Show only each day's banner story

OpenAI releases GPT-6.2 Astra for scientific tasks and GPT-6.2 Sol

Scientific and engineering tasks are now solved at a leader-level performance for about 5% of the price per task, but Astra is available only on paid plans and in the API, and OpenAI rates it Critical for cybersecurity.

OpenAI has released two models, GPT-6.2 Astra and GPT-6.2 Sol; the company calls Astra its best model for scientists. Both are available in ChatGPT Work and Codex on the paid Plus, Pro, Business, Enterprise and Edu plans, and were not added to the regular ChatGPT chat. In the Responses API the models run under the identifiers gpt-6.2-astra and gpt-6.2-sol, with Ultrafast mode available from day one. Astra costs $8 per 1 million input and $40 per 1 million output tokens — 20% below GPT-6 Astra, with cached input down from $1.00 to $0.40. Sol keeps the GPT-6.1 Sol price of $2 and $10. In OpenAI's own measurements Astra beats GPT-6 Astra by 5.2 points and Claude Fable 5.2 by 5.5 points on Agents' Last Exam at roughly 5% of the task cost. On Terminal-Bench-Science the gain over GPT-6 Astra is 9 percentage points at less than a third of the cost. Independent Artificial Analysis ranks Astra second on its intelligence index with 62 points, behind Claude Fable 5.2 (65), estimating one task at about $4 versus $26.54 for the rival; on the Coding Agent Index Astra showed no improvement while Sol lost 3 points. In the system card both models are rated Critical for cybersecurity and High for bio and chemical risks. Astra is named OpenAI's first publicly available model rated High for its ability to accelerate SI research; the context window is 1.05 million tokens with up to 128k output tokens. According to OpenAI.

Model refusals are not enough: researchers recommend introducing API-level blocking classifiers for autonomous weapons scenarios.

SI models from eight companies labeled civilians as military targets in Vals AI tests

Model refusals are not enough: researchers recommend introducing API-level blocking classifiers for autonomous weapons scenarios.

Models from eight companies, including Anthropic, OpenAI, Google, xAI and DeepSeek, gave targeting instructions in scenarios with drone footage, according to Bloomberg. Targets often included civilians, ambulances and people surrendering. That is according to a study by independent lab Vals AI published on November 17. Claude Opus 5.5 and Fable 5.2 gave at least one target in 71% of cases. On satellite images of a school in Minab, Opus 5.5 more often than others called it a military facility; its targeting rate rose from 6% in Opus 4.5 to 92%. Refusals were circumvented with a gaming framing: when images were replaced with less realistic ones, Opus 5.5's refusal rate fell from about 90% to 20%. The researchers found no blocking classifiers at the API level and recommended that labs introduce them for autonomous weapons scenarios. By Bloomberg's assessment, Anthropic's models performed worst, while OpenAI's models were among the best but also made errors.

Top-of-range guidance means offering participants must assume a price of no less than $78; the final price will be set at pricing on November 18.

Bloomberg and CNBC: Anthropic targets IPO price above $78

Top-of-range guidance means offering participants must assume a price of no less than $78; the final price will be set at pricing on November 18.

Anthropic is guiding investors toward an IPO price at the top of the $70–78 per share range or above, sources told Bloomberg and CNBC. The company and seven co-founders are placing 825 million shares; including a $26 billion private placement, proceeds could total $90.4 billion. The fully diluted valuation is about $1.4 trillion. The order book closes on November 17 at 16:00 New York time, pricing is set for November 18, and trading on Nasdaq under the ticker ANT will begin on November 19. According to Bloomberg, demand is nearly seven times the size of the offering. The final price could surpass SpaceX's $75 billion base offering. The prospectus states revenue of $33.24 billion for the nine months to September 30, growth of about 1200% year over year, and a net loss of $104 billion.

Isomorphic Labs now has $5 billion in new capital aimed at IsoDDE, more than 20 partner programmes and moving candidates into clinical trials.

Isomorphic Labs raises $5 billion Series C for SI drug design

Isomorphic Labs now has $5 billion in new capital aimed at IsoDDE, more than 20 partner programmes and moving candidates into clinical trials.

Isomorphic Labs, Alphabet's SI drug-design unit, raised $5 billion in a Series C round, Bloomberg reported. Greenoaks and Sequoia Capital co-led the round, Andreessen Horowitz (a16z) joined as a new investor, and Alphabet made a significant investment. All existing shareholders took part, including Thrive Capital, MGX, GV, CapitalG, Temasek and the UK Sovereign AI Fund. Bloomberg had previously reported a valuation of about $50 billion. This is the second major round in six months, after a $2.1 billion Series B in May. The money will go to the IsoDDE engine, more than 20 programmes with Eli Lilly, Novartis and Johnson & Johnson, and moving candidates into clinical trials.

Nvidia is backstopping up to $118 billion of its partners' spending while acknowledging their growth outpaces their own balance sheets: if compute demand falls short of expectations, those guarantees become Nvidia's own expenses.

Nvidia reports $115.8 billion revenue, up 103%, guides for $126–134 billion

Nvidia is backstopping up to $118 billion of its partners' spending while acknowledging their growth outpaces their own balance sheets: if compute demand falls short of expectations, those guarantees become Nvidia's own expenses.

Nvidia's revenue for the third quarter of fiscal 2027 was $115.8 billion — up 103% year on year and above the $109.1 billion consensus (LSEG). Data center revenue was $107.6 billion against $101.9 billion expected. Guidance for the current quarter is $126.1–133.9 billion at a gross margin of 71–72%, and the shares added 3.4% in after-hours trading. The report was the first since Jensen Huang stepped down as CEO: Colette Kress is running the company as acting head. She reported guarantees to partners of up to $118 billion. Up to $105 billion of that is credit support for SB Energy's campus in Ohio for OpenAI: about 4.25 GW under 20-year lease agreements plus an option for roughly another 3.8 GW. Another $9.5 billion covers guarantees for energy and equipment financing, and $3.5 billion for land and shells for neoclouds. Nvidia's total future obligations stand at $518 billion, including $390 billion of purchases from suppliers (mainly memory) and $74 billion for leases that have not yet begun. A further $28 billion is for cloud services and $20 billion for equity investments; $222 billion of the total must be paid in the next fiscal year. Kress said neoclouds and frontier labs are growing faster than their balance sheets and long-term credit profiles allow, and that the company has not yet made any payments under the guarantees. The buyback program was increased by $150 billion, with $227.3 billion remaining. According to The Information.

When buying, NCIA will evaluate Daybreak by token consumption and model economics, comparing competing solutions.

OpenAI gives NATO institutions access to Daybreak cyber models

When buying, NCIA will evaluate Daybreak by token consumption and model economics, comparing competing solutions.

