Show only each day's banner story

Google DeepMind releases AlphaGenome 2 to predict mutation effects in specific cells

Researchers can now find mutations that act only in specific cells more quickly and choose disease experiments more precisely.

AlphaGenome 2 takes a DNA sequence of up to one million letters together with a description of a cell type or state and predicts gene expression, histone modifications, chromatin accessibility and DNA methylation. The model was trained on more than 107 million individual human and mouse cell profiles across 68 tissues. Unlike the first version, which averaged measurements across tissues, it can work with new cellular states without fine-tuning and reconstruct missing measurements. According to DeepMind, on the TenK10K dataset the model identified 30% of causal variants versus 16% for the fine-tuned first version, and determined the direction of gene effects in 43% of cases versus 23%. In a test, it reproduced a known mechanism of inflammatory bowel disease: a variant increases ETS2 gene activity specifically in gut macrophages. AlphaGenome 2 is available for non-commercial research through the API, with commercial access in private preview. It is a research tool, not an approved clinical or diagnostic system.

SI agents now connect the entire path from intelligence data to preparation of the operator's decision in one European system, while the military retains control of the data and every combat decision.

Helsing unveils Clarity, an SI-agent system for military intelligence and target preparation

SI agents now connect the entire path from intelligence data to preparation of the operator's decision in one European system, while the military retains control of the data and every combat decision.

Clarity combines satellite and sensor data, drone feeds, open sources and reports. Hundreds of specialized agents link evidence in parallel, identify military equipment and positions, present them to an operator and prepare data for the selected weapon system. The models were trained on real operational, synthetic and simulated data. According to Helsing, the platform runs on the customer's infrastructure, follows MOSA principles and NATO standards, and can connect to weapons from different manufacturers. The operator reviews the evidence, selects each object and separately authorizes every transfer of data to a weapon; each decision is recorded against a named person. The company says a recommendation is produced only when independent sources agree and system confidence is high. The Decoder notes that such organizational safeguards do not by themselves eliminate data-fusion errors or perfunctory operator review.

The new Qwen generation can already be tested on long documents, video and agent tasks, although this is a temporary cloud preview rather than a full open release.

Alibaba opens preview access to Qwen4-Max with a one-million-token context

The new Qwen generation can already be tested on long documents, video and agent tasks, although this is a temporary cloud preview rather than a full open release.

Qwen4-Max-Preview is the first available model in the new Qwen4 generation. It accepts text, images and video, supports reasoning, function calling, web search, structured responses and APIs compatible with OpenAI and Anthropic. Its context is 1 million tokens, with a combined reasoning and final-response limit of 393,216 tokens. The preview is available only through a Token Plan subscription, as well as in QwenWork, Qoder, Qwen App and QwenStudio. API testing runs from November 11 to November 25 in UTC+8, after which this frozen version will be retired. Pay-per-token access is not available. Alibaba describes a focus on programming, office work, interfaces and complex agent tasks, but its publications provide no comparative model results.

The SI-generated proof is now available to everyone for machine and scientific scrutiny; an important part of the fundamental problem is confirmed, but not the entire prize problem.

David Gross's group publishes and confirms formal verification of OpenAI's Yang–Mills proof

The SI-generated proof is now available to everyone for machine and scientific scrutiny; an important part of the fundamental problem is confirmed, but not the entire prize problem.

The OpenAI-appointed physics group led by Nobel laureate David Gross released the model-generated Lean code and concluded that the result proves the existence of pure Yang–Mills theory in the formulation accepted in physics and mathematics. Lean verification replayed 100,168 project declarations for 17 main theorems and found only standard logical axioms. The result addresses the first part of the Millennium Prize Problem: constructing a mathematically rigorous theory underlying the Standard Model. The second part — proving a positive mass gap — remains unsolved, so the Clay prize problem is not closed. Gross and his colleagues explicitly say they do not yet fully understand the proof; publishing the code allows independent researchers to examine it and verify it more deeply.

Agents will be able to select tools and check hundreds of conditions almost instantly without spending time and money generating unnecessary text.

Google releases Gemini Decision 1 for fast answers to closed sets of questions

Agents will be able to select tools and check hundreds of conditions almost instantly without spending time and money generating unnecessary text.

The new Decisions API accepts a shared context of up to 256,000 tokens — text, images, audio, video, or PDF — and up to 1,000 independent questions. Instead of generating an answer token by token, the model returns probabilities in a single pass for yes/no options, a choice among 2–255 answers, and an ordered rubric score, including confidence from zero to one. Google reports latency of 150–250 ms and a price of $0.10 per million input tokens with no charge for output tokens. According to the company's internal tests, this is nearly five times cheaper than Gemini 3.5 Flash-Lite across a shared suite of routing, reasoning, ranking, and multimodal understanding tasks. The API is available in Google AI Studio, including a free tier for prototypes; Google already uses it to select models in Gemini Notebook, accelerate design in Stitch, and run safety checks for MCP tools in Search.

A company central to SI infrastructure faces the risk of a sudden leadership change, although the board has not yet made a decision.

Nvidia board discusses ousting Jensen Huang after reports of a second affair

A company central to SI infrastructure faces the risk of a sudden leadership change, although the board has not yet made a decision.

