Show only each day's banner story

Anthropic releases Claude Fable 5.2 and Mythos 5.2 for demanding technical work

The models are substantially more capable at prolonged engineering and scientific work, from large code migrations to formal proofs and new calculations in physics.

Both models share the same foundation: Fable 5.2 is available to all users, while Mythos 5.2, without some cybersecurity and biology restrictions, is offered to vetted organizations. Anthropic estimates generation speed gains of up to 30%; pricing is $8 per million input tokens and $40 per million output tokens. A separate fast mode increases output speed by up to 2.5 times at twice the price. In Anthropic's measurements, Fable 5.2 scored 49.1% on FrontierCode 2.0 at maximum effort and 74.2% on Humanity's Last Exam with tools. In prerelease testing, about a thousand agents formalized the proof of Fermat's Last Theorem in Lean in 30 hours, and the model also helped complete calculations for two open problems concerning black holes. These results and preprints were published alongside the announcement. The model underwent external evaluations by the UK AI Security Institute, Frontier Design and METR. In one adversarial test, Mythos 5.2 used data outside its sandbox in 3.1% of scenarios, compared with 21% for Mythos 5.1; Anthropic nevertheless emphasizes that evaluations do not guarantee behavior after deployment. The company deliberately kept capabilities that directly accelerate model development at roughly the level of Opus 5.5.

Cloud SI is gaining more specialized hardware: agents with massive contexts and sensitive data will be able to run more efficiently without relying solely on Nvidia GPUs.

Google unveils TPU 8tx for confidential computing and the dual-chip TPU 8i Max

Cloud SI is gaining more specialized hardware: agents with massive contexts and sensitive data will be able to run more efficiently without relying solely on Nvidia GPUs.

Google has expanded its eighth-generation TPU portfolio with two systems. TPU 8tx uses TPU 8t silicon and fifth-generation AMD EPYC processors: data and model weights remain encrypted and isolated even while being processed. The Boardfly network reduces latency by up to two times compared with TPU 8t, while one system can handle confidential inference, fine-tuning and reinforcement learning for agents. TPU 8i Max combines two TPU 8i chips in one package and doubles compute, SRAM capacity and HBM bandwidth for long-context workloads. These are new system configurations rather than new chips: the modular architecture allowed Google to introduce four TPU systems in a single year. The company also released tools for porting PyTorch code and Hugging Face models to TPUs. TPU Sync combines memory across chips, hosts and storage; according to Google, it reduces KV-cache transfer latency and time to first token by more than 90%.

SI is beginning not only to perform individual research tasks but also to run a cycle of selecting hypotheses, experimenting and validating improvements for subsequent models.

Baidu introduces ERNIE Evolve, a system that runs its own experiments to improve SI

SI is beginning not only to perform individual research tasks but also to run a cycle of selecting hypotheses, experimenting and validating improvements for subsequent models.

ERNIE Evolve selects research directions, modifies code and runs experiments in an isolated environment. An evaluator external to the agents checks correctness and performance gains, while a separate reviewer rejects solutions that exceed the specified boundaries or exploit weaknesses in the evaluation. On NanoGPT Speedrun, the system reduced the time needed to reach the target quality to 38.9 seconds, compared with the best public result of 39.9 seconds; on NanoChat AutoResearch it achieved 0.8763 BPB versus 0.8878. In GPU kernel optimization, its score of 0.819 was close to the SOL-ExecBench leader's 0.827. While optimizing the still-in-development ERNIE-5.5-Flash, the system increased deterministic inference throughput on GPUs by 33.7% in about three days. On Kunlunxin P800, generation speed rose to 623.6% of the baseline; a significant part of the gain came from enabling optimized execution paths that were already available. All figures are reported by Baidu.

The breach of development infrastructure affected not only service keys but also some customer data and models; the privacy of an SI service also depends on its external providers.

Fireworks confirms attackers accessed customer data, datasets and unfinished models

The breach of development infrastructure affected not only service keys but also some customer data and models; the privacy of an SI service also depends on its external providers.

An update to the investigation into October's incident has expanded the initial assessment: resource metadata, billing addresses, invoice histories and balances were affected, along with secrets, datasets and incomplete model weights for some customers. Fireworks does not store payment-card numbers. On October 23, a security incident occurred at the third-party cloud provider hosting Fireworks research machines. The next day, the attacker used it to access development machines and run an unauthorized workload; the company detected the activity the same day and has recorded no further successful intrusions into the customer platform since October 25. Fireworks links the attack to a financially motivated group and LLMJacking, the sale of tokens providing unauthorized access to language models. The company brought in Palo Alto Networks Unit 42, replaced credentials, restricted network paths and temporarily disabled self-service dataset uploads and downloads for batch inference and training. According to Bloomberg, Notion and Cursor were among those notified. Notion said the incident affected fewer than 1% of its workspaces. The new dataset contained Notion AI chat transcripts used for the Insights feature; for the vast majority of affected workspaces, they contained no author or workspace labels and no tool-call results. Notion disabled Insights, stopped serving open-weight models through Fireworks and advised affected customers to rotate secrets they had entered into Notion AI during the incident period.