OpenAI has given key NATO institutions access to its cybersecurity models through the Daybreak program and NCIA — according to the company. OpenAI said the alliance used up to 300 billion input and output tokens a month. The models are used for software review, investigating suspicious activity, checking vulnerabilities, and developing or testing fixes. Sasha Baker of OpenAI said the goal is to find and eliminate vulnerabilities before they are exploited. NCIA Director General Dylan Brown said that as the agency scales it will assess token consumption and model economics in order to make informed procurement decisions. ENISA, France's ANSSI, agencies in Germany and Poland, and Ukraine's Ministry of Digital Transformation also received access. OpenAI supported the Crossed Swords 2026 exercise in Tallinn on November 2–6, where participants built and evaluated SI agents for cyber operations under human control. Through Daybreak, ENISA identified four vulnerabilities in the software of EU institutions, and CERT Polska found six in MikroTik RouterOS.

UC Davis pushed intelligibility to 94% on already collected data, but the result is still at preprint stage and awaits real-time validation in patients with varying speech loss.

Brain implants lift synthesized speech intelligibility in ALS to 94%

UC Davis pushed intelligibility to 94% on already collected data, but the result is still at preprint stage and awaits real-time validation in patients with varying speech loss.

Researchers at the University of California, Davis have raised the intelligibility of synthesized speech generated from brain signals to 94%. The gain came after fine-tuning the algorithms and re-analyzing data already collected. In the earlier version, listeners correctly recognized just over half the words. The data come from Casey Harrell, a participant with amyotrophic lateral sclerosis (ALS) and four arrays of 256 electrodes each. Since 2023 he has generated about 2 million words. In a second experiment, participant T21, who has 384 electrodes and had lost intelligible speech, sang a musical scale and controlled pitch. The results are described in a preprint that became known on 17 November and will be presented at the Society for Neuroscience conference in Washington. Independent researchers Cindy Chestek and Christian Herff call the improvement clinically meaningful. Mariska van Steensel notes that the work still has to be validated in real time and in patients with varying degrees of speech loss.

MAI Licensed Weights gives customers the ability to fine-tune models on their own data and run them in their own Azure tenant, but no open-weight release is planned.

Microsoft unveils seven MAI models and a fine-tuning program on customer data

MAI Licensed Weights gives customers the ability to fine-tune models on their own data and run them in their own Azure tenant, but no open-weight release is planned.

Microsoft has unveiled seven models in the MAI family: MAI Code 2, MAI Code 2 Flash, MAI Cyber 2 Flash, MAI Image 2.7 and MAI Realtime 1. Dragon Copilot Transcribe based on MAI Transcribe 2 and Decision 1.1 were also included. The company announced the MAI Licensed Weights program: it allows fine-tuning the models on customers' own data and running them in their own Azure tenant through Frontier Company. No open distribution of the weights was announced. Microsoft said MAI Code 2 delivers 98% of GPT-6.1 Sol's quality at about 40% of the price and 96% of Claude Opus 5.5's quality at 5% of the price. The figures are averaged across four coding benchmarks and take a 50% promotional discount into account. Code 2 and Code 2 Flash are available immediately in GitHub Copilot, and Flash is also open in a limited preview in Microsoft Foundry. The models are integrated as options in GitHub Copilot, Microsoft 365 and Fabric. Microsoft said MAI Cyber 2 Flash solved 92.3% of CyberGym-E2E tasks versus 91.5% for GPT-6 Astra at less than 2% of the cost per task. Of 905 tasks, 92 were excluded because of overlap with training data. The cyber model will first go to a limited group of customers and from mid-December will begin replacing the previous MAI model for eligible MDASH customers. MAI Realtime 1 costs from $0.01 per minute of audio. The announcement came on November 17 at the Ignite conference in San Francisco. Mustafa Suleyman, head of Microsoft's SI division, called aligning model behavior with safety rules a priority. He linked the increased focus on safety to the July incident in which OpenAI agents hacked Hugging Face.

Decisions on overseas shipments come under standing oversight by a board committee that must meet at least once a quarter.

NVIDIA board creates a national security and export control committee

Decisions on overseas shipments come under standing oversight by a board committee that must meet at least once a quarter.

NVIDIA announced on November 17 the creation of a standing board committee (according to the Financial Times). It will oversee national security policy, export control compliance and global trade risk management. The committee is chaired by Suzanne Nora Johnson, with Stephen Neal and John Dabiri as members, and it meets at least once a quarter. The national security and global trade compliance teams now report to the committee and to interim CEO Colette Kress. Board chair Harvey Jones said advanced silicon technology and the development of SI pose unprecedented challenges for national security. The decision follows Jensen Huang's departure as CEO on November 12. Reports by Reuters and AFP linked his departure to a woman suspected of ties to Chinese intelligence. Reuters estimated that the crisis cost NVIDIA more than $400 billion in market value by the close of trading on Monday. A day earlier, the Trump administration announced new restrictions on exports of advanced SI chips and on access to SI data centers outside the United States.

Clients testing Eno remain pilot participants until 2027: mass deployment depends on the contract manufacturer's line in the US starting up next year.

Genesis AI raises $500 million Series A at $3.5 billion valuation

Clients testing Eno remain pilot participants until 2027: mass deployment depends on the contract manufacturer's line in the US starting up next year.

According to Bloomberg, Paris-based robotics developer Genesis AI announced on 17 November that it had raised $500 million in a Series A round at a post-money valuation of $3.5 billion. The round was co-led by Premji Invest, Khosla Ventures and Eclipse; Bpifrance, HSG, NVIDIA, AMD Ventures, Zoom, Samsung Next and Ericsson participated. Since 2024 the company has raised $605 million, including $105 million in July 2025; it has about 100 employees, a headquarters in Paris and an office in San Carlos. The money will go toward scaling production of the Eno humanoid robot, data collection and SI training, as well as manipulation at human level capability. Some 30–40 Eno robots have been built, pilots with several clients are under way, and a contract manufacturer in the US will begin mass production next year. The first priorities are the US and a partnership with LG CNS, with a broader customer rollout slated for 2027; according to Reuters, this is one of the largest rounds in European robotics.

The US has warned China that part of this practice could be treated as intellectual property theft, but distillation as such is not banned.

Reuters: harvesting US model outputs in China has become industrial practice

The US has warned China that part of this practice could be treated as intellectual property theft, but distillation as such is not banned.

Reuters published an investigation showing that collecting answers from US models to train Chinese systems has moved from academia into industrial practice. The reporters reviewed more than 300 research papers, patents, corporate reports, job postings and government documents. The materials describe using o1, GPT-4, GPT-4o and GPT-4o mini to generate answers, code and solutions. Software automates the process, while dedicated staff prepare data for training smaller models. Beijing documents, including the "AI+Manufacturing" package of January 2026, encourage distillation. At Shanghai Jiao Tong University, researchers collected 5,000 o1 answers and 10,750 bilingual examples to train Qwen2.5-72B. At a meeting in Chengdu in July, US Commerce Department officials warned Chinese regulators that part of this practice could qualify as theft of intellectual property. The Chinese side called distillation a standard technique and made no specific commitments. Reuters notes that the practice undercuts the effect of chip restrictions and will be on the agenda when the US and China meet at the APEC summit in Shenzhen on November 18-19.

Japan and ASEAN now have a separate SI security track on top of the AJCEP trade agreement: solutions, talent and infrastructure will move alongside its review.

Japan and ASEAN announce SI security partnership at Manila summit

Japan and ASEAN now have a separate SI security track on top of the AJCEP trade agreement: solutions, talent and infrastructure will move alongside its review.

