Show only each day's banner story

Google shows agent teams solving open scientific problems with /teamwork

A multi-agent system with independent result verification has moved beyond code generation: it helps settle longstanding mathematical problems, check proofs and carry out multi-stage engineering work.

Google has published results from applying Gemini 4 and Antigravity's multi-agent /teamwork workflows to mathematics, chemistry and systems engineering. Specialized agents explore options in parallel while independent checkers use Lean 4, tests, simulators and hardware counters; coordinators delegate tasks to fresh agents with limited context. Working with scientists, the system proved the 2013 Courtade–Kumar conjecture in full generality and formalized the proof in Lean 4. Other work settled questions in adaptive data analysis, quantum queries, algorithms and data structures. /teamwork-LEAP autonomously formalized the Weil conjectures in approximately 1.8 million lines of Lean, while 30 agents using a physics-based spectral simulator achieved 95% top-1 accuracy in determining complex molecular structures. In engineering experiments, the system rewrote parts of Zircon and libgav1 in Rust while passing the original tests, optimized TPU kernels and reduced preemption for some TPU workloads by 24% through changes to the Borg scheduler. /teamwork, /boost and Deep Review are available to external users; Google plans to add specialized workflows in upcoming updates.

A personal SI agent can now not only choose a product but also pay for it with the owner's permission, turning everyday requests into completed actions.

OpenAI launches ChatGPT Wallet for purchases through personal agents

A personal SI agent can now not only choose a product but also pay for it with the owner's permission, turning everyday requests into completed actions.

ChatGPT Wallet lets dots agents buy products on a user's behalf. The service was built with American Express, Mastercard and Visa: the card is entered once, and the owner confirms every purchase with a separate security check. The agent receives a payment token scoped to the specific transaction rather than the card details; card rewards and protections remain intact. At launch, the feature is available to adult ChatGPT Pro subscribers in the US, and OpenAI is not taking a cut of purchases for now. Alongside the launch, the company published tools and guidance to help merchants make their websites work with these agents.

A model's published weights can reveal the generation of hardware used to create it, providing a new technical way to inspect training infrastructure without access to a data center.

Researcher finds numerical traces of Nvidia Blackwell in six Chinese models

A model's published weights can reveal the generation of hardware used to create it, providing a new technical way to inspect training infrastructure without access to a data center.

Former OpenAI researcher Will DePue reported that the published weights of DeepSeek-V4-Flash and V4-Pro, Kimi K3 and K3.1-Pro, MiniMax-M3 and Ling-3.1-flash retained characteristic initialization patterns from Blackwell GPUs. The method uses differences in how random starting values are distributed across compute units: the B200/B300 have 148 SM units, while the GB200/GB300 have 152. The B300 445-of-469 check reproduced selected DeepSeek-V4-Flash weight rows bit for bit. The GB200 was used for Kimi K3. The method also correctly identified the hardware of all 24 control models with known equipment. The authors estimate that the five largest runs may have required about 12 thousand Blackwell GPUs, but this is a calculation rather than a confirmed figure. The analysis shows which hardware initialized the weights, but does not establish the GPUs' location, owner, means of access or any violation of export restrictions.

Anthropic's offering combines exceptional demand with geographic restrictions on investor access: technology controls and export rules are shaping not only the laboratory's products but also its capital structure.

Anthropic IPO: demand six times supply as Chinese investors are excluded

Anthropic's offering combines exceptional demand with geographic restrictions on investor access: technology controls and export rules are shaping not only the laboratory's products but also its capital structure.

According to Bloomberg, orders for Anthropic shares were approximately six times the number available. The company intends to raise up to $64.4 billion on Nasdaq and another $26 billion through a private placement at a target valuation of about $1.4 trillion. The public offering covers 825 million shares priced at $70–78, with trading under the ANT ticker scheduled to begin on November 19. According to The Decoder, Anthropic's annualized revenue grew from $71 billion in September to $81 billion by the end of October, while Nvidia is participating in the private placement with $10 billion. Bloomberg later reported that several banks are rejecting orders from institutional investors in mainland China and Hong Kong or consider allocations to them unlikely, citing US ITAR rules. Goldman Sachs and Morgan Stanley are leading the offering alongside 20 other banks; Anthropic declined to comment on the restrictions.