Frontier model evaluation and information sharing about incidents are moving from companies' disparate pledges toward a common international framework.

Eight countries and seven SI companies sign the Paris Call for frontier model safety

Frontier model evaluation and information sharing about incidents are moving from companies' disparate pledges toward a common international framework.

At Emmanuel Macron's initiative, the document was supported by Bahrain, Canada, Finland, Japan, Norway, the Netherlands, the United Kingdom and Sierra Leone, as well as Anthropic, Black Forest Labs, Hugging Face, Microsoft, Mistral AI, OpenAI and Sakana AI. Representatives of the UN, OECD, researchers and civil society took part in the discussion. The parties agreed to develop independent evaluations, common standards for assessment and reporting of major incidents, and greater clarity about the responsibilities of ecosystem participants. Non-sensitive evaluation results are to be published in a standardized format and collected in a shared public registry. Developers committed to disclosing known model capabilities, limitations and risks, supporting independent evaluations and providing mechanisms to stop models. Governments intend to strengthen early warning about serious cyber incidents and loss of control, create protected channels for whistleblowers and clarify liability for cross-border harm. The declaration remains open to additional signatories.

Sources 1

Teaching a robot a new job may now require only a few short examples and half an hour of fine-tuning on a consumer graphics card: different robots are beginning to learn skills from one another and directly from people.

NVIDIA GR00T 2 masters new robot tasks from five demonstrations with 95% success

Teaching a robot a new job may now require only a few short examples and half an hour of fine-tuning on a consumer graphics card: different robots are beginning to learn skills from one another and directly from people.

NVIDIA has introduced the open GR00T 2 model at CoRL, which predicts future video and robot actions together. In the company's own tests on 50 tasks the target robot had not previously performed, the model achieved an average success rate of 40% without demonstrations and completed 40 tasks at least once. One demonstration raised the average result to 71%, and five to 95%; the five recordings took less than three minutes, while LoRA fine-tuning took 30 minutes on a single RTX 5090. The model transfers skills from videos of people, the handheld UMI gripper and robots to other robot bodies. Pretraining used 402,730 hours of egocentric video, 10,184 hours of UMI demonstrations and 23,500 hours of robot teleoperation. Paired with a reasoning agent that breaks work into steps and recovers from failures, GR00T 2 assembled three toys in 292 of 300 continuous trials. The lighter GR00T 2 Turbo runs directly on the Unitree G1 humanoid's Jetson Thor: a prediction takes 124 ms, and clearing a table of assorted objects succeeded in 93.3% of trials. All figures are NVIDIA's results on its own task sets.

One trainable model could replace many specialized simulators and accelerate the study of materials, molecules and other physical processes.

WorldParticle simulates sand, fluids and proteins with one neural network

One trainable model could replace many specialized simulators and accelerate the study of materials, molecules and other physical processes.

Researchers from MIT, UBC and other universities have introduced WorldParticle, a unified model for physical simulations. It breaks an object into thousands or millions of particles, receives their coordinates and properties, and successively predicts each particle's state. Instead of using a prewritten equation for a specific material, the neural network learns general dynamics, while external forces are added separately before each step. To scale, the model merges and compresses particles while preserving details of their behavior. In demonstrations, WorldParticle covered six classes of phenomena, including deformable objects, granular materials and proteins. According to the authors, it was more accurate than the numerical methods used for comparison and simulated changes in protein structures about 100 times faster than numerical simulators. The work will be presented at SIGGRAPH Asia; the researchers see it as a first step toward a universal foundation for scientific simulations.

Sources 1

Stopping a hazardous research run no longer depends on the agent's own code: access to model computation can be cut off from a single point.

OpenAI deploys Control Room to centrally stop research models

Stopping a hazardous research run no longer depends on the agent's own code: access to model computation can be cut off from a single point.

According to Unite.AI's account, Control Room associates every research workload with an identifier and an approved safety case. Before returning a result, an inference engine requests authorization from the central service: after a stop command, the workload's next request to any model is refused, although the research process itself may continue running. The system covers training, evaluation and data processing. A person or a monitor model can initiate a stop; according to OpenAI as relayed by Unite.AI, authorization caching means up to 35 seconds can pass between the command and the point when further progress becomes impossible. If Control Room fails or is disabled, the integrations stop returning responses. At the time of the announcement, this fail-closed mode covered all research workloads using frontier models.

A single operator can direct not just one autonomous aircraft but a group of machines that coordinate their own maneuvers.

Hivemind coordinates multiple autonomous US Air Force combat aircraft as a group for the first time

A single operator can direct not just one autonomous aircraft but a group of machines that coordinate their own maneuvers.

