Show only each day's banner story

Brookfield and Drax plan an SI data center exceeding 1 GW in North Yorkshire

British researchers and companies could gain domestic computing capacity on the scale needed for frontier models, but the project still needs approvals and an anchor customer.

Drax, Brookfield and its cloud platform Radiant have agreed to jointly develop a computing complex at a power station site in North Yorkshire. The first phase of up to 100 MW is due to come online in Q1 2028, another 500 MW is planned for 2028–2031, and the site is ultimately intended to exceed 1 GW. Brookfield estimates total investment, including chips, at more than £20 billion ($26.8 billion), including £2.5 billion for the first phase. The initial launch depends on planning and grid-connection approvals and on securing a tenant; a final investment decision has not yet been made. The complex will initially draw power from the national grid, with direct supply from Drax expected by the early 2030s.

Sources 3

cnbc.comtheguardian.comaa.com.tr

Access to advanced computing increasingly depends not only on buying chips, but also on which customers foreign clouds are allowed to serve.

Nvidia asks cloud partners to reduce their share of Chinese customers

Access to advanced computing increasingly depends not only on buying chips, but also on which customers foreign clouds are allowed to serve.

According to The Information, Nvidia has recommended that some cloud customers, including neocloud providers it backs, take fewer orders from Chinese companies and disclose the end users of computing capacity. The Decoder reports that the recommendation is not mandatory and applies to future rather than existing contracts. According to the Financial Times, it was communicated verbally through sales representatives and regional teams. Renting capacity abroad provides access to advanced chips without importing them into China, while current US export rules primarily restrict hardware shipments rather than remote access.

Sources 2

the-decoder.com3dnews.ru

SI accelerates the discovery of infrastructure weaknesses, making timely updates and basic security even more important; the same capabilities help defenders find vulnerabilities.

UK NCSC considers an SI-powered malware attack on critical infrastructure highly likely by 2028

SI accelerates the discovery of infrastructure weaknesses, making timely updates and basic security even more important; the same capabilities help defenders find vulnerabilities.

The UK National Cyber Security Centre's annual State of the Cyber Nation report assesses that SI-powered malware is highly likely to be used in a ransomware attack on the country's critical infrastructure by 2028. The NCSC expects SI primarily to increase the scale and efficiency of established methods: vulnerability discovery, social engineering and extortion. NCSC chief executive Richard Horne urged organisations to address basic weaknesses, including outdated systems, unnecessary internet connections and underinvestment in security. The agency also highlighted SI's defensive potential: the government previously announced £60 million for the first phase of Cyber Shield, £70 million for the NERV vulnerability-discovery centre and a new centre of excellence for SI cybersecurity in Manchester.

Sources 3

ncsc.gov.uktheguardian.comeuronews.com

SI has moved from simulation to controlling a production fighter in test scenarios; the next step is to verify the agents' reliability across a broader range of flights.

Saab tests its own SI agents on a Gripen E fighter jet

SI has moved from simulation to controlling a production fighter in test scenarios; the next step is to verify the agents' reliability across a broader range of flights.

Saab has unveiled a family of in-house SI agents for future aviation systems and confirmed that they have flown on a Gripen E. According to the company, an agent autonomously controlled the aircraft during tests in beyond-visual-range air combat scenarios. The agents are intended to make tactical decisions, assist pilots and operate unmanned platforms alongside crewed aircraft. They are being developed through the AI Fighter Pilot Academy: training takes place in simulations spanning millions of virtual flight hours, and the infrastructure supports new scenarios and platforms. Saab says it controls the entire chain from training and verification to flight testing; computing capacity is provided by Sweden's Sferical AI initiative.

Sources 2

saab.combreakingdefense.com

A model evaluation is not a permanent property of the technology: repeating measurements on new versions helps distinguish durable patterns from limitations that have already disappeared.

Stanford researchers show that findings about language models quickly become outdated

A model evaluation is not a permanent property of the technology: repeating measurements on new versions helps distinguish durable patterns from limitations that have already disappeared.

A research group led by James Zou retested 50 widely cited findings about language models using 20 models from eight developers released between July 2024 and September 2026. Most of the claims either no longer held or showed substantially weaker effects within 15 months of publication. One example is the well-known finding that GPT-3.5 fabricated 55% of its bibliographic citations: the newer GPT-6.1-Sol and Claude Sonnet 5.5 rarely produced such citations for the same prompts, although some models continued to make errors. The study covered sycophancy, guardrail bypasses, long context and cultural bias, showing that measurements must be tied to a specific generation of models.