Access to advanced computing is now subject to direct parliamentary scrutiny: the hearing answers will influence future export rules for SI chips.

US Senate summons Jensen Huang and commerce secretary for chip export hearings

Access to advanced computing is now subject to direct parliamentary scrutiny: the hearing answers will influence future export rules for SI chips.

Senate Banking Committee Chairman Tim Scott asked former Nvidia CEO Jensen Huang and Commerce Secretary Howard Lutnick to testify at a public hearing on December 3, promising subpoenas if they refuse. Three House committees separately requested that Huang and Nvidia General Counsel Tim Teter appear at hearings on December 1–2. Lawmakers are examining compliance with export restrictions and the Commerce Department's decisions concerning Chinese companies. Lutnick agreed to testify, while Huang agreed to a closed-door interview, according to House Foreign Affairs Committee Chairman Brian Mast. Nvidia says it complies with US export laws; China denies the espionage allegations and warns that new restrictions would harm supply chains.

The long-term lease of an entire campus gives Anthropic a substantial power reserve and shows how former mining sites are becoming infrastructure for frontier models.

Riot Platforms leases the 910 MW Corsicana campus to Anthropic

The long-term lease of an entire campus gives Anthropic a substantial power reserve and shows how former mining sites are becoming infrastructure for frontier models.

According to AI Weekly's description, Riot Platforms has signed a 15-year lease with Anthropic for its entire Corsicana campus in Texas. The agreement covers 910 MW of prepared power capacity and is valued at approximately $28 billion over the initial term. Two five-year extensions could raise the total value to $54 billion. Anthropic is the tenant for the site's frontier SI computing operations.

Long-term contracts with a major customer allow an SI cloud provider to borrow more cheaply and continue expanding computing capacity.

CoreWeave is raising a $4 billion loan backed by Meta contracts

Long-term contracts with a major customer allow an SI cloud provider to borrow more cheaply and continue expanding computing capacity.

CoreWeave has approached the credit market seeking $4 billion. Morgan Stanley and Mitsubishi UFJ will syndicate the loan to lenders, with Meta contracts serving as collateral, according to Semafor. The company is targeting an all-in interest rate of about 7.5%, while the yield on its unsecured bonds approached 16%. A source told the publication that the offer received a positive response during early marketing.

Sources 1

The transition to a wireless neural interface has been delayed: development of the technology for people with paralysis and ALS continues, but the next version requires further work.

Synchron pauses trials of its second-generation neural interface

The transition to a wireless neural interface has been delayed: development of the technology for people with paralysis and ALS continues, but the next version requires further work.

Synchron has paused clinical trials of the wireless version of its brain-computer interface. The first-generation device is delivered by catheter through veins and places a stent near the motor cortex to transmit brain activity to digital devices. Company representatives told STAT that the pause is not safety-related. CEO Tom Oxley previously spoke about a delay involving one component that slowed the submission of documents to the FDA. The company is also cutting jobs in New York City and working on a new device, according to WIRED.

Sources 1

Local SI gains a predictable foundation: applications will find it easier to update without unexpected breakage, and models will run on personal devices.

llama.cpp reaches stable version 1.0.0

Local SI gains a predictable foundation: applications will find it easier to update without unexpected breakage, and models will run on personal devices.

The ggml-org team released llama.cpp and ggml libraries version 1.0.0. Incompatible changes to the public C API and ABI will now be reflected in the major version, backward-compatible features and model support in the minor version, and fixes in the patch version. The local multimodal model runtime supports more than 150 architecture variants and the main speculative decoding methods. The team is also expanding prebuilt packages for Linux, macOS, Windows, Android and iOS, package managers, and change-testing infrastructure on hardware from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel.