The Wall Street Journal reported that the board is preparing to discuss the future of Nvidia's co-founder and CEO and may remove him from office. Shortly beforehand, The New York Times reported that Huang had a relationship with healthcare vice president Kimberly Powell, who reports directly to him. No official decision on his dismissal had been announced at the time of publication. This develops an independent investigation launched earlier into another relationship acknowledged by Huang, with Megaspeed chief Alice Huang, whom US intelligence believes is linked to Chinese intelligence services. The board previously said it had seen no evidence that company secrets or national security had been compromised. Nvidia shares lost about 8–9% in after-hours trading following the new reports. Huang has led the company for 33 years; sources note the absence of an obvious successor and approximately 60 direct reports.

Different enterprise agents receive shared verifiable definitions and access to permitted data instead of searching the entire company independently.

Databricks opens Genie Ontology and business context to third-party SI agents

Different enterprise agents receive shared verifiable definitions and access to permitted data instead of searching the entire company independently.

Genie Ontology is now generally available: it automatically extracts business rules from dashboards, SQL queries, notebooks and external BI models, while employee-approved definitions take precedence over learned ones. Genie Search is also generally available and indexes documents, tickets and wikis together with the source systems' access permissions. Plugins carry this context into ChatGPT, Claude, Microsoft Copilot Cowork and Grok Bot. An agent can hand an entire question to Genie and receive SQL analysis with citations, or request only the missing information and continue working with its own tools. This supplements the autonomous Genie launched on the same day.

New inference capacity can be deployed next to enterprises in weeks using electricity already supplied to the site instead of waiting years for a new connection.

Torus raises $500 million and launches modular data centers using spare capacity in existing grids

New inference capacity can be deployed next to enterprises in weeks using electricity already supplied to the site instead of waiting years for a new connection.

Torus Foundry combines computing, energy storage, cooling and networking in prefabricated modules rated at 1–20 MW. They are installed behind the meter at enterprise sites and use spare capacity at existing transformers without a new grid connection. Torus Reserve flywheels and batteries smooth demand peaks and provide backup power. The company promises launch 12–16 weeks after contract signing, about a week of site preparation and one or two days of installation. The systems are designed for inference and already use thousands of GPUs; Torus plans more than 47,000 by the end of 2027. Hydra Host is the first named customer. Magnetar led the $500 million round with participation from ICONIQ, Origin Ventures, Ridgeline Partners and Trinity Capital. The funds will expand the GigaOne factory and Foundry deployments in the United States; the production line is designed for 500 MW of IT load with plans to expand to 1 GW.

The largest contract chipmaker is becoming a co-owner of two leading laboratories and tying demand for its capacity even more closely to the development of their models.

TSMC to invest up to $1 billion in Anthropic and OpenAI shares

The largest contract chipmaker is becoming a co-owner of two leading laboratories and tying demand for its capacity even more closely to the development of their models.

TSMC's board approved purchases of up to $500 million in Anthropic common shares at the IPO price and up to $500 million in OpenAI preferred shares. Separately, the company will create an investment vehicle with up to $500 million in capital for investments in semiconductor companies and related fields. The board also approved a capital budget of about $40.93 billion for advanced manufacturing expansion, fab construction and research. According to Focus Taiwan, the company links expected demand partly to the development of agentic SI applications. Quarterly dividends were raised to a record NT$9 per share.

Universities will have to test independent knowledge where SI can do the work for a student while also preparing graduates to use these tools thoughtfully.

England's regulator tells universities to redesign assessment for SI

Universities will have to test independent knowledge where SI can do the work for a student while also preparing graduates to use these tools thoughtfully.

Following a joint study with Advance HE of more than 300 students and staff at 127 institutions, the Office for Students said widespread SI use makes some coursework almost impossible to assess reliably. The regulator reminded universities that they must create valid and reliable ways to assess knowledge or risk breaching registration conditions and facing sanctions. Recommendations include in-person written and oral exams, presentations, assessment of the work process, and critiquing and editing SI output. The OfS is not requiring universities to remove the technology from teaching: institutions are encouraged to teach responsible SI use while separating it from assignments intended to test a student's own skills. Universities UK said many institutions have already changed assessment formats.

The pace of SI infrastructure construction increasingly depends on power grids: on-site generation speeds up launch but creates a large long-term source of emissions.

BloombergNEF estimates proposed on-site gas generation for data centers at 181 GW

The pace of SI infrastructure construction increasingly depends on power grids: on-site generation speeds up launch but creates a large long-term source of emissions.

In BloombergNEF's published study of projects in the United States, the United Kingdom and mainland Europe, data-center developers are considering on-site gas power plants with a combined capacity of 181 GW because of queues for national grid connections. About 90% of this capacity is in the United States. If every plan is implemented, emissions there are estimated at about 345 million metric tons of CO₂ equivalent per year. BloombergNEF forecasts electricity consumption by SI data centers to almost quadruple by 2035, reaching 2,354 TWh, or about 7% of global demand. The analysts emphasize that gas generation described as temporary could become long-lived because of the economics of plants already built. Faster grid connections and additional cleaner generation will determine the overall climate impact.