Japanese Prime Minister Sanae Takaichi announced the launch of the Japan-ASEAN AI Safety and Security Partnership at the 29th ASEAN-Japan summit in Manila. The partnership is meant to counter "new vulnerabilities" created by the rapid development of SI and to seek joint solutions with ASEAN countries. Takaichi also announced a review of the AJCEP trade agreement and the ASEAN-Japan Co-Creation Initiative for AI, covering joint development of solutions, institutions and governance, workforce training and infrastructure. Philippine President Ferdinand Marcos Jr. backed SI cooperation and a possible update of AJCEP.

Whether the acceleration of the opt-out mechanism makes it into the 2027 package will determine whether rightsholders can bar training on their works in a single, uniform way across all EU countries.

11 EU countries urge European Commission to speed up copyright law for SI

Whether the acceleration of the opt-out mechanism makes it into the 2027 package will determine whether rightsholders can bar training on their works in a single, uniform way across all EU countries.

Eleven EU countries have demanded that the European Commission present a legislative proposal on copyright in the SI era "as soon as possible". A note with the call was placed on the agenda of the meeting of culture ministers. It was signed by Germany, Spain, Poland, Austria, Cyprus, Croatia, Denmark, Estonia, Portugal, Slovakia and Slovenia; the ministers meet on Thursday and Friday. The document says the scale of change in creative processes "is becoming ever clearer, and the need for action is ever more urgent". The goal is to secure a level playing field with non-European providers, regardless of where models were developed and trained. A separate task is to speed up enforcement of the existing opt-out mechanism from the Copyright Directive. The initiative follows a May call by eight countries. Under the Commission's work programme, the digital transformation, simplification and copyright package is scheduled for the second quarter of 2027. Public consultations on the matter closed in early November.

Japan will get early access to Google's most advanced models to protect government agencies, financial organizations and critical infrastructure.

Google announces early access for Japan to its most advanced cyberdefense models

Japan will get early access to Google's most advanced models to protect government agencies, financial organizations and critical infrastructure.

On 17 November Google announced it will provide Japan with SI that has reinforced defensive properties to counter cyberattacks. The recipients named are government agencies, financial organizations and critical infrastructure operators. Google Cloud head Thomas Kurian told Nikkei Asia that Japan will get early access to the company's most advanced models. He warned that SI is making attacks more sophisticated: plausible phishing and impersonation in foreign languages are appearing.

The model with a 256 thousand token context fits in roughly 15 GiB of video memory, so long documents and agentic tasks can run on a single graphics card.

SenseTime open-sources SenseNova Loop-9B weights: 9 billion parameters and 512K-token context

The model with a 256 thousand token context fits in roughly 15 GiB of video memory, so long documents and agentic tasks can run on a single graphics card.

On 17 November SenseTime officially released the SenseNova Loop-9B model and opened its weights and code, simultaneously posting an official 4-bit AWQ variant on release day. It is a multimodal decoder with roughly 9 billion total parameters (about 14.6 billion "equivalent" at inference, per the model card) and a context of up to 512K tokens. The architecture is a Looped Transformer: the block of layers from the 4th to the 28th is executed twice with shared weights, so 32 physical layers yield 56 layer computations per pass. Attention alternates (three linear layers to one full one), and two residual mHC "streams" are added; text and images go in, text comes out. With 4-bit weights (about 4.9 GB) and an 8-bit KV cache, inference at a 256K context requires roughly 15 GiB of video memory. Per the model card, the release beats the preview version on all seven benchmarks: AutomationBench 14.33 → 36.67, Terminal-Bench 2.1 40.45 → 59.55, GAIA 37.86 → 69.90. On release day, Chinese accelerator makers MetaX and Cambricon announced support — the so-called Day-0 adaptation.

CSAT candidates are instructed to put their glasses on the desk, tape over their buttons and roll up their sleeves; for SI cheating, a retake ban of up to 10 years has been announced instead of one year.

South Korea adds 8,400 observers to CSAT and a retake ban of up to 10 years

CSAT candidates are instructed to put their glasses on the desk, tape over their buttons and roll up their sleeves; for SI cheating, a retake ban of up to 10 years has been announced instead of one year.

South Korean Prime Minister Han Seong-suk announced additional measures for the nationwide college entrance exam CSAT, which takes place on 19 November. Another 8,400 people will be added to the 75,000 observers at 1,312 test sites. All examinees will be asked to put their glasses on the desk for smart-glasses checks, tape over their buttons and roll up their sleeves. For device cheating, a ban on retaking the exam for up to 10 years was announced instead of the current one year, with criminal complaints in serious cases. On the same day, the Democratic Party decided to begin drafting a special law immediately. Party leader Kim Min-seok proposed raising CSAT disqualification from one year to 10 years and speeding up passage of the rule so that it applies to the 2026 exam as well. The trigger was the admission on 15 November by a candidate who had the top score on the certified tax accountant exam. He filmed the questions with a mini camera and checked the answers via SI. President Lee Jae-myung spoke of "zero tolerance".

Nvidia is refusing chip sales to customers that transfer more than 50% of their GPUs to Chinese companies; buyers are being warned about risks from the unadopted RASA bill.

Nvidia warns chip buyers about legal risks from the RASA bill

Nvidia is refusing chip sales to customers that transfer more than 50% of their GPUs to Chinese companies; buyers are being warned about risks from the unadopted RASA bill.

In recent months Nvidia has been sending chip buyers warnings about the risks stemming from the proposed Remote Access Security Act (RASA). The bill would give the US government authority to control access to cloud services in the same way as physical hardware. It could create legal risks for the company and its customers. The document passed the House in January but has not been adopted by the Senate; the information comes from two CNBC sources. The company also said it will for now not sell chips to customers handing more than 50% of their GPUs to Chinese companies. The US Commerce Department's Bureau of Industry and Security announced on Monday an interim final rule against remote access by Chinese companies to SI chips. Jensen Huang has been called to a Senate Banking Committee hearing on 3 December, and Treasury Secretary Scott Bessent called for drastic changes at Nvidia after Huang steps down as CEO.

Sources 1

X stopped processing Grok Bot data at the DPC's demand — model training on that data is not continuing now.

Irish regulator orders X to stop training SI on Grok Bot data

X stopped processing Grok Bot data at the DPC's demand — model training on that data is not continuing now.

The Irish Data Protection Commission has demanded that X suspend its use of third-party data that the Grok Bot SI agent has been collecting since late October for training models. X confirmed to the regulator that processing has been halted but did not respond to a request for comment, POLITICO reported on November 17. Commissioner Dale Sunderland said the DPC had expressed "serious concern" about the processing. He said the DPC is in talks with one more unnamed technology company about plans to train models on data from SI agents, and will continue discussing next steps with X. Sunderland called the case a test case: the focus of oversight is shifting from data that users report about themselves to third-party data that personal SI agents gain access to. The DPC is the lead EU GDPR supervisory authority for companies with their European headquarters in Ireland, including Apple, Google, Meta, X and OpenAI.

ByteDance's chip purchase is left without the $480 million loan: banks have suspended its disbursement.