Ordinary images can now be turned into editable three-dimensional worlds and transferred to working tools, accelerating the creation of spaces for video, games and robots.

World Labs opens the Atlas world model and 3D-scene tools

Ordinary images can now be turned into editable three-dimensional worlds and transferred to working tools, accelerating the creation of spaces for video, games and robots.

Atlas combines the generation and reconstruction of three-dimensional worlds with applications and APIs. Viewfinder builds a scene from photographs or generated images and lets users plot a camera path through it, while Chisel defines the spatial structure and updates every view after objects are moved. The result can be converted from a splat cloud into a mesh with geometry and transferred to games, simulations, Blender and other 3D tools. World Labs proposes uses in film and visual effects, architecture, design, and the creation of environments for robot training.

Sources 1

Autonomous transport gains a path from trials to routine travel nationwide, with defined liability and mandatory technical safeguards.

Russian government introduces bill for permanent autonomous-vehicle rules

Autonomous transport gains a path from trials to routine travel nationwide, with defined liability and mandatory technical safeguards.

The document covers vehicles at automation levels three, four and five and sets admission and insurance conditions plus obligations for manufacturers, owners, drivers and remote-support operators. Liability for a collision is assigned to whoever controlled the vehicle when the harm occurred; if there was no driver or the driver cannot be identified, responsibility passes to the owner. In level-four and level-five vehicles, a driver may be absent from the cabin while a remote operator monitors the journey. Autonomous vehicles must have a black box, GLONASS and protection against outside interference; after the experimental regimes in place since 2018 end, they would be admitted to freight and passenger transport, taxis and scheduled routes.

OpenAI is betting on general-purpose models as a shared robot-control mechanism rather than on a standalone robotics division with its own specialized research.

OpenAI disbands its standalone robotics team

OpenAI is betting on general-purpose models as a shared robot-control mechanism rather than on a standalone robotics division with its own specialized research.

Robotics researchers have moved to the main research team under Mark Chen, while applied robotics and hardware groups have joined Greg Brockman's Applied organization. OpenAI explained the reorganization by saying that general-purpose models, including GPT-6 Astra, are already being applied directly to robot control, so separate specialized models are no longer considered a necessary path. The company confirmed to Fortune that it continues developing its own humanoid, with a demonstration planned by Sam Altman for 2027. In an Nvidia study, an Astra-powered agent achieved a 74.8% success rate in a simulated manipulation test, compared with 51% for the best specialized model on the leaderboard. Specialists nevertheless note robotics companies' advantage in real-world operational data and the need for small onboard models where cloud connectivity is unavailable.

Sources 1

The personal assistant gains the user's financial context and turns statements into analysis and budget plans, while money transactions remain outside its authority.

Gemini learns to analyze spending through Google Wallet and bank accounts

The personal assistant gains the user's financial context and turns statements into analysis and budget plans, while money transactions remain outside its authority.

Opt-in integration with Google Wallet is rolling out to all Gemini app users in the US on Android, iOS and the web. Bank accounts and credit cards can also be linked to Wallet through Plaid. When prompted with @Google Wallet, the assistant creates charts of cash flow, asset allocation and budgets, finds recurring payments and subscriptions whose prices have risen, and proposes budgets based on past spending. Gemini has read-only access and cannot transfer money, manage accounts or invest; the connection can be disabled. Personal Intelligence has also gained a Memory Summary section for viewing, editing and deleting information used for personalization.

Growing SI computing is increasing the value not only of chips but also of heat-removal systems, making cooling a strategic resource in its own right.

Samsung discusses acquiring data-center cooling supplier Modine

Growing SI computing is increasing the value not only of chips but also of heat-removal systems, making cooling a strategic resource in its own right.