The two leading SI powers now have a direct way to warn each other about dangerous model failures and align testing practices without pausing technology development.

United States and China create a notification channel for serious incidents involving frontier models

The two leading SI powers now have a direct way to warn each other about dangerous model failures and align testing practices without pausing technology development.

US Treasury Secretary Scott Bessent described a new dialogue mechanism on superintelligence: the two sides will exchange pre-release testing practices for frontier models, discuss keeping humans in charge of decisions and warn each other about serious incidents. He said the arrangement is intended to reduce shared risks and misunderstandings between the two systems. The mechanism does not limit model development: Bessent emphasized that cooperation will continue alongside technological competition. The countries' leaders are next expected to meet on the sidelines of the APEC summit on November 18–19. China's ambassador to the United States, Xie Feng, called joint work necessary to improve SI governance.

Sources 1

If the claimed result holds up in practice, SI will sharply reduce the cost of switching between chips and weaken laboratories' dependence on one manufacturer's software ecosystem.

Anthropic says its own model ported Claude to AMD MI355X in three days

If the claimed result holds up in practice, SI will sharply reduce the cost of switching between chips and weaken laboratories' dependence on one manufacturer's software ecosystem.

Anthropic co-founder and Chief Compute Officer Tom Brown said in an investor presentation that the Mythos model ported or wrote almost all the compute kernels needed to run Claude on AMD MI355X with almost no human involvement. The company says usable performance was achieved in about three days instead of the six to twelve months required for traditional adaptation to a new accelerator; The Decoder emphasizes that these figures have not been independently verified. Anthropic trains and runs Claude on four platforms — AWS Trainium, Google TPU, Nvidia GPU and AMD GPU — and distributes tasks among them as a shared computing pool. The company also plans to deploy up to two gigawatts of AMD Instinct MI455X capacity in Helios systems, with the first 1 GW phase expected in the first half of 2027.

A breach at one SI infrastructure provider spread to another service's user conversations, so secrets entered in chats need to be changed even without a breach of Notion itself.

Notion reports possible exposure of Notion AI chats through the Fireworks contractor breach

A breach at one SI infrastructure provider spread to another service's user conversations, so secrets entered in chats need to be changed even without a breach of Notion itself.

Notion Labs notified users that an incident at contractor Fireworks.ai may have exposed Notion AI chat records from the Insights analytics feature. According to the company as relayed by ITmedia, less than 1% of workspaces were affected; most records had no label identifying a specific user, but the text itself could contain information about them. Notion contacted affected users and recommended changing passwords, API keys and other secrets if they were entered in potentially exposed chats. This is a new consequence of the previously disclosed Fireworks breach, which affected its customers' data.

Anthropic shares at IPO are becoming more accessible to retail investors outside the United States, although the final price and demand outcome are still ahead.

Anthropic opens retail IPO subscription in Australia until November 17

Anthropic shares at IPO are becoming more accessible to retail investors outside the United States, although the final price and demand outcome are still ahead.

Retail investors in Australia can apply through CommSec, Bell Potter, Macquarie, Morgan Stanley Wealth Management, Stake and other participating brokers. Subscription opened on the morning of November 11 and closes at 17:00 AEDT on November 17. New Class A common shares are offered under an Australian prospectus; Anthropic does not plan a separate listing on the Australian Securities Exchange. The main Nasdaq offering under ticker ANT retains a price range of $70–78 per share and potential proceeds of up to $64.4 billion, with a parallel private placement of up to $26 billion. Australia joins Japan, Canada and parts of Europe among the small number of markets outside the United States where retail investors will be offered direct access to the shares.

Sources 1

Cross-language conversations now preserve not only a person's voice but also their changing manner of speech, so less meaning and emotion are lost.

DeepL Aure preserves intonation, emotion and speaking pace in real-time translation

Cross-language conversations now preserve not only a person's voice but also their changing manner of speech, so less meaning and emotion are lost.

Aure continuously analyzes the source audio and translated text, so the translated voice changes with the speaker: it preserves emphasis, pauses, pace and sarcasm rather than fixing intonation from an opening sample. During training, DeepL separates voice identity, melody and speech content so they can be transferred into another language independently; reinforcement learning is used to preserve accuracy and reduce accent. The model is automatically enabled in DeepL Voice for online meetings, in-person conversations and the API. At launch it supports English, German, Spanish, French, Italian, Japanese, Portuguese and Korean. DeepL says meeting recordings and translations are not stored permanently and are deleted from memory when the session ends.

Restrictions on direct shipments do not eliminate access to advanced computing through chip resale and server rental within China.

Defense-linked Chinese university buys two Nvidia H200 chips through a local reseller

Restrictions on direct shipments do not eliminate access to advanced computing through chip resale and server rental within China.

Reuters examined a procurement record published on October 21 by Northwestern Polytechnical University: two H200 chips for its management school cost 874,700 yuan, about $123,000, and were bought from Shaanxi Bosuan Network Technology rather than directly from Nvidia. The university has been on the US Entity List since 2001, so export licenses for it are normally denied; permission for selected H200 shipments to China since January did not remove that regime. Procurement documents also revealed five university contracts to rent computing on H200 and other restricted Nvidia processors from Chinese companies. In total, Reuters reviewed 26 procurement and corporate records showing Chinese universities and research organizations acquiring such chips or renting capacity through domestic suppliers.