Sources 1

zou-group.github.io

In future chatbot disputes, the key issue may be whether an answer replaces the original article by reproducing its text rather than merely using facts; a final ruling is still to come.

EU court adviser ties Gemini copyright infringement to reproducing articles' expression, not facts

In future chatbot disputes, the key issue may be whether an answer replaces the original article by reproducing its text rather than merely using facts; a final ruling is still to come.

Maciej Szpunar, an advocate general at the Court of Justice of the EU, has issued an opinion in Hungarian publisher Like Company's case against Google. He proposed considering whether the chatbot accessed articles unlawfully, for example by bypassing access restrictions, and whether its answer reproduces the publication's recognisable form of expression to such an extent that the user no longer needs to open the original. Facts reported in the news are not themselves protected by the publisher's right. Szpunar also noted that model training is permitted under the EU's text and data mining exception, and that the case contains no evidence that Like Company's articles were used to train Gemini. The opinion is not binding on the judges. Google said Gemini learns linguistic patterns to produce original answers and that rights holders can opt out of having their material used.

Sources 3

infocuria.curia.europa.eumorningstar.comturkiyetoday.com

OpenAI has gained a leader with Army modernization experience who will translate government agencies' requirements into tasks for its technologies.

OpenAI hires former US Army Secretary Dan Driscoll for national security work

OpenAI has gained a leader with Army modernization experience who will translate government agencies' requirements into tasks for its technologies.

Driscoll has taken the newly created role of senior counselor for national security modernization. He will work with OpenAI's leadership, government agencies, companies, universities and international partners, supporting the use of SI in national security, cybersecurity and defence. Within the company, he will explain the needs of relevant government bodies to product teams; externally, he will advise on changes to policies and strategies. Driscoll previously oversaw US Army modernization and resource allocation. In a statement about his move, he linked it to the potential to accelerate information processing and reduce administrative work for security personnel.

Sources 2

cnbc.comaa.com.tr

Companies are gaining tools not only to observe and test agents in advance, but also to disable them immediately when their behaviour becomes dangerous.

CrowdStrike acquires Fiddler AI and Dreadnode and launches an emergency stop for SI agents

Companies are gaining tools not only to observe and test agents in advance, but also to disable them immediately when their behaviour becomes dangerous.

Both deals were announced at the Fal.Con Europe conference and are expected to close in CrowdStrike's fiscal Q4 2027, subject to customary conditions. Fiddler tracks agents' actions and the reasons for their decisions; its technology will become part of Falcon Observe. Dreadnode tests models, applications, agents and MCP servers using more than 100 attack methods, including prompt injection; its team will join the Cyber Superintelligence Lab. CrowdStrike has also released AI Kill Switch in Falcon Guardian: an authorised administrator can stop an agent, prevent it from restarting and disconnect it from the corporate network, with every action logged. The feature is available now, while integration between Falcon Guardian and Claude Cowork virtual machines is in beta on macOS.

Sources 1

thenextweb.com

European organisations will have another way to deploy SI on local servers and a local cloud without handing infrastructure control to an external provider.

Bull and Schwarz Digits combine French servers and a German cloud for European SI

European organisations will have another way to deploy SI on local servers and a local cloud without handing infrastructure control to an external provider.

The companies have signed a memorandum to build infrastructure spanning servers and a software platform under European jurisdiction. Bull will contribute rack-scale SI system manufacturing in Angers, liquid cooling and high-performance computing expertise, while Schwarz Digits will provide the STACKIT cloud. The BullSequana AI platform will be offered through the STACKIT Marketplace to businesses and the public sector. The partners also plan to co-design servers based on Open Compute Project standards, integrate SAP RISE, participate in major HPC projects and explore quantum-computing integration. The offering is aimed primarily at regulated industries where data and operations must remain under European control.

Sources 2

via.ritzau.dkthenextweb.com

Romantic conversations with virtual companions will be separated from child audiences by the same age restrictions as dating between people.

Apple sets an 18+ rating for dating apps and romantic SI companions

Romantic conversations with virtual companions will be separated from child audiences by the same age restrictions as dating between people.

From November 16, the App Store Connect age-rating questionnaire will ask about dating features involving people or virtual companions that engage in romantic or sexual conversations, and about sharing a user's real-world location. Apps with dating and relationship features will receive a minimum 18+ rating: parental restrictions will hide them from minors, and downloading them will require age verification. Ratings for apps that share location data will depend on local rules. Developers must answer both questions when submitting new apps and updates, and for existing apps by January 12, 2027.