Samsung Electronics is in talks to acquire US-based Modine Manufacturing, Bloomberg reports. A deal could be agreed within weeks, but no final decision has been made, and the companies declined to comment. Modine shares rose 28% in extended trading following the report. The company's data-center unit grew revenue by an average of 86% a year over the past three fiscal years. For Samsung, the acquisition would expand its presence in SI infrastructure beyond memory; a year earlier, the company bought cooling-system manufacturer FläktGroup for €1.5 billion.

SI can accelerate engineering calculations while leaving the specialist in control of the shape, making strong structures more practical to manufacture.

MIT teaches an SI copilot to improve structures together with an engineer

SI can accelerate engineering calculations while leaving the specialist in control of the shape, making strong structures more practical to manufacture.

MIT researchers added a human to the topology optimization process: rather than merely producing a finished design, the algorithm shows the engineer regions during the calculation that can be made thicker or thinner. A machine-learning model selects those regions using a synthetic dataset of likely feedback on thousands of structures. In the paper published in Communications Engineering, the collaborative approach increased linear buckling load by 39% and simplified manufacturing while adding only 15 seconds to total design time compared with the fully algorithmic version. For now, user actions are limited to changing thickness in the proposed regions.

Sources 1

Biological models can be fine-tuned faster and more reproducibly on existing GPUs, allowing scientists to test more hypotheses on the same equipment.

Open-source TurboPairFormer speeds up OpenFold3 fine-tuning by up to 1.73 times

Biological models can be fine-tuned faster and more reproducibly on existing GPUs, allowing scientists to test more hypotheses on the same equipment.

Lambda, Stevens Institute of Technology and the OpenFold team released TurboPairFormer, open-source GPU kernels for triangle attention and multiplication operations in biomolecular structure prediction models. On 16 Nvidia H100 GPUs with a crop size of 768 tokens, OpenFold3 training became up to 1.73 times faster than with the original Triton backend. BF16 rounding compensation reduced mean gradient error by 28–47% in controlled experiments, while a fixed accumulation order produced bit-identical gradients in all 600 repeated-run tests. The package is available through PyPI and supports H100 and H200; the speedup enables more frequent fine-tuning of models for proteins, nucleic acids and drug discovery on the same GPUs.

Sources 1

SI helps find chemical patterns in experimental databases that are difficult to spot manually, opening a path to new materials and medicines.

St Petersburg State University and Samara State Technical University use SI to find unusual bonds in crystals

SI helps find chemical patterns in experimental databases that are difficult to spot manually, opening a path to new materials and medicines.

The chemists combined topological analysis of a crystal-structure database with a large language model to search for non-covalent interactions without predefined distances and angles. ToposPro determined atomic neighbors from geometry, after which the model helped identify statistical patterns while researchers verified its conclusions. The paper published in Coordination Chemistry Reviews examined 1553 contacts in 473 crystals of hypervalent halogen salts and identified five arrangements of particles. These include perpendicular contacts that conventional selection could miss; a single iodonium center proved capable of holding up to three partners. The authors plan to clarify the effect's electronic nature through quantum-chemical calculations.

Sources 1

The EU is trying to agree on a common course: accelerate its own SI industry while preparing for incidents involving autonomous systems.

EU leaders gather in Dublin for the first summit dedicated to SI

The EU is trying to agree on a common course: accelerate its own SI industry while preparing for incidents involving autonomous systems.

The meeting at Dublin's Convention Center addresses Europe's role in developing SI, competition with the US and China, and protecting people and businesses from risks. Mistral CEO Arthur Mensch and researcher Geoffrey Hinton are due to address the participants. Estonia and Finland called for an EU-wide mechanism for responding to serious SI incidents, while Mistral opposes slowing development and urges the EU to continue advancing the technology. European Council President António Costa justified a dedicated leaders' meeting by the speed of SI development and its influence on the economy and society.

Sources 1

Common notification rules will help detect dangerous behavior by autonomous agents sooner and protect critical infrastructure with shared data.