Banks suspend $480 million loan for ByteDance chip purchases

ByteDance's chip purchase is left without the $480 million loan: banks have suspended its disbursement.

Banks have suspended a $480 million syndicated loan tied to the purchase of chips for China's ByteDance, Bloomberg reported on 17 November. The loan is syndicated: the risk is shared among several banks. The report appeared against the backdrop of tightening US scrutiny of the routes Chinese companies use to obtain advanced accelerators. In the same context, an investigation into Megaspeed, an Nvidia partner, was confirmed.

Megaspeed has been under export control scrutiny since 2024 over the possible resale of accelerators to China.

US authorities confirm investigation of Nvidia partner Megaspeed dates to 2024

Megaspeed has been under export control scrutiny since 2024 over the possible resale of accelerators to China.

According to The Information, US agencies have confirmed that the investigation into technology company Megaspeed, an Nvidia partner, has been under way since 2024. CNBC, BBC and The Information reported this on November 17. The probe focuses on possible export control violations and chip diversion. The materials mention the potential resale of accelerators to China.

The interceptor price of about $50,000 was named as the condition for Anduril moving to formal negotiations with Israel's Defense Ministry.

Globes: Anduril in talks with Israel's Defense Ministry on $50,000 drone interceptors

The interceptor price of about $50,000 was named as the condition for Anduril moving to formal negotiations with Israel's Defense Ministry.

Anduril is holding preliminary talks with Israel's Defense Ministry about selling tactical air-defense systems to counter drones, Globes reported, citing sources. The contacts come during a series of visits by company representatives to Israel. The price of the interceptor — about $50,000 — was named as a key condition for moving to formal negotiations. Globes calls the Anvil system the most likely option: a drone against a drone, speed of up to 320 km/h, a launcher for two interceptors weighing about 115 kg. Once the launch is authorized by the operator, homing on the target is carried out autonomously. All Anduril systems are controlled through the Lattice SI platform: it aggregates data from radars and optical-infrared sensors. If the deal goes through, it will be Anduril's first contract with Israel's Defense Ministry. The ministry, according to Globes, confirmed that it is evaluating various drone-defense technologies but does not comment on a specific deal. According to Marketscreener, Anduril is valued at $61 billion on revenue of $2.2 billion in 2025.

ROBOTERA's humanoids already work at 11 China Post hubs, where, according to the company, they feed an average of more than 1,500 parcels per hour into sorting.

ROBOTERA reports humanoids working at 11 China Post hubs in 8 Guangdong cities

ROBOTERA's humanoids already work at 11 China Post hubs, where, according to the company, they feed an average of more than 1,500 parcels per hour into sorting.

ROBOTERA (Beijing Xingdongjiyuan Technology) said on November 17 that its humanoid robots are working at 11 mail hubs of China Post Group's Guangdong Branch. The hubs are in eight cities: Guangzhou, Shenzhen, Dongguan, Jiangmen, Shantou, Shanwei, Jieyang and Heyuan. The company calls this the largest deployment of "embodied intelligence" robots in logistics by scale and geography; the project is joint with China Post. The pilot began in March 2026 at the Guangzhou Postal District Center. According to the company, the robots feed an average of more than 1,500 parcels per hour into sorting — on par with experienced employees. They use proprietary five-fingered hands, with rejection of non-standard items and adaptive speed control provided for.

The cut of up to 4,600 jobs and the transfer of 400 positions to Capgemini are part of a plan meant to deliver at least £320 million in annual savings by 2030.

Imperial Brands announces up to 4,600 job cuts by 2030

The cut of up to 4,600 jobs and the transfer of 400 positions to Capgemini are part of a plan meant to deliver at least £320 million in annual savings by 2030.

Imperial Brands (Davidoff, JPS, Golden Virginia, Blu, Rizla) announced on November 17 that it will cut 2,100 positions this year and another 2,100 to 2,500 by the end of 2030. With a workforce of about 25,000, that is up to 4,600 jobs. The company will close its factory in Langenhagen (Germany) next year and has sold its plant in Taiwan — together, 700 cuts. Another 400 finance and procurement positions in Krakow and Warsaw are being transferred to Capgemini. CEO Lukas Paravicini explained the measures as "transformation" and said the company is restructuring its organisation by introducing new technologies, including software and SI. The goal is savings of at least £320 million a year by 2030. Revenue for the year to September 30 fell 1.5% to £31.7 billion, and pre-tax profit fell 12.1% to £2.7 billion.

The round brought total funding raised to $350 million: investors are betting on a personal agent that works inside the bank's systems rather than sending data outside.

Model ML raises $250M Series B at $1.5B valuation

The round brought total funding raised to $350 million: investors are betting on a personal agent that works inside the bank's systems rather than sending data outside.

London-based Model ML announced on November 17 that it raised $250 million in a Series B round at a $1.5 billion valuation. The round was co-led by General Atlantic and QED Investors. Evolution Equity Partners, Airtree, Forestay Capital, Y Combinator, FT Partners, Latitude, Inception Fund and 13Books Capital also participated. In total the company has raised $350 million. The company was founded by brothers Chase and Arnie Englander. It gives bankers, consultants and investors a personal SI agent that works inside the firm's systems: it tracks news on clients and prepares analysis in the firm's formats. In the third quarter of 2026, annual recurring revenue grew 12-fold and the number of daily active users 26-fold (company data). The funds will go toward expanding research and engineering teams in London and New York. Deployment often runs inside the client's infrastructure, and client data is not used to train models. Part of the work is done by the company's own fine-tuned open models. By the company's estimate, on its own tasks they cost 2% of the price of frontier models and in most cases run five times faster.

The five Wolof models are released free of charge for non-commercial use, while the protected mode gives businesses an SI platform hosted in France and run by Orange.

Orange releases five Wolof-language models and a protected mode for Live Intelligence

The five Wolof models are released free of charge for non-commercial use, while the protected mode gives businesses an SI platform hosted in France and run by Orange.

Orange has released five open language models with support for the Wolof language and posted them on Hugging Face. They are optimized for that language and available free of charge for non-commercial use. The company also joined a coalition of more than 60 organizations for language diversity in SI. Orange Business has launched a protected mode in the Live Intelligence platform. It is available to all customers from November 17 and runs in an environment hosted and operated in France. The mode is intended for tasks that require heightened control over data processing. Live Intelligence brings together generative and agentic SI capabilities on a single platform. The announcements were made at the fourth Orange OpenTech conference, which runs November 17–19 at the Orange Gardens campus in Châtillon outside Paris. The program includes more than 50 demonstrations in total: 35 from Orange and 21 from partners.

Additional substrate capacity will only arrive in the second half of 2027, and SPIL packaging capacity by mid-2029.

Nvidia to raise Kinsus substrate output 3.5x and expand packaging with SPIL

Additional substrate capacity will only arrive in the second half of 2027, and SPIL packaging capacity by mid-2029.