Russian organizations will find it harder to use Claude through foreign teams and intermediaries, so workflows will need to move to accessible models.

Anthropic to extend Claude restrictions to employees of Russian companies abroad

Russian organizations will find it harder to use Claude through foreign teams and intermediaries, so workflows will need to move to accessible models.

From November 12, Anthropic is clarifying how regional restrictions apply: Claude cannot be used by organizations registered or headquartered in unsupported countries, including Russia, or by organizations under majority direct or indirect control of people or entities from those countries. An employee's physical location in a supported country will no longer remove the restriction. According to Vedomosti, the rules cover Claude.ai, Claude Code, Cowork, the API and models accessed through cloud platforms and authorized intermediaries. This clarifies existing restrictions rather than introducing the first ban on access from Russia.

Sources 1

Schoolchildren will be systematically taught not only to use SI but also to verify its answers and recognize technology-related risks.

More than 1,000 Russian schools begin piloting a superintelligence and information security course

Schoolchildren will be systematically taught not only to use SI but also to verify its answers and recognize technology-related risks.

The extracurricular course "Superintelligence and Information Security" is being piloted for students in grades 5–11. Pilot schools will send feedback to developers by the end of 2026; the course is planned for all Russian schools from September 1, 2027. Weekly lessons will include videos, practical work and review questions. The program covers safe SI use, recognizing fakes and digital hygiene. Textbooks are being prepared with Russian IT companies, alongside a separate platform for practical assignments. Regional coordinators will help schools with teaching methods, scheduling and assessment of their technical infrastructure.

Training environments have become a product in their own right and a channel for transferring know-how between countries, while export rules still primarily regulate chips rather than this resource.

ChinaTalk estimates US sales of RL environments to Chinese labs at about $500 million

Training environments have become a product in their own right and a channel for transferring know-how between countries, while export rules still primarily regulate chips rather than this resource.

According to a ChinaTalk investigation by Aqib Zakaria and Quinn Ennis, Mercor, Surge AI, AfterQuery, Turing and other US providers sell Chinese laboratories about $500 million worth of training data. This includes reinforcement-learning environments: simulations of programming, financial and office work with verifiable outcomes. The authors identified at least nine US companies with Chinese customers. The investigation says buyers choose products already used by Anthropic or OpenAI to avoid repeating expensive development and unsuccessful training runs, which is especially important given computing and memory constraints. ChinaTalk notes that export rules barely cover such environments and proposes extending existing controls to dual-use variants in biology and cybersecurity. At the same time, the authors argue that borrowing others' work leaves Chinese labs at least several months behind their US counterparts, while frontier companies increasingly build the most sensitive environments in-house.

Simple personal agents can run locally on a small model, keeping data on the device and avoiding continuous charges for cloud requests.

Tether releases BotPsy, a 0.8-billion-parameter agent model for on-device use

Simple personal agents can run locally on a small model, keeping data on the device and avoiding continuous charges for cloud requests.

BotPsy is built on Qwen3.5-0.8B and trained to call tools for search, documents, calendars, messaging and device control. According to Tether AI Research, the model leads on five of six agent benchmarks among compared systems with up to 1.2 billion parameters and beats Qwen3.5-2B-Think on five; its average across six tests rose from 24.7 for the base model to 45. The largest improvements concern multi-step conversations and refusing to call an unsuitable tool. Weights are being published under Apache 2.0 in safetensors and GGUF formats with bf16, Q8 and Q4 variants. The developers warn that the model is intended for low- and medium-difficulty tasks, makes mistakes and should not autonomously make decisions in medicine, finance or other sensitive fields. Open-ended search remains a weakness: even BotPsy's best BrowseComp-Plus result is only 4.5 points.

Inexpensive speech synthesis increases the cost of verifying emergency calls, but services cannot simply ignore a threat report even when forgery is suspected.

Synthetic voices make hoax emergency calls harder to identify

Inexpensive speech synthesis increases the cost of verifying emergency calls, but services cannot simply ignore a threat report even when forgery is suspected.

NPR describes a false bomb report made in August to a school in California's Yuba County: police responded overnight, classes were canceled, and the caller's voice was generated by SI. Dispatcher associations APCO and NENA say natural pauses, intonation and background sounds make such false reports harder to recognize. In a June survey of 766 emergency communications professionals, 14% said they had encountered attacks using SI. Dispatchers are being trained to ask about details that cannot be quickly extracted from public sources, but services will continue to treat every report of a threat to life as real. Such hoax calls divert police and rescuers from other incidents and can lead to injuries and deaths.

Sources 1

Purchases through personal SI agents have run into responsibility for payments and mandatory acknowledgements, so the airline is leaving the final step to a person for now.

United Airlines bans third-party SI agents from booking flights independently

Purchases through personal SI agents have run into responsibility for payments and mandatory acknowledgements, so the airline is leaving the final step to a person for now.