White House task force discusses reporting cyberattacks by SI agents

Common notification rules will help detect dangerous behavior by autonomous agents sooner and protect critical infrastructure with shared data.

Participants discussed national standards under which developers would report SI security incidents and share threat information, an administration spokesperson told Reuters. A separate topic was a mechanism allowing banks, energy companies and emergency services to disclose vulnerabilities and attacks without fear of legal consequences. Task force vice chair Emil Michael reported progress on such standards. The discussion followed incidents involving autonomous agent attacks and comes as the Cybersecurity Information Sharing Act of 2015, which protects companies when they share threat data, approaches expiration.

Armed ground robots are moving from demonstrations to real security operations, making rules for their use and human control a practical question.

El Salvador's president shows armed robots during an operation in Haiti

Armed ground robots are moving from demonstrations to real security operations, making rules for their use and human control a practical question.

Nayib Bukele published video of armed robots during an operation against gangs at the main stadium in Port-au-Prince. The stadium had been under gang control for the previous two years, according to Al Jazeera. The published report gives no information about the operation's outcome.

Cohere is expanding its sovereign platform beyond models into secure development and autonomous cyber defense while strengthening the European side of its business.

Cohere completes AISLE and cubic acquisitions for cybersecurity and code review

Cohere is expanding its sovereign platform beyond models into secure development and autonomous cyber defense while strengthening the European side of its business.

Cohere has completed its acquisition of Czech company AISLE, which develops autonomous vulnerability discovery and remediation systems, and welcomed the team behind British code-review platform cubic. According to BetaKit, all cubic employees are joining Cohere, while AISLE's 50-person team is joining a workforce of more than 800. AISLE will form the foundation of Cohere Cyber, and Prague will become the company's global center for SI cybersecurity. The cubic team will establish a secure coding vertical: its agents review changes, find bugs and vulnerabilities, and can operate in on-premises and air-gapped environments. These are Cohere's third and fourth strategic acquisitions of 2026, following Reliant AI and the agreement to merge with Aleph Alpha; the companies did not disclose the terms of the new deals.

The Pentagon's central digital technology and SI unit has a new name while retaining its previous mission and leadership.

Pentagon renames CDAO the Super Intelligence Office

The Pentagon's central digital technology and SI unit has a new name while retaining its previous mission and leadership.

The US Department of Defense has renamed the Chief Digital and Artificial Intelligence Office as the Super Intelligence Office. The department said the unit's mission is unchanged: the rename follows President Donald Trump's September directive for federal agencies to replace references to SI with the term super intelligence. Office head Cameron Stanley linked the new name to the transition from experimental prototypes to using sovereign systems for military superiority across every domain.

Claude is moving from advising manufacturing specialists to acting inside factory systems, while an equipment-connection standard is intended to simplify the next step toward physical processes.

Anthropic connects Claude to industrial software with Claude for Manufacturing

Claude is moving from advising manufacturing specialists to acting inside factory systems, while an equipment-connection standard is intended to simplify the next step toward physical processes.

The new plugin for Team and Enterprise customers connects to Autodesk Fusion, Oracle NetSuite, Microsoft 365 and Infor; other systems can be integrated through file exports to Microsoft 365. Seventeen skills let Claude draft quotes from actual inventory, compare suppliers, generate quality reports and design variants, and analyze sensor readings and manuals during equipment downtime. Through the Infor plugin, Claude can order from a backup supplier and reschedule production after human approval; results include links to source documents for employee review. The plugin does not yet control physical machines. For such connections, Anthropic is separately developing the Model Hardware Standard: its research preview uses standardized drivers for robotic arms and microscopes and, by the company's estimate, cuts setup from weeks or months to hours or minutes while preserving human oversight.

The model helped cover an entire archive that researchers had been reading at roughly one page a week, while a formal verifier distinguished confirmed theorems from false ones and those needing refinement.

