judge will become the reviewer.

Russia's Supreme Court approves a plan: in three years SI will prepare draft court rulings

The justice system, one of the most conservative spheres, is also moving to SI: the court's routine work will soon be done by neural networks, and the judge will become the reviewer.

Supreme Court Chairman Igor Krasnov approved a concept for introducing SI into court proceedings: a judicial competence center for SI is to be created in the third quarter of 2026, and a pilot starts in the first quarter of 2027. By 2030, 95% of judges are to use SI services, neural networks are to prepare at least 90% of hearing transcripts and 30-50% of draft procedural documents, and the time to resolve routine cases is to fall by 20-30%. The final decision remains with the judge; retired judge Sergei Pashin warned that SI may cite nonexistent norms.

Sources 2

gov.cnews.ruvbr.ru

Enterprise software is ceasing to be a separate program with windows: people will work with it through the SI assistant they already use.

Salesforce builds its CRM directly into Claude and Slack and shows its own model, Koa

Enterprise software is ceasing to be a separate program with windows: people will work with it through the SI assistant they already use.

At the opening of Dreamforce, Marc Benioff presented AIforce, a layer that brings Salesforce's data, processes and access rights into Claude, Slack and other SI tools so that agents and employees don't need to open the CRM itself; data is not stored with model providers. It includes Claudeforce with 37 ready-made skills for sales, Slackforce and Agentforce Coworker, which 100 thousand users activated in 35 days. Also shown is Koa, the company's first "reasoning" CRM model, based on Nvidia Nemotron, currently in pilot.

Sources 2

salesforce.comsalesforcebreak.com

An infrastructure of trust is growing around SI agents, with audits and certificates that will let them into real business faster.

Staff from early Anthropic and METR raise $40 million to "certify" SI agents before they are cleared for work

An infrastructure of trust is growing around SI agents, with audits and certificates that will let them into real business faster.

The Artificial Intelligence Underwriting Company (AIUC) of Rune Kvist and Rajiv Dattani closed a $40 million Series A led by Ribbit Capital ($55 million in total). The company runs agents through about 5,000 tests for jailbreaks, hallucinations and data leaks under the AIUC-1 standard, modeled on SOC 2, and issues a security report. Its clients include Cursor, Lovable, Harvey and ElevenLabs. "The smarter the SI, the harder it is to deploy and control," says Kvist.

Sources 1

techcrunch.com

Even the most profitable companies in the world are ready to sacrifice profit for SI: the bet on it matters more than today's money.

ByteDance's half-year profit falls to $20 billion as SI spending eats it up, even though revenue grew 30%

Even the most profitable companies in the world are ready to sacrifice profit for SI: the bet on it matters more than today's money.

According to The Information, the net profit of TikTok's owner in the first half of 2026 fell by a single-digit percentage year over year, to $20 billion, even though revenue grew about 30%, to $120 billion. The reason is large investment in SI infrastructure: cloud, the Doubao and Seedance models, and its own inference chip. The company is private.

Sources 2

aiweekly.cotheinformation.com

Humanoids are stepping out from behind the fences: a robot that can be trusted to work side by side with a person is a step toward mass physical automation.

Agility Robotics shows Digit 5 humanoid that can work alongside people without safety fences

Humanoids are stepping out from behind the fences: a robot that can be trusted to work side by side with a person is a step toward mass physical automation.

The robot, 180 cm tall and weighing 129 kg, lifts up to 23 kg (40% more than its predecessor), runs for about 90 minutes and charges in about 9 minutes, and has swappable grippers. The main thing is the new safety architecture: SI algorithms and a set of sensors continuously track people nearby, and the robot itself steps around them, stops or sits down. Orders for Digit 5 already exceed $300 million, and broad availability is expected in late 2027.

Sources 2

prnewswire.comtherobotreport.com

SI capabilities are becoming a separate product: people are now being offered to pay for more "intelligence" in familiar social networks.

Meta launches paid Meta One subscription with expanded SI limits across all its apps

SI capabilities are becoming a separate product: people are now being offered to pay for more "intelligence" in familiar social networks.

Subscriptions for individual apps start at $2.99 a month (WhatsApp Plus), bundles for individual users cost $7.99 (Core) and $19.99 (Premium), and plans for businesses and creators run from $14.99 to $499. Subscribers get more generations in Meta AI, including video on Muse models, and business plans add analytics, scheduled posts and expanded access to Meta Business Agent. Meta promises that basic features and Meta AI will remain free.

Sources 1

about.fb.com

Generative video is learning to "feel" time: another step toward clips indistinguishable from real footage, and in open access.

Sber open-sources Time Adapter, a module that teaches video neural networks to control the speed of action in a frame

Generative video is learning to "feel" time: another step toward clips indistinguishable from real footage, and in open access.

The module, under 1% the size of the main model, separates two previously merged parameters, frame rate (from 15 to 60 fps) and the pace of the action itself, so slow motion or speed-up can be produced without retraining the base model. It was trained on 40 thousand clips; together with Kandinsky 5.0 Video Lite, naturalness of motion rose by 29%, visual quality by 8% and text alignment by 19%. The code is posted on Hugging Face under the MIT license, and the work was presented at ICML.

Sources 2

banks.cnews.ruhabr.com

SI agents have become such familiar visitors to websites that attackers are already disguising themselves as them: traffic can no longer be trusted by a bot's "name" alone.

Hackers attack Russian company websites by disguising themselves as SI bots ChatGPT, DeepSeek and Claude

SI agents have become such familiar visitors to websites that attackers are already disguising themselves as them: traffic can no longer be trusted by a bot's "name" alone.

Solar WAF, a service of the Solar group, has recorded malicious requests with a spoofed DeepSeekBot User-Agent header since August 12, and with ChatGPT-User, Perplexity-User, Claude-User and GrokBot since August 27. The bet is that site owners filter SI assistants' traffic more leniently as a valuable source of customers. Among such requests, 53% are DNS rebinding attempts, 12% data exfiltration, 4% path traversal, and the rest SQL injections and other techniques.

Sources 2

anti-malware.rusafe.cnews.ru
Previous day
Follow on Telegram