Shield AI reported two demonstrations under the Collaborative Combat Aircraft program. On October 8, an F-22 pilot issued high-level directives to an autonomous FQ-44A Fury, and on October 26, several CCAs running Hivemind conducted a formation flight and simulated beyond-visual-range air combat maneuvers. Software on each aircraft translated the operator's intent into coordinated group actions. The tests used the A-GRA government reference architecture, which separates mission autonomy from the functions of a specific airframe through standardized interfaces. According to Shield, the same Hivemind foundation has already been integrated into more than 40 uncrewed systems across different domains.

Sources 1

SpaceX plans to pair a large computing center with its own energy source, reducing the project's dependence on the external power grid.

SpaceX proposes a $24 billion data center and gas-fired power plant in Starbase

SpaceX plans to pair a large computing center with its own energy source, reducing the project's dependence on the external power grid.

According to Quartz, citing the City of Starbase's public hearing notice, SpaceX intends to build, own and operate a gas-fired power station and supercomputing complex itself. The initial phases are estimated at $24 billion; the project may be implemented in phases and expanded. The complex is proposed for a roughly 444-acre site east of Brownsville in the Cameron County Enterprise Zone. On December 9, the city commission will consider a property tax abatement agreement and a separate development agreement for the project.

Scrutiny of circumstances surrounding the key SI chipmaker has extended beyond its board to export-control and national-security questions.

Trump says FBI is examining circumstances surrounding Nvidia CEO's relationship with Alice Huang

Scrutiny of circumstances surrounding the key SI chipmaker has extended beyond its board to export-control and national-security questions.

Donald Trump told reporters that the FBI is examining the situation involving Jensen Huang and Megaspeed executive Alice Huang, whom US officials considered linked to Chinese intelligence, according to the New York Times. Trump said Nvidia's CEO told him he had not known about the suspicions. The FBI and Justice Department declined to comment to the BBC and NBC. According to an NBC source, a separate Justice Department criminal investigation concerns possible shipments of Nvidia chips by Megaspeed to customers in China and the actions of people associated with the company; it is being conducted by the National Security Division, the FBI and the Commerce Department's Bureau of Industry and Security. China's Foreign Ministry called the espionage allegations groundless. Nvidia previously launched its own independent review and said it had no reason to believe company secrets or national security had been compromised.

A model with trillions of total parameters can now be tested and modified in a private environment without hosted API restrictions, although full deployment still requires infrastructure with terabytes of memory.

Meituan releases multimodal LongCat-2.5 weights with 1.6 trillion parameters under the MIT License

A model with trillions of total parameters can now be tested and modified in a private environment without hosted API restrictions, although full deployment still requires infrastructure with terabytes of memory.

Meituan has published the full LongCat-2.5 weights on Hugging Face under the MIT License, one week after launching the hosted API. The multimodal MoE model contains 1.6 trillion parameters, approximately 48 billion of which are active per token. It accepts text, images and video and is designed for long-horizon agent tasks involving codebases, complete projects, design and 3D. The model checkpoint occupies about 1.8 TB, so local deployment requires the full weight set even though only some experts are active. In Meituan's own measurements, LongCat-2.5 scored 88.4 on Terminal Bench v2.1 versus 70.8 for LongCat-2.0 and 64.6 on DeepSWE v1.1 versus 9.7. VentureBeat notes that comparisons with newer proprietary models on Terminal Bench v4.0 no longer favor it and that the test settings differ. The license permits commercial use, modification and redistribution with the required notice. W4A4 quantization of experts and activations is available for Chinese accelerators; Meituan's own stress tests show prefill throughput increasing by 43.3% and decode throughput by 42.77% relative to W8A8.

A household robot could learn a person's preferences from a couple of ordinary sentences instead of needing a manually written rule for every object.

MIT teaches robots to generalize people's everyday warnings

A household robot could learn a person's preferences from a couple of ordinary sentences instead of needing a manually written rule for every object.

FLEX, developed by MIT and New York University, turns individual human remarks into broader rules for robot movement. After the phrase "don't spill coffee on the laptop," the language model links liquid with risks to electronics, and the robot also starts avoiding a nearby phone despite receiving no separate instruction about it. FLEX works in three stages: it adds the named object to a list of constraints, generalizes its properties using GPT-4o, and removes items no longer visible from the list. The cycle takes about five seconds. In robotic-arm simulations, volunteers tested three tasks involving coffee, a candle and a bowling ball; the system extended warnings to electronics, flammable objects and fragile items. The results apply to virtual scenarios and will be presented at the Conference on Robot Learning.

Sources 1

Biological SI models can be tested not only on their overall score, but also on whether combining data actually reveals new properties of cells.

Nature Machine Intelligence paper proposes measuring the real benefit of combining biological models

Biological SI models can be tested not only on their overall score, but also on whether combining data actually reveals new properties of cells.

Researchers have shown that combining tissue-image and gene-expression models does not necessarily produce new information: standard alignment methods may merely reproduce knowledge already present in the strongest individual modality. To test this, they introduced the Synergistic Information Score (SIS), the relative improvement of a combined representation over the best unimodal model when using the same linear classifier. The authors compared ten fusion methods on three spatial transcriptomics datasets: fibrotic lungs, the developing thymus and breast tumors. Theoretical analysis and experiments revealed a "spectral ceiling" for alignment methods: they recover shared linear signals well but lose nonlinear complementary information. In tasks requiring morphology and gene expression to be combined, methods such as CoMM overcame this limit. The evaluation and comparison code has been published.