Claude helps recover hundreds of theorems from Kurt Gödel's workbooks

The model helped cover an entire archive that researchers had been reading at roughly one page a week, while a formal verifier distinguished confirmed theorems from false ones and those needing refinement.

A team led by logician Jan von Plato at the University of Helsinki applied Claude to transcriptions of 16 Gödel notebooks from the early 1940s: 1,009 pages of dense writing, including the obsolete German Gabelsberger shorthand. Within days, the model produced a 650-page analysis that researchers then checked page by page. At the team's request, Claude extracted 663 theorems with proofs, proof sketches or proof ideas: according to The Guardian, about 400 were verified in Coq as stated, roughly 200 required clarification and about 50 were false. The reconstructed proof of the independence of the axiom of choice also passed Coq verification. External expert Joel David Hamkins noted that historians still need to separate Gödel's contribution from the model's reconstruction and that a similar chain of reasoning alone does not prove Gödel had a completed result.

Most new spending is going to computing infrastructure rather than applications: it is forecast to exceed combined spending on services, software and agents.

Gartner forecasts global SI spending will reach $4 trillion in 2027

Most new spending is going to computing infrastructure rather than applications: it is forecast to exceed combined spending on services, software and agents.

Gartner's new forecast puts global SI spending in 2027 at $4.05 trillion, 47.5% above the projected $2.746 trillion for 2026. Infrastructure will account for 59% of spending: $2.397 trillion, up from $1.563 trillion this year. Spending on SI-optimized servers, the largest infrastructure segment, is expected to grow from $527.6 billion to $957.2 billion. The SI-agent segment is forecast to expand from $30 billion to $67 billion, SI cybersecurity from $51 billion to $86 billion, and SI services from $576 billion to $746 billion. Gartner analyst John-David Lovelock linked the growth to users scaling workloads and to the increasing complexity of model training.

Sources 1

SI is creating work beyond software development: more specialists are needed for data and deployment, while basic model skills are becoming widespread.

Sberbank nearly triples hiring for SI roles

SI is creating work beyond software development: more specialists are needed for data and deployment, while basic model skills are becoming widespread.

In the first nine months of 2026, Sberbank hired almost three times as many specialists to work with SI as in all of 2025: 279 people for ML roles, 95 for other SI roles and 616 for data annotation. Basic training in working with SI has been completed by 47 thousand frontline employees, about 40% of that workforce. Applications for Middle ML Engineer vacancies rose almost fourfold, and more than fivefold for Senior positions. According to Sberbank's analysis, generative SI skills were required in 4.7 thousand vacancies in September, almost 20% more than a year earlier, with seven of the ten professions requiring such skills being non-technical.

Sources 1

Schoolchildren gain a personal assistant that guides them through solutions instead of doing the work for them, making individual support more accessible.

Alice AI begins helping schoolchildren with math in the My School public-services app

Schoolchildren gain a personal assistant that guides them through solutions instead of doing the work for them, making individual support more accessible.

The module for students in grades 5–11 is available in 20 pilot regions where the app serves as the main electronic gradebook. It explains topics, analyzes the steps used to solve a particular problem, selects similar exercises and helps prepare for tests. Instead of providing a finished answer, Alice AI asks guiding questions and offers hints so that the student completes the solution independently. Yandex Education, teachers and methodologists are developing the tool; only the parent and child have access to the data.

A unified assistant could make complex financial contracts easier to understand and give customers comparable terms regardless of their chosen bank.

Bank of Russia expects a unified SI assistant for choosing financial products in 2027

A unified assistant could make complex financial contracts easier to understand and give customers comparable terms regardless of their chosen bank.

The assistant based on the central bank's methodology will be embedded in the apps and websites of banks and insurance organizations, according to Bank of Russia Deputy Governor Mikhail Mamuta. In plain language, it will explain the main terms of deposits, mortgages, credit cards and comprehensive vehicle insurance and highlight departures from standards. Recommendations will be voluntary for the first six months; if testing succeeds, the regulator may enshrine the requirement in its rules. Testing since late last year involved the ten largest banks, including VTB, T-Bank and Sovcombank. The Bank of Russia expects recommendations that are equally distant from all market participants and provides a procedure for discussing disputed answers with financial organizations.