The restriction covers United's website and app, as well as bookings through travel agencies and aggregators. Agents can show users flights but cannot complete a purchase on their behalf. The airline cites the risk of exposing payment and account credentials, unintended purchases and account changes, and the need for passengers personally to acknowledge mandatory aviation declarations. United is willing to reconsider once adequate safeguards exist. CEO Scott Kirby also said he does not want to create another intermediary between the airline and its customers. eBay, banks, and other companies previously introduced similar restrictions; Southwest, by contrast, offers a plugin for booking through ChatGPT.

Teams will be able to integrate Codex into their existing AWS rules instead of creating a separate access and accounting system for SI development.

OpenAI Codex gains support for models through Amazon Bedrock

Teams will be able to integrate Codex into their existing AWS rules instead of creating a separate access and accounting system for SI development.

Codex continues to run on the developer's computer, but requests to supported OpenAI models can pass through Bedrock and AWS's existing identity, access, and observability mechanisms. AWS describes three paths: direct access through IAM Identity Center, the managed AgentCore Gateway with routing and Guardrails, or a custom LLM gateway to control requests. The examples use GPT-5.6 Sol; the availability of Sol, Terra, and Luna depends on the region and inference path. AWS has published an open-source repository with configurations for Native AWS Access, AgentCore, LiteLLM, Portkey, and OpenTelemetry. The recommended rollout starts with a small least-privilege pilot, a separate two-user isolation check, auditing, and a way to revoke access.

Voice agents are becoming a corporate system with their own memory and training on real interactions, rather than merely a speech synthesizer.

ElevenLabs will customize models on customer conversations and opens ElevenAI workplace app

Voice agents are becoming a corporate system with their own memory and training on real interactions, rather than merely a speech synthesizer.

In the coming months, companies will be able to work with ElevenLabs researchers to fine-tune its voice models and certain open models on their own customer interaction data, setting goals such as fewer errors or greater satisfaction. ElevenAgents will gain memory of returning customers and follow-up routines, such as calling an applicant back. ElevenAI records meetings for employees with participants' consent and turns them into notes and task lists; free early access is open. ElevenLabs also launched a partner program and a future app marketplace, reported more than 550 partners in 44 countries, and expanded its agreement with Nvidia by 15 MW of computing capacity for large customers.

Smartwatches could move from recording the past to making personalized forecasts of which manageable sleep and exercise routine will provide the greatest benefit.

Google trains model to predict how sleep and movement affect health using Fitbit data

Smartwatches could move from recording the past to making personalized forecasts of which manageable sleep and exercise routine will provide the greatest benefit.

Sensor World Model was trained on 2.28 million de-identified person-weeks from approximately 100,000 consenting Fitbit users. From two weeks of baseline measurements, it simulates changes in 20 physiological markers over the next 12 weeks under different sleep, step, and cardio routines. An optimizer selects realistic scenarios, and a language model turns them into a personalized plan. The associated "coachable health age" metric uses resting heart rate, steps, sleep, and active minutes. In the All of Us cohort of 22,748 participants, it improved predictions of hypertension, cardiovascular disease, and type 2 diabetes compared with standard clinical measures. In a blinded comparison, coaches preferred plans grounded in simulated alternatives in 71.4% of evaluations, and physicians rated 99.2% of plans as safe in exercise intensity. These are research prototypes evaluated on retrospective data, not clinical diagnostic tools.

Evaluating dialogues and agents against custom criteria could become cheaper and more accessible for local deployment without a gigantic judge model.

Sber opens Pollux Judge, a bilingual 4-billion-parameter judge model

Evaluating dialogues and agents against custom criteria could become cheaper and more accessible for local deployment without a gigantic judge model.

Pollux evaluates answers in Russian and English using criteria, scales, and formats defined by the user. It handles not only a prompt-answer pair but also dialogues, RAG answers with retrieved context, and chains of tool calls. The model was trained with GRPO on bilingual data and can be deployed locally or through an OpenAI-compatible API. The accompanying benchmark contains 35 task groups and 2,115 expert-written prompts. According to Sber, on an independently expert-labelled set, Pollux's Spearman correlation with specialists was 0.744, higher than the tested general-purpose models with up to 1.1 trillion parameters. The code, demo, criteria, and labels are open; DeepEval and LightEval are supported.

Sources 1

Field laboratories receive an offline early-warning tool for unfamiliar pathogens, although its signals still require independent assessment.

Radical Numerics introduces offline Omnii Defense model for detecting dangerous pathogens

Field laboratories receive an offline early-warning tool for unfamiliar pathogens, although its signals still require independent assessment.

The model runs on a laptop with a USB sequencer without an internet connection and estimates the probability that a sequence is pathogenic or artificially generated, as well as its closest taxonomic group. A combined threat score from zero to one combines signs of unnaturalness with functional predictions instead of searching only for exact matches in a catalog of known organisms. In the company's experiment, after the Bundibugyo virus was removed from training, the model flagged 33% of its sequences as pathogenic; after the entire genus was removed, it flagged 22%. Access is limited to vetted external testers, including national laboratories and government agencies. Broad Institute expert Bronwyn MacInnis noted the approach's potential but emphasized the difference between detecting a signal and proving real danger or artificial creation of a pathogen.

Sources 1

Expanding SI infrastructure increasingly depends on verifiable electricity and water commitments, not only on the availability of land and chips.