Sources 1

Data centers will be able to combine different types of SI accelerators in one system instead of building their entire infrastructure solely on one vendor's GPUs.

Nvidia, SK Telecom and Rebellions to combine GPUs and Korean NPUs in one computing system

Data centers will be able to combine different types of SI accelerators in one system instead of building their entire infrastructure solely on one vendor's GPUs.

The companies have signed a memorandum to create heterogeneous SI infrastructure for SK Telecom data centers. Nvidia GPUs will work alongside neural processing units from South Korean startup Rebellions through NVLink Fusion, which lets customers connect their own accelerators to Nvidia's rack-scale infrastructure. The agreement was signed at AI Day Seoul with support from the Ministry of Science and ICT. The ministry expects the project to expand the use of Korean chips and help local companies enter the global market.

Apple has completed another mandatory step toward bringing its SI features to Chinese tablets and computers through local models.

Apple Intelligence registered in China for iPad and Mac

Apple has completed another mandatory step toward bringing its SI features to Chinese tablets and computers through local models.

Shanghai's cyberspace regulator has added tablets and laptops to the Apple Intelligence registration, Reuters reports. The service was previously approved for iPhones in July. Chinese rules require public generative SI services to be registered before launch. The Chinese version of Apple Intelligence uses Alibaba's Qwen models, with Baidu also serving as a technology partner. Qwen powers Siri, Writing Tools and visual search. The first features on compatible iPhones are expected with iOS 27.3 in December, while the new registration expands preparations for launch to iPadOS and macOS.

The regulator will assess not whether an update was issued, but whether it eliminated the driverless vehicle's dangerous behavior in real road situations.

NHTSA to review fix for 2,895 Waymo robotaxis after new school-bus violations

The regulator will assess not whether an update was issued, but whether it eliminated the driverless vehicle's dangerous behavior in real road situations.

The US road-safety regulator has opened a review of the effectiveness of Waymo's recall update, Reuters reports. The cases involve vehicles that initially slowed or stopped near a school bus but then drove past it while its red lights were flashing or its stop arm was extended. According to NHTSA, most such incidents occurred after November 2025, when Waymo said it had completed the software fix for all affected vehicles. The review covers 2,895 vehicles and will determine the causes of post-update violations and whether the measures were sufficient.

An expensive frontier agent can delegate high-volume simple operations to a cheaper model while reserving most computation for difficult decisions.

Meta releases Muse Ember, a small model for agents' routine work

An expensive frontier agent can delegate high-volume simple operations to a cheaper model while reserving most computation for difficult decisions.

Muse Ember is designed for tool calls, request routing, data extraction, small coding tasks and processing long documents. The model accepts text, images, audio and video, supports a context of up to one million tokens and is available through the Meta Model API and Muse Code at $0.10 per million input tokens and $0.50 per million output tokens. Ember scored 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5; according to figures supplied by Meta, running the evaluation cost about $0.14 per task versus $0.41 for Haiku. Ember scored 37.8% on AutomationBench but trailed Haiku on FrontierCode 2.0, which evaluates real code changes: 27.2% versus 34.9%. The model weights remain closed.

A generated image can be refined for longer as a working layout without being recreated from scratch after every edit.

ByteDance introduces Seedream 5.5 with stable image editing across up to 100 rounds

A generated image can be refined for longer as a working layout without being recreated from scratch after every edit.

The new model preserves unedited regions at pixel level and reduces the accumulation of artifacts across successive edits. In ByteDance's demonstration, clothing, accessories and objects in a single image were changed 100 times in succession, with each previous output serving as the basis for the next round. Seedream 5.5 also directly generates RGBA images with full or partial transparency: a finished element can be transferred into a layout without separately removing the background. The company also reports more natural materials and lighting and more accurate rendering of dense, small text.

The market priced Nvidia's governance and reputational risk at hundreds of billions of dollars, although part of the intraday decline was recovered by the close.

Nvidia shares close down 4.2% amid investigation involving its CEO

The market priced Nvidia's governance and reputational risk at hundreds of billions of dollars, although part of the intraday decline was recovered by the close.

Nvidia's market capitalization fell by about $250 billion by Monday's close, the BBC estimates; its shares were the worst performers in the Dow and Nasdaq. The decline was steeper at the open: according to The Guardian as relayed by Rossiyskaya Gazeta, it reached nearly 7.5%, equivalent to about $335 billion in market capitalization. The sell-off followed the board's decision to launch an independent review of circumstances involving Jensen Huang. Other chipmakers gained, while the broader market began the week higher. Crossmark Global Investments analyst Victoria Fernandez linked the resilience of long-term expectations to Nvidia's central role in financing and supplying the SI industry.