Sources 1

Autonomous delivery is expanding to products that are often needed urgently: medicines can arrive at the entrance without waiting for a courier.

Yandex robots begin delivering medicines across Moscow

Autonomous delivery is expanding to products that are often needed urgently: medicines can arrive at the entrance without waiting for a courier.

The 36.6 pharmacy chain and Yandex Delivery launched a trial of robotic delivery for non-prescription medicines and other pharmacy goods in eight Moscow districts. Orders are placed through the 36.6 and Gorzdrav websites and apps, a pharmacist hands the package to the robot, and the customer tracks it through Yandex Go or the pharmacy website. The robot brings the order to the building entrance; a combined format is also available in which a courier takes it to the apartment. Following the trial, the companies plan to launch the service in St Petersburg. According to Yandex's own figures, more than 1500 of its robots were operating in 13 cities by the end of the third quarter.

Sources 1

Real car-sharing journeys will become training data for local robotaxis and autonomous buses.

Hyundai, Socar and Apex Mobility will collect road data for autonomous driving

Real car-sharing journeys will become training data for local robotaxis and autonomous buses.

Hyundai signed a memorandum with car-sharing company Socar and Apex Mobility, its joint venture with Krafton. Hyundai will supply vehicles and a DevKit for integrating autonomous systems, while the partners collect and analyze data on South Korean roads and traffic. The goal is to adapt autonomous-driving platforms to local conditions and train SI on real journeys. Apex Mobility, established in May, eventually plans commercial robotaxis and autonomous buses.

Sources 1

Military-autonomy development is being localized in India: local teams will gain tools for their own uncrewed aviation systems.

Shield AI opens Bengaluru center to develop autonomy for India's defense systems

Military-autonomy development is being localized in India: local teams will gain tools for their own uncrewed aviation systems.

Indian engineers at the new office will develop and test autonomy software for Indian aircraft and defense systems on the Hivemind platform. Shield AI is recruiting specialists in autonomy, flight software, simulations, systems integration and testing. This is the company's second office in the country after the New Delhi office opened in April, which works with India's Ministry of Defence, armed forces and industry.

Sources 1

SI has accelerated patch writing, but human review is not scaling at the same rate, so open-source code quality needs new methods of selection and review.

Linux maintainer scales back reviews because of the flood of SI-generated code

SI has accelerated patch writing, but human review is not scaling at the same rate, so open-source code quality needs new methods of selection and review.

Lorenzo Stoakes, a Linux contributor since 2014 and a memory-management specialist, stopped additional reviewing in the memory management, core and Rust areas, retaining only the parts of the kernel he directly maintains. He explained that he lacked the resources to review the growing number of SI-assisted patch series properly and redistributed his previous workload among four contributors. By late 2026, up to 25% of pull requests to the kernel were created by neural networks, according to CNews. Linus Torvalds had previously criticized the mass duplication of generated vulnerability reports and the disruption such reports caused to the development process.

Sources 1

Schools are trying to preserve the right to choose SI tools separately from basic access to knowledge; Google's answer will determine whether Search and generative features can be separated.

New York City Council demands that Google preserve search access for schools that ban SI

Schools are trying to preserve the right to choose SI tools separately from basic access to knowledge; Google's answer will determine whether Search and generative features can be separated.

Speaker Julie Menin and Council Member Shekar Krishnan sent Google CEO Sundar Pichai a letter demanding an explanation by November 19 of whether Google Search will remain available in schools that prohibit students from using generative tools. The letter follows a rule announced on November 3 that ties search access to permission for products such as Gemini. In September, New York City introduced a one-year moratorium on student use of SI from 2-K through eighth grade. The officials ask about educator choice, answer accuracy, protection against misinformation, and the collection and use of student data. A Google representative previously told The New York Times that the company cannot alter its infrastructure to separate its products from SI, including Search.