Microsoft pledges not to pass data center electricity price increases on to neighbors

Expanding SI infrastructure increasingly depends on verifiable electricity and water commitments, not only on the availability of land and chips.

Microsoft president Brad Smith said the company will pay for its own electricity, prevent its facilities from raising tariffs for neighboring consumers, and return more water to local communities than it uses. The pledge comes amid growing resistance to data center construction after the US midterm elections. Using Quincy, Washington, as an example, CBS describes both tax revenue that helped fund a school and hospital and the displacement of some residents for facilities. Local rates have risen about 55% over 20 years, less than the state and national averages. Microsoft employs about 700 people there; Smith acknowledged the need to demonstrate benefits to communities and manage the consequences of growth.

Sources 1

Export-control plans have encountered SI data centers' dependence on Chinese network optics: restrictions could delay the connection of chips already purchased.

Microsoft asks US to drop ban on Chinese optical transceivers for data centers

Export-control plans have encountered SI data centers' dependence on Chinese network optics: restrictions could delay the connection of chips already purchased.

As relayed by iTnews, Microsoft and industry groups are warning the US administration that banning new Chinese optical transceivers could slow SI infrastructure construction. These components connect servers, routers, and switches over fiber optics; a single facility can require millions of them. Microsoft participated in a September meeting with representatives of the White House and National Security Council. The FCC is considering restrictions because of risks involving data theft, malware, and service disruption. China accounts for about two-thirds of the global market, while US companies Coherent and Lumentum lack the scale to replace supplies entirely, according to the Foundation for American Innovation. One option under discussion is a 3.2 Tbps threshold for future models whose mass production is expected later. Microsoft declined to comment.

Agents are beginning to perform salespeople's preparatory work rather than simply replacing the interface of a corporate customer database.

Rox raises $100 million at a $1.3 billion valuation for sales SI agents

Agents are beginning to perform salespeople's preparatory work rather than simply replacing the interface of a corporate customer database.

General Catalyst and March Capital co-led the Series B round; Sequoia and GV returned, and Prosperity7 joined for the first time. Rox builds agents that research customer accounts, select priority contacts and prepare outreach rather than merely storing information in a traditional CRM. Early customers include Ramp, New Relic, Trafigura, MongoDB and Groq. The company told Fortune that annual recurring revenue has grown more than tenfold since its first year after launch. The funds will support research and development, and the US sales and marketing team is planned to grow fivefold next year.

Sources 1

Agents gain faster access to large corporate datasets without a proportional increase in infrastructure costs.

Edelbyte emerges from stealth with $70 million for search and analytics for SI agents

Agents gain faster access to large corporate datasets without a proportional increase in infrastructure costs.

Edelbyte raised $70 million across seed and Series A rounds led by Sequoia and Quiet Capital, with participation from Greenoaks, Swish Ventures and Startpoint Capital. The company is developing a search and analytics platform for the growing number of queries that SI systems make against data that can remain in different stores. According to Globes, the platform already supports petabyte-scale deployments at Fortune 500 customers and allows some customers to run ten times as many queries within the same budget. The 40-person team in Israel, the United States and Europe previously spent more than a decade building data systems for Microsoft products together.

Schools will gain a specialized domestic SI assistant, but its real value will depend on knowledge testing and how well it helps students reason independently.

Russia's Ministry of Education and Sber to present an educational neural network by the end of the academic year

Schools will gain a specialized domestic SI assistant, but its real value will depend on knowledge testing and how well it helps students reason independently.

According to Sergey Kravtsov, the system is being developed by the Ministry of Education, Sber and other Russian companies together with teachers, scientists and engineers. It is intended to help teachers and schoolchildren, including students who are falling behind and children studying at home, and to rely on proven teaching methods and reliable information. Before a deployment decision, the neural network will be tested for answer accuracy and child safety. Kravtsov described it as a supplement to traditional education: the ministry's goal is not merely to give students a ready-made answer but to develop critical thinking and improve subject learning.

A fast open model for long agent tasks can now be called through familiar AWS APIs with data-residency requirements and a cheaper background tier.

Kimi K3.1 Pro Highspeed is now available on Amazon Bedrock

A fast open model for long agent tasks can now be called through familiar AWS APIs with data-residency requirements and a cheaper background tier.

Moonshot AI's model is available through global, US and India cross-Region Bedrock profiles. It retains the open 3-trillion-parameter-class architecture and a 1-million-token context; according to Moonshot, Highspeed generates responses 2–3 times faster than Kimi K3.1 Pro. Tool calling, structured output, images, streaming and configurable reasoning are supported. Existing Kimi K3 applications can switch by changing the model identifier. AWS promises zero retention of prompts and responses, no data sharing with Moonshot and no use of the data for training. The Flex tier costs half as much as Standard in exchange for potentially higher latency, while explicit caching of repeated context reduces the cost and duration of long agent conversations.

Teams can verify whether a new model, prompt or tool actually improved an agent instead of rebuilding metrics for every framework.

AWS open-sources UAEF with 36 metrics for comparing SI agents across frameworks

Teams can verify whether a new model, prompt or tool actually improved an agent instead of rebuilding metrics for every framework.