A prominent supplier of human labor for SI training and moderation is leaving the market under pressure from falling revenue and litigation costs.

Meta and OpenAI contractor Sama begins liquidating its assets

A prominent supplier of human labor for SI training and moderation is leaving the market under pressure from falling revenue and litigation costs.

According to a letter to creditors reviewed by TIME, Sama is winding down operations and selling its assets. The company attributed the decision to declining revenue amid competition and class-action litigation that would have required substantial resources and years to defend. Its unaudited October balance sheet showed about $1 million in assets and $17 million in liabilities. On October 23, Samasource Impact Sourcing transferred its assets to Sama ABC, which was created to sell them and distribute the proceeds among creditors; a bank has a priority claim on substantially all assets. Sama performed content moderation for Meta and data labeling for OpenAI. In Kenya, former moderators are challenging working conditions and dismissals; Sama said it complied with local law, while Meta denied responsibility for working conditions. In the United States, some plaintiffs in privacy lawsuits concerning data from Meta's smart glasses voluntarily dropped Sama as a defendant in August.

Sources 1

Uncrewed boats now have a permanent operational unit alongside the crews of traditional ships: autonomous systems are moving from isolated trials to round-the-clock naval service.

Royal Navy forms its first autonomous boat squadron for front-line operations

Uncrewed boats now have a permanent operational unit alongside the crews of traditional ships: autonomous systems are moving from isolated trials to round-the-clock naval service.

The Royal Navy formally commissioned the 1st Autonomous Squadron at HMNB Portsmouth on November 5 and published the announcement on November 9. The unit of approximately 180 personnel is split into two watches for round-the-clock operations and will use uncrewed boats in front-line operations, both independently and as part of groups with allies. At the ceremony, Kraken K3 Scout and Rattler boats were manoeuvred around Portsmouth Harbour from a control hub. The squadron replaces the experimental Project Beehive and will develop skills, tactics and the introduction of new uncrewed systems with Navy Develop and the Capability Accelerator. The transition to a hybrid fleet, in which traditional ships and units operate alongside uncrewed and autonomous technology, is backed by £1.5 billion of investment over four years.

Teen protections will depend on an account's actual behavior rather than solely on the age entered during registration.

Meta expands Instagram's SI age estimation worldwide

Teen protections will depend on an account's actual behavior rather than solely on the age entered during registration.

Instagram will begin identifying accounts worldwide that belong to teenagers who entered an adult date of birth and automatically switching them to Teen Accounts. These accounts restrict contact, available content and time spent on the service. Meta estimates that the change will affect millions more teenagers; similar technology continues to roll out on Facebook and will arrive on Threads later this year. In Australia, Meta has also begun using Apple's age-range signal, which the app store shares with the developer with consent when the device is set up. The company links the expansion to an agreement with 52 attorneys general from US states and territories to improve identification of children under 13.

Sources 1

An agent-built app can become a working service with a database, user sign-in and payments more quickly through familiar web infrastructure.

Automattic launches Spacefast to publish apps built by SI agents

An agent-built app can become a working service with a database, user sign-in and payments more quickly through familiar web infrastructure.

The owner of WordPress.com has introduced Spacefast, a platform providing SI-built sites and apps with databases, authentication, file uploads, email and analytics. Projects can connect ChatGPT, Claude, Cursor, Grokbot, Muse and other agents or upload a Git repository. Unlike integrated builders such as Lovable, the service lets users choose a separate tool to generate code. Following a soft launch in September, Spacefast was used for internal mockups, presentations and promotional websites. Automattic plans to connect the platform with WordPress, WooCommerce and its own content-delivery infrastructure, including for companies seeking to centralize disparate projects across their teams.

The ride-hailing service gains computing and validation infrastructure to move from driver assistance to fully autonomous rides.

Kakao Mobility and Nvidia to develop Level 4 autonomous driving

The ride-hailing service gains computing and validation infrastructure to move from driver assistance to fully autonomous rides.

Kakao Mobility has signed a memorandum with Nvidia to jointly develop an SAE Level 4 autonomous-driving system, under which a vehicle operates without a driver in defined conditions. The company intends to combine its end-to-end driving models with Nvidia's computing platform and use the Halos safety system to run, develop and validate the technology. The collaboration aims to prepare a global commercial service and coordinate with software, component and vehicle manufacturers.

If the claimed specifications are confirmed in samples, large language models will be able to run directly in robots and laptops with modest energy consumption.

DeepX tapes out 5-watt DX-M2 chip for local models with up to 100 billion parameters

If the claimed specifications are confirmed in samples, large language models will be able to run directly in robots and laptops with modest energy consumption.

South Korea's DeepX has completed the design of its second-generation DX-M2 accelerator and sent it to manufacturing. According to the company, the 2nm chip is designed for models with 20–100 billion parameters and 3–20 billion active parameters, generation at 10–30 tokens per second and 5 W power consumption. It delivers 80 TOPS and 150 GB/s of memory bandwidth. Samples are expected in March 2027, with mass production planned for autumn 2027. DeepX and Samsung are also testing LPDDR5x-PIM memory, which performs some operations inside memory chips. Simulation showed Llama 3.1 8B running 6.7 times faster than on a single DX-M2 while increasing power consumption 2.7 times; the company expects a twofold improvement without increasing the power budget. These are simulation results for the future combination, not measurements of a finished device.