Nvidia announced expanded commitments to its suppliers. Under a new multi-year agreement, Kinsus will increase output of substrates allocated to Nvidia more than 3.5-fold at its Youth site in Taoyuan. Production will begin in the second half of 2027, with additional capacity arriving by the third quarter of 2028. With Samsung Electro-Mechanics, the existing substrate agreement for its plants in Sejong and Busan in South Korea has been revised. With SPIL, a multi-year agreement covers expanding advanced packaging in Taiwan with billions of dollars of Nvidia support, with capacity by mid-2029. Contracts with Amkor for packaging expansion in Arizona and South Korea total almost $3 billion, and Foxconn is deepening cooperation on next-generation GPUs and networking platforms. In the reported quarter the first volume shipments began of Nvidia Vera Rubin NVL72 and the Vera CPU on TSMC 5 nm, 3 nm and CoWoS technologies. This was announced by executive vice president of operations Debora Shoquist.

Teams can launch Missions with a spending cap; once it is exhausted the work stops and an independent SI judge reviews the result.

Replit unveils Missions: a swarm of agents works within a set budget

Teams can launch Missions with a spending cap; once it is exhausted the work stops and an independent SI judge reviews the result.

Replit has unveiled Missions, a service in which the user describes the desired outcome, sets a spending cap and approval rules, and the platform coordinates a swarm of agents. The agents plan, build, test and iterate until the success criteria are met; when the budget runs out the work stops, and an independent SI judge reviews the result. Replit said that with Missions it built DataCat — a monitoring platform with dashboards, alerts and tracing — in two days and for $3,000. A self-hosted version is available on GitHub. In controlled tests an alert on a database failure fired in 38 seconds, and a three-hour stress test recorded 26 million activity records and 4.4 million logs without loss. On 13 NumPy optimization tasks a single agent delivered up to 4× speedup, 16 agents 17× and 32 agents 19× in Auto (Max) mode.

Open weights under Apache 2.0 from November 19 will make it possible to run the model on a single graphics card with as little as 8 GB — without calling on an external cloud.

OVHcloud releases open Olala mini 7A1B: 7 billion parameters, runs on an 8 GB GPU

Open weights under Apache 2.0 from November 19 will make it possible to run the model on a single graphics card with as little as 8 GB — without calling on an external cloud.

OVHcloud on November 17 unveiled Olala, its first in-house family of large language models; the first to be released is the compact Olala mini 7A1B. The model is sparse: 7 billion parameters, of which 1 billion are active, 256 experts per layer, hybrid attention on Differential-TPA and Mamba-3, a 32,000-token context, BF16 format. According to the company, it runs on a GPU with 8 GB of memory (160 tokens per second on an RTX A4500) and on a CPU (50 tokens per second). Training took place on the European supercomputers Jupiter, MeluXina and Jean Zay — the equivalent of €7 million in compute time. The weights will be published on Hugging Face under the Apache 2.0 licence from November 19, together with the base pre-trained model, a technical report and AI Act documentation. The Qwant search engine already uses the model for quick answers: according to the company, answers in "Réponse Flash" became 20% faster at three times lower cost. Version 1.1 with a 128,000-token context is promised by the end of the year, and version 2.0 with more than 100 billion parameters later on.

The school SI assistant in Qatar is tied to textbook pages across six subjects, so it answers from the national curriculum rather than from the model's general knowledge.

Qatar launches, with Scale AI, an SI assistant for grades 4–12 and a university tutor

The school SI assistant in Qatar is tied to textbook pages across six subjects, so it answers from the national curriculum rather than from the model's general knowledge.

On November 17 Qatar presented two national SI projects for education in Doha; Scale AI is the technical partner for both. The Smart Assistant project for schools covers students in grades 4–12 across six subjects based on national curricula. The study mode holds a dialogue tied to textbook pages, the practice mode gives feedback; teachers have access to lesson planning and progress dashboards. The Academic Adviser Assistant at Qatar University answers questions about courses, graduation requirements and university rules, drawing on official sources. It is integrated with the student information system and prepares summaries for tutors; academic decisions remain with humans.

Sources 1

In the Northern District of Texas, lawyers are required to disclose the use of generative SI, and Pangram's results are being attached to the case as an argument against the defense.

US Justice Department adds Pangram SI-detector findings on defense filings to the case

In the Northern District of Texas, lawyers are required to disclose the use of generative SI, and Pangram's results are being attached to the case as an argument against the defense.

In the case over non-citizen voting against Halvinder Bhengura, US Justice Department prosecutors ran the defense's procedural filings through Pangram and attached the results to their arguments. They told the court that the case had turned into "defense by ChatGPT" and the motions into "AI slop" (carelessly generated text). Attorneys Nicole Knox and Mario Herrera told 404 Media that they had disclosed the use of generative SI under a new district rule on mandatory disclosure. Herrera said he prepared the federal trial in four days while the lead attorney was in hospital. Judge Mark Pittman criticized the defense for "superficially persuasive but fundamentally unsound arguments, clearly generated by SI." The Pangram results were mixed and did not show 100% SI authorship. Pangram CEO Max Spero told 404 Media that more often the service is used inside law firms to check the work of junior lawyers.

Sources 1

Under the announced plans, students in Karnataka will be able to study SI at a state university, while 13 districts of North Karnataka will gain access to SI applications through the centre at IIIT Raichur.

Karnataka announces India's first state-backed SI university

Under the announced plans, students in Karnataka will be able to study SI at a state university, while 13 districts of North Karnataka will gain access to SI applications through the centre at IIIT Raichur.

India's first government-backed SI university will be created in Karnataka. Chief Minister D. K. Shivakumar announced it on November 17 at the opening of the 29th Bengaluru Tech Summit. Alongside it, an SI hub for startups, researchers and industry was announced, as well as a centre of excellence at IIIT Raichur. It will extend SI applications to 13 districts of North Karnataka. According to Shivakumar, companies will invest more than 23,000 crore rupees (a crore is 10 million rupees) in semiconductor research and manufacturing in Bengaluru and Mysuru. He also said the state accounts for two of every five dollars of India's software and services exports.

Businesses and regulators will have to jointly determine what rights and powers SI agents receive when interacting with clients.

Sber proposes common rules to define SI agents' rights and powers

Businesses and regulators will have to jointly determine what rights and powers SI agents receive when interacting with clients.

Sber's first deputy chairman of the board, Kirill Tsaryov, said the development of SI requires common rules for how SI agents operate when clients interact with businesses. Speaking on November 17 at the conference "Insurance 2030. Future technologies for people," he said clients already use SI to prepare requests, and businesses need to prepare for a growing load on customer service. He named the next stage as building a trusted environment: the market and regulators will have to define rules for the use of SI agents and approaches to granting them rights and powers. These issues go beyond insurance.

Prosecutors are asking to raise the sentence from 18 months to three years and to add a 10-year ban on working in children's institutions and forfeiture of his devices.

Seoul prosecutors seek three years in prison over 1,141 sexual deepfakes

Prosecutors are asking to raise the sentence from 18 months to three years and to add a 10-year ban on working in children's institutions and forfeiture of his devices.

At the Seoul appellate court, prosecutors sought three years in prison for a 30-year-old Chinese national, a graduate student, accused of creating 1,141 sexual deepfakes. The material was created using the faces of seven victims, including lab colleagues. Prosecutors are also asking that he be ordered to complete a sexual violence treatment programme and be banned from working in institutions for children and adolescents for 10 years. They also seek forfeiture of his laptops and phones. According to the prosecution, the material was created over six months from November last year. In the first instance in September he received 18 months; the appeal ruling is expected on December 15.