Universal Agent Evaluation Framework converts traces from LangGraph, LangChain, Strands Agents, Amazon Bedrock AgentCore and Langfuse into a common AgentTrace format and evaluates them identically regardless of the original SDK. Its 36 built-in metrics cover tool calling, response quality, safety, performance, multi-turn conversations, agent coordination and reasoning. It includes deterministic assessments and model-as-a-judge evaluation, which must cite an exact span of the response. UAEF stores experiments in DynamoDB and S3, compares versions against a baseline and classifies regressions by size from 5% to more than 20%. The Python library and deployable web interface use the same engine, support personal-data redaction and integrate RAGAS and DeepEval. The project is open source under Apache 2.0; customers across 12 organizations in telecom, healthcare and financial services have already used it before releasing new agent versions.

Long model training runs will have a way to stop a rare error on one GPU before it silently corrupts the entire model and wastes weeks of computation.

AWS proposes three checks for detecting silent GPU errors during model training

Long model training runs will have a way to stop a rare error on one GPU before it silently corrupts the entire model and wastes weeks of computation.

Silent data corruption occurs when a GPU returns a plausible but incorrect result without a hardware error; in distributed training, it enters the gradients and weights of every replica. AWS added three checks to a PyTorch DistributedDataParallel loop: detection of non-finite gradients, cross-replay of a microbatch on a GPU on another node with comparison of losses and gradient fingerprints, and comparison of weight fingerprints after the optimizer step. When a check fires, the system identifies the suspect node, does not apply the corrupted step, rolls back to a checkpoint, and hands the node to SageMaker HyperPod for replacement. The method was tested with software-injected faults on a cluster of 32 Nvidia H100 GPUs and repeated on the minimum three-instance configuration. The code is available on GitHub.

Reviewers will be able to spot more quickly where a generated agent policy actually changed permissions rather than merely expressed them in different words.

AWS open-sources semantic comparison tool for formal SI agent policies

Reviewers will be able to spot more quickly where a generated agent policy actually changed permissions rather than merely expressed them in different words.

The tool compares the structure of policies rather than their text: anti-unification finds their common part, while an e-graph accounts for specified algebraic laws. Reordering conditions or writing an external address as "not internal" no longer masks substantive differences such as replacing "and" with "or", dropping a path check, or changing a size threshold. The method compares several formalizations of the same requirement produced by a language model; the model itself plays no part in computing the differences. Examples cover SMT-LIB, Cedar, Dogwood, OpenShell, and CEL. The open-source Rust engine preserves both divergent variants for review by a person, validator, or solver. Its core uses state-restoration operations verified in Verus, but the complete solver does not yet have end-to-end formal proof.

SI agents and training jobs will be able to store state and vector indexes next to GPUs without a separate database in another cloud.

CoreWeave launches managed PostgreSQL alongside SI computing

SI agents and training jobs will be able to store state and vector indexes next to GPUs without a separate database in another cloud.

CoreWeave Database is now generally available: PostgreSQL-compatible databases run on CPU nodes inside a CoreWeave Kubernetes Service cluster alongside training and inference. The service uses the CloudNativePG operator, a primary instance and read replicas on different hosts, distributed file storage, and the pgvector extension. The company manages failover, patches, TLS, and continuous backups with point-in-time recovery. Customers retain control of access and queries through PostgreSQL roles and Kubernetes network policies. Migration is possible with standard Postgres tools; placing the database in the same cluster eliminates calls from agents, RAG systems, and experiments to another cloud provider.

A group will be able to obtain its own small app for shared planning from a simple description and modify it together as it works.

Google Labs introduces Tabletop, shared spaces with tools built from prompts

A group will be able to obtain its own small app for shared planning from a simple description and modify it together as it works.

In Tabletop, participants add links, files, and ideas to a shared space, and a request to @tabletop turns them into an interactive trip itinerary, poll, expense splitter, tournament bracket, or another tool. The group works with it simultaneously through an invitation link, as in a Google Doc, and can ask for rules or features to change without creating a new space. The experiment is gradually opening through a waitlist to users over 18 in the United States. Google offers ready-made templates and scenarios ranging from wedding planning to household chores with game points.

Organizations will be able to run more SI workloads on accelerators they already own instead of leaving GPUs assigned to different teams idle.

Basis introduces Basis Air for sharing GPUs among teams

Organizations will be able to run more SI workloads on accelerators they already own instead of leaving GPUs assigned to different teams idle.

The platform allocates a shared pool of accelerators according to departmental priorities and quotas, assigning spare capacity to tasks above their quotas. It manages clusters at different sites, supports whole-GPU allocation, sharing compute and video memory among tasks, and MIG profiles. Kubeflow, MLflow, and various storage systems are supported for training, inference, and Jupyter and VS Code environments. According to the developer, GPU utilization can rise from 30% to 80%, while preparing a working environment can fall from a month to one or two hours. Basis Air runs in the customer's own data center, including an isolated network, and does not require other Basis products. The company says about 80% of the platform's code is its own development.

Sources 1

Doctors will have an additional barrier against transmitting patients' personal data to generative medical services.

MDCalc launches CrossCover to remove unnecessary personal data from doctors' SI queries

Doctors will have an additional barrier against transmitting patients' personal data to generative medical services.