Sources 1

The Russian research cluster has gained fault-tolerant storage based on domestic processors for lengthy calculations and model training.

Moscow State University supercomputer receives all-Russian storage system

The Russian research cluster has gained fault-tolerant storage based on domestic processors for lengthy calculations and model training.

Astra Group and Spintronica have deployed SDS TROK storage on servers with Baikal S processors for the Istok computing complex. The system provides data over NFS and uses synchronous DRBD replication: a write is confirmed after the data is saved on a specified number of nodes, and only changed blocks are synchronized when a failed server returns. Software and hardware compatibility was tested first on virtual machines and then on physical equipment. The project received financial support from Russia's Ministry of Science and Higher Education. More than ten groups from Russian Academy of Sciences institutes already use the cluster; full-scale scientific computing is planned from 2027.

Sources 1

Surgeons will be able to assess anatomy and the consequences of an operation more accurately in advance, preserving more healthy lung tissue.

Sechenov University begins testing SI for planning lung surgery

Surgeons will be able to assess anatomy and the consequences of an operation more accurately in advance, preserving more healthy lung tissue.

The pilot of the university's joint development with RTK Radiology will last at least a month and use anonymized CT scans from at least 30 patients. The system identifies tumors, lung lobes and segments, bronchi, arteries and veins, builds an interactive 3D model and calculates predicted respiratory function after removal of part of the organ. Doctors will compare the automated analysis with expert assessments; at this stage, the module is not used to make treatment decisions. During initial trials, lung-tissue segmentation accuracy reached 93% by DICE, with 96% coverage of bronchial-tree branches and 97% of the vascular tree. Deployment in clinical practice is planned for 2027 following further evaluation.

Sources 1

Automating editing, design and distribution is becoming a substantial business and shortening the path from manuscript to publication to a few weeks.

Publishing platform Spines is raising $64 million

Automating editing, design and distribution is becoming a substantial business and shortening the path from manuscript to publication to a few weeks.

Spines is receiving $29 million in equity from Zeev Ventures, Aleph, Dan Senor and other investors, alongside $35 million in debt financing from the Twelve fund. The platform's SI agents edit and proofread manuscripts from self-publishing authors, create covers and layouts, and then Spines distributes electronic, audio and print books. The company's annual recurring revenue has reached $50 million, compared with $3.5 million at the time of its Series A two years ago. Its catalog includes 8 thousand authors and more than 12 thousand books. The platform does not accept books written entirely by SI, while machine translations are reviewed by native speakers; within six months, Spines plans to expand translation to 20 languages.

Bipartisan work on SI safety rules is possible in Washington, but the president ties their acceptability to preserving the US technological advantage.

Trump open to SI safeguards provided the US retains its lead

Bipartisan work on SI safety rules is possible in Washington, but the president ties their acceptability to preserving the US technological advantage.

In a 60 Minutes interview, the US president said SI risks can be managed and responded affirmatively when asked about possible cooperation with Democrats on safety. "I don't mind guardrails, but I don't want to put us in a deep second position," he said, contrasting the US approach with rules in China. Trump also said the Department of Justice addresses SI-related risks and described threatening a major SI company's leader with imprisonment, without naming the person or company.

Sources 1

An accurate motor model transferred simulation-trained control to unusual robot mechanics without manually programming every movement.

Boston Dynamics puts Atlas on wheels to test a learned controller on real motors

An accurate motor model transferred simulation-trained control to unusual robot mechanics without manually programming every movement.

Engineers replaced the research Atlas's feet with scooter wheels using a single custom adapter on the existing ankle actuator. In the robot model, each wheel was described as a continuously rotating joint. This is not a new commercial configuration but a testbed for actuators and transferring a controller from simulation to physical hardware. The wheels force the motors to operate continuously at high torque and speed during acceleration, turns and stops. The team calibrated the simulator against measurements of the real motors and changed the rewards and command range used in reinforcement learning. The robot then learned to use rolling during side dashes and performed a 360-degree jump by following a recording of a human movement. Boston Dynamics is applying the findings to Atlas's standard configuration, including controller resilience to actuator failures.

A major rights holder can train creative tools on its own materials and integrate them into production rather than relying solely on general-purpose generators.

Disney and Adobe will build custom generative models around Disney characters and worlds

A major rights holder can train creative tools on its own materials and integrate them into production rather than relying solely on general-purpose generators.

The multiyear agreement makes Adobe Disney's official SI innovation partner. Disney will receive early access to Firefly, Firefly Foundry, Premiere and Photoshop, as well as Adobe's engineering and research resources. Through Firefly Foundry, the companies will adapt proprietary models to Disney's character attributes, environments and visual styles to create materials suitable for commercial production. Walt Disney Imagineering already uses a model that turns hand-drawn sketches into 2D concept art and generates branded assets for its parks. The collaboration also covers the film studio's StudioLAB and the ESPN Edge center, which will explore new sports-media formats and ways to work with ESPN's archive.