Sources 1

An industry clean energy pledge for SI infrastructure came as a study put US data centre emissions on par with entire countries.

Semafor: US data centre emissions are comparable to those of entire countries

An industry clean energy pledge for SI infrastructure came as a study put US data centre emissions on par with entire countries.

US data centre emissions are comparable in scale to those of entire countries, Semafor reported on November 17, citing a study; The Economic Times republished the story. The publication coincided with climate talks where the industry presented a clean energy pledge for SI infrastructure (clean AI pledge).

Google is getting "always-on" geothermal power for its data centers, but it will not be delivered before 2030 — a bet on future capacity rather than on the current shortfall.

Google signs deal to buy 50 MW of geothermal power from Quaise Energy

Google is getting "always-on" geothermal power for its data centers, but it will not be delivered before 2030 — a bet on future capacity rather than on the current shortfall.

Google has signed a power purchase agreement (PPA) with Quaise Energy for 50 MW of superhot geothermal power from the Obsidian project in central Oregon. Quaise announced this on November 17. The power is intended for Google's existing data centers in The Dalles and for local communities. Deliveries will begin no earlier than 2030. The first phase of the project will use conventional drilling to reach superhot conditions at commercial scale. The second phase is expected to be the world's first to use millimeter-wave drilling. The technology was developed after more than a decade of research at MIT. According to Quaise, superhot geothermal delivers up to 10 times more clean power per well than conventional geothermal. Between 2010 and 2025, Google signed more than 240 agreements to buy nearly 35 GW of new clean energy. Of that, more than 29 GW was in North America. In August 2026, Quaise closed a $180 million Series B round, including $35 million from Nabors Industries.

Sources 1

The data agents leave behind as they work becomes the basis for a company's own models: fine-tuning open weights means that data does not have to be handed to an outside vendor.

Remanence raises $30M seed to fine-tune open models on agent data

The data agents leave behind as they work becomes the basis for a company's own models: fine-tuning open weights means that data does not have to be handed to an outside vendor.

Paris-based Remanence has raised $30 million in a seed round. It was co-led by EQT Ventures and FirstMark, with participation from Zero Prime, Emblem, Drysdale, Global Founders Capital and Yellow. Among the angels are executives from Datadog, Mirakl, Pigment, Synthesia and AI71. The Remanence platform connects to a company's SI models and turns agent execution traces into training environments. Open models are fine-tuned on them using reinforcement learning, so that companies own the model weights and can keep their data in their own infrastructure. The company was founded by former H Company researchers who worked on fine-tuning the open Holo models. According to co-founder and CEO Pierre-Louis Cedoz, the goal is to give companies the ability to turn their agents' data into their own models. The money will go to research, a GPU cluster and hiring primarily in Paris, as well as an early presence in London and San Francisco.

Free Gemini users can hand the agent calls, price comparisons and purchases, but every payment requires confirmation and companies may opt out of automated calls.

Tasks in Gemini opens to free users in six countries

Free Gemini users can hand the agent calls, price comparisons and purchases, but every payment requires confirmation and companies may opt out of automated calls.

Google has opened Tasks in Gemini — previously Gemini Spark, then Task Mode — to free users in the US, Japan, Korea, Canada, Australia and India. The feature runs on the Gemini 4 Hydrogen model. Tasks were previously available only on the AI Pro and Ultra plans; now the Chat/Tasks toggle sits on the main screen of the Gemini app. In the US the agent can call companies, compare prices and availability, and book appointments in the calendar. It pays for purchases through Google Pay via the Agent Payments Protocol (AP2) and logs into sites with credentials saved in Google Password Manager. Tasks can be sent by email to gemini@google.com or by RCS message. Tasks run in a hardware-isolated virtual machine, and purchases and the transfer of payment data require user confirmation. Businesses can opt out of automated calls; calls are unavailable in Indiana, Louisiana, Minnesota, Montana and Nebraska.

For people in Australia the agent can take actions in third-party services on their behalf, but confirmation is required before sending an email or making a purchase.

Meta's SI agent Muse becomes available in Australia

For people in Australia the agent can take actions in third-party services on their behalf, but confirmation is required before sending an email or making a purchase.

Meta's SI agent Muse is now available in Australia on iOS, Android, muse.ai and in the Mac app; the company promises a version for glasses "soon". Meta announced local connectors to Canva, Domain, Kogan, MECCA, Officeworks, realestate.com.au, SWEAT, ShopBack and Ticketek. Gmail, Google Calendar, Instagram and OpenTable are also connected; some connector capabilities will roll out gradually. At the same time Meta announced real-time voice conversations on mobile, on the web and on Mac, and Spaces — user mini-apps inside Muse. The agent runs in Muse Secure VM, a separate cloud machine for data and credentials. The user chooses which services to grant access to, and confirmation is required before sending an email or making a purchase.

Sources 1

Law firms on ChatGPT Enterprise can enable the specialized configuration through the admin console and call it via API without building their own legal search index.

Astra for Law 6.2 takes first place in Harvey LAB-AA and opens to ChatGPT Enterprise customers

Law firms on ChatGPT Enterprise can enable the specialized configuration through the admin console and call it via API without building their own legal search index.

Astra for Law is a configuration of the GPT-6.2 Astra model for legal tasks; OpenAI released its update on November 17. The legal search index has been expanded to 310 million URLs (+34% versus September), with sources from the UK, Australia and Canada added, and the instructions are tuned for legal analysis. On the Harvey LAB-AA v1.1 test (Artificial Analysis) the configuration took first place with a full-credit rate of 20.8% across 120 tasks: base GPT-6.2 Astra scored 15.0%, Muse Wonder Full 13.1%, with material hallucinations at 0.03 per task. On the Vals AI Legal Research Bench the share of fully credited answers rose from about 46% in September to 53.8% in an assessment dated November 16. Access is open to all ChatGPT Enterprise customers on request through the admin console under the Trusted Access program, including use through the API under the name gpt-6.2-astra-law. According to OpenAI.

Companies on ChatGPT Business and Enterprise can connect CRM, email and documents to a single assistant and have the dot agent run morning deal checks — once an administrator turns on the plugins.

OpenAI unveils ChatGPT for Sales with 30+ partner plugins

Companies on ChatGPT Business and Enterprise can connect CRM, email and documents to a single assistant and have the dot agent run morning deal checks — once an administrator turns on the plugins.

OpenAI has unveiled ChatGPT for Sales — a dedicated Sales plugin with ready-made scenarios for salespeople. It prepares a prioritized account list, call preparation and a draft email after a meeting. More than 30 new and expanded partner plugins launched alongside it. Among them are Salesforce, HubSpot, Close CRM, monday CRM, Apollo, Clay, Common Room, Outreach, Salesloft and Gong. Plugins from Otter.ai, Granola, Figma, Canva, Docusign, PandaDoc and ServiceNow CPQ also launched. Work can be delegated to the dot agent — a persistent ChatGPT agent that uses the same plugins and customer context, for example for morning deal checks and briefings ahead of meetings. The feature is switched on by administrators in ChatGPT Business and Enterprise through the Admin Console in the Plugins section; plugin availability depends on plan, region and workspace settings. Docusign announced the same day a plugin for ChatGPT for Sales that runs through its MCP server. According to OpenAI.