According to STAT's available description, CrossCover is intended to increase trust in medical chatbot answers while preserving patient and provider privacy. Unlike tools that expand the amount of clinical and insurance information in a query, MDCalc's approach removes identifying and protected health information before a request reaches a generative model. MDCalc's head of SI, Graham Walker, presents the service as a way to reduce the risks of clinical SI use without registering doctors by national identification number or requiring agreements that permit protected patient data to be entered.

Sources 1

The largest change in this forecast is not the disappearance of work altogether, but the need for hundreds of millions of people to master new tasks and ways of working.

Gartner forecasts net gain of 1.7 million jobs from SI by 2030

The largest change in this forecast is not the disappearance of work altogether, but the need for hundreds of millions of people to master new tasks and ways of working.

Gartner expects SI to create 10.6 million new jobs by the end of the decade and cause the elimination of 8.9 million positions. According to the company, employers postponed opening 1.3 million vacancies because of SI in 2026, while 2.2 million jobs were lost and 1.6 million created over the past year. By 2030, job losses due to SI will affect between 0.4% and 1.8% of workers worldwide, according to the forecast. Meanwhile, 435.5 million employees, just over 20%, will need to adapt to new ways of working. Gartner characterizes companies advanced in adoption as using SI to transform operations and rebalance human and technical capabilities, rather than merely replacing people.

Sources 1

iPhone users now have voice access to work data and extended Gemini conversations directly from their car display.

Gemini arrives as a standalone Apple CarPlay app

iPhone users now have voice access to work data and extended Gemini conversations directly from their car display.

iOS users can launch Gemini Live from their car display and hold a continuous voice conversation, continuing a chat from their phone or starting a new one. The assistant works with Calendar and Gmail, helps prepare for meetings, summarizes emails, finds nearby places, and allows complex topics to be discussed hands-free. The global rollout has begun for all supported Gemini Live languages. Connecting requires the latest version of the Gemini app, microphone permission, and at least one previous Gemini Live session on the iPhone; conversation history syncs with the phone.

The EU is discussing a move from general safety rules to a practical system of joint action when a frontier model or agent causes a serious incident.

Finland and Estonia propose EU-wide response mechanism for serious SI incidents

The EU is discussing a move from general safety rules to a practical system of joint action when a frontier model or agent causes a serious incident.

In a letter ahead of the first European Council meeting devoted to SI, the countries called for response arrangements at the EU level. They believe it is possible to mobilize capital and computing power to compete at the frontier while governing the technology responsibly. The discussion follows incidents involving autonomous agents, cyberattacks, and the discovery of code vulnerabilities. Hungary previously proposed a network of national contact points in case a provider loses control of a model. SI safety and security are on the leaders' meeting agenda.

Sources 1

Opposition to SI infrastructure growth over bills and resources is becoming bipartisan, but this remains a proposal rather than an active ban.

Senator Josh Hawley backs nationwide data center construction moratorium

Opposition to SI infrastructure growth over bills and resources is becoming bipartisan, but this remains a proposal rather than an active ban.

As relayed by Gizmodo, the Missouri Republican called on Senate colleagues to make a moratorium a priority for the next six weeks. According to Politico's description, the proposed pause would last until Congress guarantees that taxpayers will not face higher utility bills. Hawley became the first Senate Republican openly to support such a moratorium, previously championed mainly by progressive Democrats including Bernie Sanders. This is a political call, not an enacted restriction. It followed Republican losses in the midterm elections and successful local initiatives to restrict data center construction.

Before an agent enters a real company, it can be tested on synthetic copies of corporate systems to find costly errors in advance.

Plurai raises $13 million for corporate digital twins that test SI agents

Before an agent enters a real company, it can be tested on synthetic copies of corporate systems to find costly errors in advance.

Plurai builds a digital twin of a customer's knowledge bases and tools, simulates possible agent action sequences and creates edge cases for finding and fixing errors. The platform works alongside Claude Code, Codex, Copilot and Cursor. The company also uses data gathered in simulations to fine-tune small open-weight models for a specific business. Team8 led the seed round with participation from U&I, Velocity and Mercer. Founders Ilan Kadar and Elad Levi previously worked at Nexar; Kadar worked on simulations and validation of Waymo's autonomous driving systems. Kadar says joint tests with Nvidia and Nebius showed better quality than mid-tier frontier models at substantially lower cost, but TechCrunch presents this as a company claim.

Coding agents now have a shared source of current browser rules and can use modern built-in capabilities instead of outdated code more often.

Chrome, Edge, Firefox and MDN pool modern web guidance for coding agents

Coding agents now have a shared source of current browser rules and can use modern built-in capabilities instead of outdated code more often.

Modern Web Guidance installs a set of skills in an agent with recommendations from the Chrome, Microsoft Edge, Mozilla Firefox and MDN teams. The CLI finds the relevant rule, retrieves it into context and scans code for outdated patterns. The library describes 250 web-platform capabilities and 203 practical scenarios, from modern CSS and accessibility to WebMCP and local browser models. The project accounts for current Baseline compatibility and provides experimental Origin Trial and rapidly changing features only with the developer's explicit permission. Its goal is to reduce models' tendency to repeat old JavaScript and polyfills where the browser already offers a built-in solution.