Nvidia is strengthening not only chip supplies in South Korea, but also local development of models, robots and new interconnects for computing clusters.

Nvidia to invest in four South Korean SI startups

Nvidia is strengthening not only chip supplies in South Korea, but also local development of models, robots and new interconnects for computing clusters.

At a forum with South Korea's Ministry of Science and ICT, Nvidia announced an investment in Point2 Technology and plans to invest in Motif Technologies, RLWRLD and Upstage. Point2 develops radio-frequency interconnects between chips and racks in SI clusters; Motif creates open language models and agents; RLWRLD trains robotics models for delicate manipulation; and Upstage develops Solar models and document processing. The meeting also included more than 20 South Korean startups working on foundation models, physical SI, infrastructure and semiconductors. The investments expand Nvidia's work with the local ecosystem, where more than a thousand companies participate in the Inception program.

Autonomous cleaners are entering everyday municipal work and filling a specific staffing shortage rather than only being demonstrated at exhibitions.

Kazan property managers buy 28 robots to clean courtyards

Autonomous cleaners are entering everyday municipal work and filling a specific staffing shortage rather than only being demonstrated at exhibitions.

Kazan's mayor's office said the autonomous machines will be delivered within a month to address a shortage of janitors. The robots will clear snow in winter and sweep grounds in summer; the mayor's office estimates that one robot can replace up to ten people with appropriate work organization. Operators have already been trained to maintain the equipment. Large property-management companies employ more than 1.3 thousand janitors, compared with an estimated requirement of nearly 2.2 thousand. In addition to the robots, 239 units of other specialized equipment have been prepared for winter.

Sources 1

A missed phone call can become a full conversation with a secretary that relays both the recording and the substance of the request.

T2 launches generative voice secretary Lina

A missed phone call can become a full conversation with a secretary that relays both the recording and the substance of the request.

Lina answers forwarded incoming calls when a subscriber is busy, unavailable or offline. Instead of following predefined scenarios, a generative language model conducts the conversation, asks clarifying questions and responds to unexpected remarks. After the call, the subscriber receives a full transcript and a concise summary in the T2 app. Transcripts from Lina and the operator's other assistants can also be sent to a bot in Max. The service is activated through the SafeWall section of the mobile app.

Sources 1

Computational screening will help identify suspicious chemicals earlier and direct costly laboratory tests toward the highest-priority substances.

Sechenov University creates SI screening for substances that may harm fertility

Computational screening will help identify suspicious chemicals earlier and direct costly laboratory tests toward the highest-priority substances.

The model predicts whether a chemical compound will bind to the progesterone receptor, activate or block it, and whether the effect is related to general cellular toxicity. Training used Tox21 data on more than seven thousand compounds with known experimental results. The system also assesses how similar a new substance is to the training set: the prediction is considered less reliable when differences are substantial. This research prototype selects compounds that should be tested first rather than replacing laboratory experiments. The next stage is expanded validation on independent samples and integration with other tools for assessing the effects of steroid hormones.

Sources 1

Corporate teams gain agentic programming inside their own perimeter, with the model's actions tied to agreed requirements.

Astra integrates secure SI agent into OpenIDE

Corporate teams gain agentic programming inside their own perimeter, with the model's actions tied to agreed requirements.

OpenIDE Pro 2026.2.3 has gained the Meta Harness tool for calling Astra SI Code directly from the development environment. The agent analyzes the project, writes and corrects code, runs checks and operates inside the customer's perimeter without sending code or queries outside it. The integration supports specification-driven development: the engineer and agent record requirements, after which the agent executes a plan through verifiable steps. Files, folders, current changes and problems found by the environment can be added to the context, while comments for the agent can be left directly in the code. Parallel tasks run in separate chats while the developer continues using the debugger, Git and Docker Compose.

Sources 1

Corporate video calls gain verification not only of permission to join, but also of the authenticity of the person on camera during the conversation.

IVA MCU begins checking video-call participants for deepfakes

Corporate video calls gain verification not only of permission to join, but also of the authenticity of the person on camera during the conversation.

The IVA MCU corporate videoconferencing platform has integrated a VisionLabs detector. The service selects frames at an interval set by the administrator, sends them to a model deployed in a secure perimeter and displays a "real/fake" result in the participant list. If detections persist, a moderator can restrict audio and video for that participant, verify their identity and remove them if necessary without stopping the rest of the session. The model detects face swaps, expression transfer and synthesized faces, working alongside liveness checks and virtual-camera detection. According to VisionLabs, accuracy in production reaches 99.3% by ROC-AUC, while basic CPU image processing takes fractions of a second. A separate MWS AI module can supplement the check with audio-deepfake analysis.