On-premises clinical SI can turn years of unstructured medical records into verifiable research data while keeping those records inside the hospital.

Synapsis AI analyzes records for 7,800 Cleveland Clinic patients on hospital servers

On-premises clinical SI can turn years of unstructured medical records into verifiable research data while keeping those records inside the hospital.

Dyania Health's system processed records for more than 7,800 patients who underwent aortic valve replacement between 2005 and 2021, including physicians' notes and test results spanning up to 16 years. Synapsis AI answered 88 clinical questions about each patient and linked every answer to its source record. The work ran on NVIDIA servers inside Cleveland Clinic's network without transferring medical records outside it. The resulting data was used to build a model predicting the risk of replacement-valve deterioration and reintervention three, five and ten years after surgery. Doctors review the evidence and make every decision; the study demonstrates dataset construction and prediction rather than improved treatment outcomes, and Dyania says external validation is needed.

SI agents now have an open lightweight browser that can be inspected, modified and run in their own infrastructure instead of a closed managed service.

Cloudflare open-sources its lightweight Kitesurf browser for SI agents

SI agents now have an open lightweight browser that can be inspected, modified and run in their own infrastructure instead of a closed managed service.

Kitesurf has been published on GitHub under the Apache 2.0 license: developers can run it locally, modify it and deploy it in their own Cloudflare Workers accounts. Without a browser container, the engine parses HTML, applies CSS, executes JavaScript and renders page screenshots, while Chrome DevTools Protocol support allows Puppeteer and agent harnesses to connect. It uses the Blitz, Stylo, Boa, Linebender and Krilla Rust packages compiled to WebAssembly. Cloudflare intends the browser to use less CPU time and memory than full Chromium and to scale across the Workers network. Self-deployment with Dynamic Workers requires a paid Workers plan; the project continues to improve compatibility and reliability.

Models are substantially better at recognizing useful advice when choosing between options than at independently inferring a user's hidden goal and redirecting their answer.

XYBench tests whether models notice a question's mistaken premise

Models are substantially better at recognizing useful advice when choosing between options than at independently inferring a user's hidden goal and redirecting their answer.

Ai2, working with Carnegie Mellon, NVIDIA and Johns Hopkins, has released a set of 8,115 programming and everyday-task questions in which the literal request Y does not serve the actual goal X. Separate metrics assess whether a solution addresses the goal, how the answer is weighted and whether the model identifies the mistaken assumption. Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B and DeepSeek-V4 Pro addressed the original request more often than the underlying problem. On Stack Overflow questions, the models focused on the actual goal in only 35–41% of answers, compared with 55% for top human answers, even though they recognized the correct goal-focused answer in 71–84% of cases when choosing between two prepared options. Explicitly stating the goal improved some metrics but did not close the gap. XYBench won the Best Resource Paper Award at EMNLP 2026.

Generative editing makes advertising more persuasive, but without clear labeling buyers lose an accurate picture of the actual property.

BBC shows how SI-enhanced property photos mislead buyers

Generative editing makes advertising more persuasive, but without clear labeling buyers lose an accurate picture of the actual property.

Journalists described the practice known as housefishing: estate agents use SI to remove structural elements, alter windows and finishes, or add imaginary details, leaving the actual property noticeably different from the listing. In a survey of 906 British homebuyers, one in five who had bought a home in the past year reported viewing a property whose photos had been digitally enhanced relative to reality. Rightmove, Zoopla and Winkworth removed the images shown by the BBC, while Howards introduced labels on all edited photos. A Trading Standards representative told the BBC that basic editing and virtual furnishing are permissible when SI use is disclosed, while changing a room's structure or hiding its features may mislead consumers and break the law.

Sources 1