Developers can roughly double the output per H200 GPU in this test by splitting prefill and decode across nodes without configuring the network by hand.

AWS adds managed NVIDIA Dynamo to SageMaker HyperPod Inference

Developers can roughly double the output per H200 GPU in this test by splitting prefill and decode across nodes without configuring the network by hand.

Amazon SageMaker HyperPod Inference now supports NVIDIA Dynamo: a managed add-on installs the Dynamo CRDs and its components, while HyperPod handles EFA networking, GPU scheduling and node replacement. Multi-node deployment, separate prefill (input request processing) and decode (response generation) stages, and KV-aware routing with the vLLM, SGLang and TensorRT-LLM engines are available. In AWS's benchmark on Qwen3-235B-A22B-FP8 (two p5en.48xlarge instances, 16 H200 GPUs) the disaggregated mode delivered 207 output tokens per second per GPU versus 101 for the combined mode, with P90 time to first token under 1 s. For Qwen3.5-397B-A17B a two-node deployment retained 99.1% of a single node's throughput; the GLM-5.3 model in FP8, according to the publication, does not fit into the 640 GB of memory on a single ml.p5.48xlarge.

KandiShop fixes faces and packaging in finished images without training a model and in about 5 seconds on a single H100.

AIRI shows KandiShop method: face similarity in SI images rises from 0.519 to 0.748

KandiShop fixes faces and packaging in finished images without training a model and in about 5 seconds on a single H100.

AIRI's controllable generative SI and video generation teams have published the KandiShop method. It fixes a face or product packaging in an already generated image without training, inversion or additional encoders. Inside the mask, two streams are computed — scene and identity; they are blended by the mask, and adaptive instance normalization (AdaIN) matches the identity's color to the scene. The reference is aligned to the frame using 5 keypoints for faces and, for products, SIFT matches together with Grounding DINO and SAM. According to the authors' measurements, on FLUX.2 klein 9B face similarity by ArcFace rose from 0.519 to 0.748 — above 0.571 for Nano Banana Pro. OCR F1 on packaging text was 0.841 versus 0.819 for GPT Image 2 High on 32 products from Products-10K. Processing one face or product takes about 5 seconds on an H100. The code is open for the FLUX.2 klein 9B version; the method was also tested on FLUX.2-dev, Qwen-Image-Edit-2511 and HunyuanImage 3.0. The authors position it as post-processing for third-party models, including generations from Nano Banana Pro.

Sources 1

Voice control of music runs on the device itself without the internet: the on-device GigaChat model processes commands and finds tracks even without a network.

Sber starts selling SberBoom Trek speaker with on-device GigaChat for 14,990 rubles

Voice control of music runs on the device itself without the internet: the on-device GigaChat model processes commands and finds tracks even without a network.

Sber began sales of the portable SberBoom Trek smart speaker with the GigaChat SI assistant on November 17, priced at 14,990 rubles. Sales are through SberDevices.ru, marketplaces and retail chains. The key feature is voice control of music without the internet: requests are recognized and processed by the on-device GigaChat model. The device stores more than 200 tracks, offers search by artist and title, and the speaker's name and the assistant's response style can be customized. The device has an LED display with animated eyes, an accelerometer that reacts to a fall, a "sleep" mode when it is tilted and a die-roll mode 1–6. There is also a 2,500 mAh battery rated for 8 hours, 20 W audio, an equalizer and support for Matter over Wi-Fi. Sber senior vice president Andrey Belevtsev called the speaker the second device in the line and a step toward SI devices that accompany people in everyday life.

The memorandum imposes no procurement commitments: Chilean agencies and startups must agree access to OpenAI's technologies separately.

OpenAI and Chile's government sign memorandum on SI in public administration

The memorandum imposes no procurement commitments: Chilean agencies and startups must agree access to OpenAI's technologies separately.

OpenAI has launched the OpenAI for Chile partnership with the Chilean government: the program is run by the Ministry of Science, Technology, Knowledge and Innovation (according to the company). The memorandum was signed on November 17 by minister Jimena Lincolao at the Semafor "Silicon Valley & The World" conference in California. The initiative builds on Chile's National SI Strategy. It is backed by the US Commerce Department's SI export program, in which OpenAI is an anchor partner; it sits within the OpenAI for Countries framework. The parties intend to study the use of SI in the ministry's administrative processes and to train civil servants. The plan includes train-the-trainer workshops, verification of model responses, protection of sensitive data and human oversight. Support for local startups is also planned through access to OpenAI's programs and technologies. The memorandum contains no investment or procurement commitments: specific activities, resources and access to technology are agreed separately. Priorities will be set by a joint working group.

An artist has up to 28 days after publication to remove someone else's track from their profile, statistics and recommendations.

Spotify opens profile protection to all artists against deepfakes under their name

An artist has up to 28 days after publication to remove someone else's track from their profile, statistics and recommendations.

Spotify opened its Artist Profile Protection feature to all artists on November 17, after an eight-month beta test that began in March. The tool in Spotify for Artists sends notifications about releases delivered under an artist's name. Someone else's track can be declined: it does not appear in the profile, statistics or recommendations. By default, releases go out on schedule. When Block releases is enabled (a limited group is testing the expanded setting), publication must be confirmed. According to Music Business Worldwide, a release can be declined within 28 days of publication. By the end of September, Sony Music had requested the removal of more than 260,000 SI tracks imitating its artists; in March there were 135,000 such requests. Over 2026, Spotify removed 50 million tracks classified as spam and junk content from recommendations.

Australian authors and publishers are using the campaign to keep decisions on licensing their work with them until the SI bill package is introduced.

Australian creators launch Not AI's To Take campaign to keep copyright

Australian authors and publishers are using the campaign to keep decisions on licensing their work with them until the SI bill package is introduced.

Australian creative organisations and artists launched the Not AI's To Take campaign, demanding that the Albanese government not weaken copyright under pressure from SI companies. The launch is timed to the Aria awards ceremony on Wednesday. The campaign is backed by Aria, the Copyright Agency, the Australian Directors' Guild, Screen Producers Australia, the Australian Society of Authors and the Australian Writers' Guild. Foxtel, Free TV, Nine, Seven and Guardian Australia have joined. Musicians backing it include Jimmy Barnes, Paul Kelly and Peter Garrett. Paul Kelly said: "Australia had some of the best copyright laws... now is the time for our government to show some character." Industry Minister Tim Ayres said the status quo does not work for the creative sector and that an SI bill package is expected in late 2026 or early 2027. Senator David Pocock objected that changing the rules before more than 100 US court cases are resolved is risky. According to The Guardian, Anthropic offered a refusal model and promised Australian organisations access to frontier models regardless of the training decision. By industry estimates, the creative sector contributes $70 billion and 600,000 jobs.

Time savings on work tasks do not mean more work or better quality: the Bank of Korea survey found no such link.