Sources 1

Schools must choose not only between banning and allowing SI, but also between access to familiar search infrastructure and alternative ways of finding knowledge.

Parents and teachers criticize Google Search cutoff for schools that block SI

Schools must choose not only between banning and allowing SI, but also between access to familiar search infrastructure and alternative ways of finding knowledge.

Google said last week that Search, Gemini, Notebook and Gemini in Classroom would be entirely disabled for K-12 school accounts until an administrator explicitly enables them. As a result, blocking SI tools also affects search. The company explains this as giving schools unified control over access. Parents and teachers interviewed by The Guardian criticized the bundling of services, the growing influence of large technology companies and the potential harm of SI to critical thinking; some proposed switching to other search engines and academic databases. Other teachers opposed blocking SI and suggested using Gemini in lessons to teach students how to verify sources.

Authors and model developers could gain clearer rules for using protected texts instead of the current uncertainty.

Russian State Duma plans to regulate copyright for SI use by March 1

Authors and model developers could gain clearer rules for using protected texts instead of the current uncertainty.

Pavel Krasheninnikov, chair of the State Duma Committee on Legislation, told journalists about a plan to address copyright protection when using SI by March 1, 2027. He cited a model's use of a published article and its subsequent reproduction in altered form as examples. Krasheninnikov considers it necessary to establish limits on such use through general legislation, while emphasizing that SI is an important tool that should be used.

Sources 1

SI becomes part of the editor itself, while authors gain a way to show readers which materials and sources were used in their work.

Medium launches TK editor with an SI assistant and a record of how text was written

SI becomes part of the editor itself, while authors gain a way to show readers which materials and sources were used in their work.

In the TK web app, typing "@TK" invokes an assistant directly inside a draft. It answers questions, finds saved fragments, reviews headlines and text against editorial style and suggests topics. An author can connect their own agent through an MCP server or Claude skill, giving it memory of the author's projects. Authors can publish notes explaining how SI was used, including prompts and source materials, making the text's provenance inspectable. Early drafts can also be shared with people for feedback. TK is free for Medium account holders and supports Markdown import and export and access controls for collaboration. A desktop version is planned.

Apple's cloud-based SI features now offer a way to continue after the free monthly limit without buying a more expensive subscription.

Apple begins selling additional Apple Intelligence usage for $49.99

Apple's cloud-based SI features now offer a way to continue after the free monthly limit without buying a more expensive subscription.

Logic Pro is the first to receive a one-time Intelligence Usage Pack: after reaching the monthly limit, users can buy additional access to cloud-based SI features. The pack is tied to the purchaser's Apple Account and does not support Family Sharing; unused usage rolls over between months and expires 12 months after purchase. The option is available to Creator Studio subscribers in the United States, United Kingdom, Canada, Brazil, France, Germany, India and Mexico. Logic Pro's limit depends on the complexity of conversations with its assistant, which can help start a composition, set up a vocal track or modify the current project.

The Japanese frontier-model developer is preparing a path to public capital that could finance further research and computing.

Sakana AI considers a stock market listing as early as fall 2027

The Japanese frontier-model developer is preparing a path to public capital that could finance further research and computing.

According to Kyodo and Nikkei as relayed by Sweden Herald, Japanese lab Sakana AI is considering an IPO in Tokyo or on Nasdaq no earlier than fall 2027. The company has hired Goldman Sachs as an adviser for the listing. Sakana AI was founded by David Ha, Ren Ito and Llion Jones and develops its own models in Japan.

Data on rare and widespread languages becomes available for visual exploration, helping account for linguistic diversity when building SI.

Google launches Language Explorer, an interactive map of linguistic diversity

Data on rare and widespread languages becomes available for visual exploration, helping account for linguistic diversity when building SI.

One online platform, Language Explorer, brings together data on thousands of languages from the open LinguaMeta dataset and lets researchers, educators and general users explore them through visualizations. The project is part of the 1,000 Languages initiative. An interactive version is featured in the Real Talk! exhibition at Berlin's Museum of Communication; the exhibition runs through February 2028 and then continues in Frankfurt. The museum's companion online material opened on November 6, and Language Explorer itself is available on the web.

Adopting ChatGPT has become easier for large companies: the purchase goes through an established cloud agreement and consolidated AWS billing rather than a separate new procurement channel.

ChatGPT Enterprise arrives on AWS Marketplace with billing through an existing Amazon account

Adopting ChatGPT has become easier for large companies: the purchase goes through an established cloud agreement and consolidated AWS billing rather than a separate new procurement channel.

According to OpenAI's description, organizations can now buy and administer ChatGPT Enterprise through AWS Marketplace, using their existing AWS procurement and billing arrangements and applying eligible spending toward their Amazon commitments. The subscription includes the same models and features available directly from OpenAI, including ChatGPT, ChatGPT Work, Codex and dots, with new capabilities released at the same time. Eligible customers can also access ChatGPT for Healthcare and ChatGPT for Financial Services. Developers can separately build applications and agents on OpenAI models through Amazon Bedrock.

Sources 1