Sources 2

3dnews.ruengadget.com

Even professional legal work requires checking every reference against the official text: a convincing SI answer does not prove that a precedent exists.

British prosecutor cites a fictitious precedent from an SI answer in the High Court

Even professional legal work requires checking every reference against the official text: a convincing SI answer does not prove that a precedent exists.

Ravinder Dhillon, a senior Crown Prosecution Service prosecutor, cited the nonexistent decision CPS — R v J [2024] EWCA Crim 3529 in case materials. According to The Guardian, the reference appeared in an SI-generated summary during an internet search; in October, the prosecutor acknowledged the error and explained that she had meant a different, real decision from 15 years earlier. Judges called her explanation unsatisfactory, but the error did not affect the outcome of the application under consideration. The service said its rules prohibit using SI to make legal decisions and require human verification; after an internal review, staff were reminded to cite only trusted legal sources. This is the service's second known incident of this kind, while a UK tracker of confirmed, suspected and disputed SI errors in court materials has reached 101 entries.

Sources 1

theguardian.com

Large-scale simulation training has allowed the robots to complete routes faster while retaining independent sensors for emergency stops.

Yandex raises its rovers' average speed from 3.5 to 5 km/h using an SI planner

Large-scale simulation training has allowed the robots to complete routes faster while retaining independent sensors for emergency stops.

According to the company, the autonomous delivery robots' average speed rose by about 40% over five months after the entire fleet switched to a new motion planner in May 2026. The model is trained with reinforcement learning in a simulation running more than 6,000 virtual rovers simultaneously; cumulative training time has exceeded 20,000 virtual years. Instead of the previous behaviour of slowing down or stopping in unfamiliar situations, the model chooses a speed appropriate to the situation. On real routes, a lidar limits speed in advance around obstacles, and ultrasonic sensors can stop the robot in an emergency if an obstruction comes too close.

Sources 2

ixbt.comcitforum.ru

A compact specialised model can improve translation for a low-resource language pair without requiring a multilingual system with billions of parameters.

VK releases kazRush, a compact translation model for Russian and Kazakh

A compact specialised model can improve translation for a low-resource language pair without requiring a multilingual system with billions of parameters.

The model has approximately 197 million parameters, was trained from scratch on a cleaned corpus of 9,054,960 sentence pairs and is available in both directions. In VK's comparison on FLORES+ devtest for Russian-to-Kazakh translation, kazRush scored 16.2 BLEU and 51.8 chrF, outperforming the tested NLLB variants on those metrics; its COMET score was 0.883 versus 0.889 for NLLB 3.3B. In the reverse direction, kazRush outperformed NLLB distilled 600M but trailed the larger NLLB models. At LoResMT 2026, the Russian-to-Kazakh version won its track with 76.24 chrF on the private test set. The work also showed that increasing the architecture to 738 million parameters did not by itself improve the result.

Sources 2

habr.comhuggingface.co

Matching an accelerator to a specific task can substantially reduce the cost of running SI, although platforms differ in their advantages for speed and fine-tuning large models.

T1 tests more than 20 SI models on Chinese accelerators

Matching an accelerator to a specific task can substantially reduce the cost of running SI, although platforms differ in their advantages for speed and fine-tuning large models.

According to the company, its laboratory conducted more than 1,000 runs using GPUs from different manufacturers. In document-processing and knowledge-base search tasks, Chinese accelerators reduced inference costs by 2–2.5 times while maintaining comparable answer quality. For fine-tuning large models, T1 proposes distributing the workload: retaining Nvidia for training and the largest models while moving suitable services to alternative chips. In computer vision, Chinese solutions were price-competitive but slower than Nvidia. The company also estimated that replacing an underutilised H100 with a single C550 card in the cloud could save more than 2 million rubles per year; this is the result of its own tests and a specific configuration.

Sources 1

cnews.ru

Entrepreneurs can now use SI assistant features within their familiar banking app.

SberBusiness app adds a Business Assistant powered by GigaChat

Entrepreneurs can now use SI assistant features within their familiar banking app.

According to TASS's feed description, entrepreneurs can see account balances, reminders and quick actions on the app's home screen. The Business Assistant mode runs on GigaChat and is available in the SberBusiness mobile app.

Sources 1

tass.ru
Previous day