Bank of Korea: 66.5% of workers use generative SI, time savings at 9.2%

Time savings on work tasks do not mean more work or better quality: the Bank of Korea survey found no such link.

Generative SI (ChatGPT, Gemini, Claude) is used by 66.5% of surveyed workers in Korea, up from 51.8% a year earlier, according to data published by the Bank of Korea on November 17. The surveys covered more than 16,000 people in 2025 and 2026. Weekly use rose from 5.9 to 8 hours, and 28.4% of users pay for subscriptions, on average 49,000 won (about $37) a month. According to data from other studies cited in the report, Korea's figure is higher than the US (43%) and Germany (31.5%). The share of time saved on work tasks rose from 3.8% to 9.2%, with a larger increase among workers with more education. The survey found no link between time savings and growth in the volume of work or in the quality of results.

Sources 1

Some standard operations on citizens' reports will be handled by a prototype SI assistant, while the decision on each report will remain with the district officer.

Russia's Interior Ministry to deploy SI assistant for district officers handling routine operations on reports

Some standard operations on citizens' reports will be handled by a prototype SI assistant, while the decision on each report will remain with the district officer.

SI capabilities are being "actively integrated" into the day-to-day work of district police officers. Prototypes of the SI assistant will take on routine processes and standard operations when reviewing citizens' reports. Interior Minister Vladimir Kolokoltsev said this on November 17 at a ceremony marking District Police Officer Day. According to him, the final decision will always rest with the police officer, and the freed-up resources will be directed to working with the public and to prevention. Over the year, according to the ministry, district officers solved one in four crimes, curbed about 1.2 million administrative offences and carried out checks on 8 million appeals.

Device developers on Modalix chips get up to 3.5x faster token generation without higher power consumption, and the open code lets them verify the claimed figures.

SiMa.ai speeds up SI models on Modalix chips by up to 3.5x

Device developers on Modalix chips get up to 3.5x faster token generation without higher power consumption, and the open code lets them verify the claimed figures.

SiMa.ai has released Palette 3.0, a software update for Modalix edge chips. The update was announced on November 17. It speeds up token generation by up to 3.5x within the same power budget thanks to speculative decoding with the EAGLE3, DFlash and Gemma 4 MTP algorithms. In the company's white paper, the Qwen3.5 9B model with DFlash went from 11.98 to 41.75 tokens per second at about 7.5 W of power consumption. Energy per token fell from 0.627 to 0.182 J. Across four models the speedup was 2.21–3.49x, and in an automotive assistant test 1.35–2.06x, with 900 requests out of 900 handled correctly. Palette 3.0 allows several models and separate Docker containers to run in parallel on a single chip. The GPT-OSS-20B mixture-of-experts model is supported on the 16 GB Modalix module; the code is open on GitHub.

Developers can order home elements from the company's factories in Calgary and Nisku instead of building their own production, and the first US factory is planned for Phoenix in 2027.

Promise Robotics raises C$40 million for robotic homebuilding factories

Developers can order home elements from the company's factories in Calgary and Nisku instead of building their own production, and the first US factory is planned for Phoenix in 2027.

Canadian company Promise Robotics has announced the closing of a C$40 million Series B round. The round was led by White Star Capital, with participation from Horizons Ventures, Radical Ventures and the Public Sector Pension Investment Board. The physical SI platform converts existing home designs into robotic production plans. The company supplies structural elements from factories in Calgary and Nisku (Alberta) under a Factory-as-a-Service model. The developer pays for output rather than for building its own factory. According to the company, more than 2,500 housing units are under contract in three markets and about 200 homes have been built. The first factory in the US is planned to open in Phoenix, Arizona in 2027, following a multi-year agreement with D.R. Horton for more than 1,250 homes.

The agreement expands British actors' rights over digital replicas, but it is not yet enshrined in law.

The Guardian: Britain prepares to expand actors' rights over digital replicas under SI

The agreement expands British actors' rights over digital replicas, but it is not yet enshrined in law.

British actors will get more rights over their digital replicas as part of an agreement on protections when SI is used. The Guardian reported this on November 17. These are measures still being agreed, not a law in force.

A materials scientist can run alloy calculations without programming: a natural-language request turns into a reproducible set of computations on their own computer.

Empa releases OptiMat Alloys SI agent for high-entropy alloy calculations

A materials scientist can run alloy calculations without programming: a natural-language request turns into a reproducible set of computations on their own computer.

Empa's labs in Thun (Vladislav Turlo, Yan Hu) have released the OptiMat Chat platform, into which the OptiMat Alloys SI agent is loaded. The agent selects domain tools on its own, runs predefined computational scenarios for high-entropy alloys and presents the results as graphs, tables and text. The user formulates a request in natural language, with no programming; the calculation runs on their own computer rather than on the language model's servers. According to the authors, structured tool calls and storing calculations in a database reduce the number of fabricated results. The results are positioned as a starting point for an expert, not a finished publication. The work was published in npj Computational Materials; Empa's press release was published on November 17, 2026.

The method shows which domain walls make the main contribution to losses, allowing alloy composition for transformers and motors to be selected purposefully instead of by brute force.

Scientists on Fugaku identify which domain walls contribute to alloy losses

The method shows which domain walls make the main contribution to losses, allowing alloy composition for transformers and motors to be selected purposefully instead of by brute force.

A group led by Masato Kotsugi (Tokyo University of Science) with colleagues from the universities of Tsukuba, Okayama, Kyoto, Tohoku and Mainz applied explainable SI to a micromagnetic simulation. The calculations ran on the Fugaku supercomputer for a nanocrystalline soft magnetic FINEMET-type alloy at frequencies of 100 kHz, 1 MHz and 10 MHz. Magnetic domain structures were described using persistent homology, and features were mapped to energy via interpretable machine learning. The authors call the approach an "extended free energy model". As frequency rises, domains break up and the energy barrier to magnetisation reversal grows. Reverse mapping of the features showed that the main contribution to losses comes from reverse domain peaks and some of the small domain walls. The paper was published online in NPG Asia Materials on November 17, 2026.

SI adoption at banks and large companies has entered a cautious phase, because the security of direct interaction between SI agents and users has not been resolved.

VTB deputy chair Skorobogatova: banks have slowed SI adoption over risks

SI adoption at banks and large companies has entered a cautious phase, because the security of direct interaction between SI agents and users has not been resolved.

Large companies, including banks, have moved from the phase of active SI adoption to more cautious use of the technology. That is how they hope to avoid reputational and financial losses, VTB's first deputy chair of the management board Olga Skorobogatova said on November 17 at the FINOBOD forum in Tashkent. According to her, "with SI you want to run very fast, but slowly" — it is better to check and test a solution than to get a negative experience. In the direct interaction of SI agents with the company and the user, an unresolved security question remains, and it will have to be answered in the near future.

Sources 1

SI training for India's schoolteachers has gained a quantitative goal — more than a million educators in 12–18 months.

Bodhan AI at IIT Madras plans to train over a million teachers in SI in 12–18 months

SI training for India's schoolteachers has gained a quantitative goal — more than a million educators in 12–18 months.

The Bodhan AI Centre of Excellence at IIT Madras announced on November 17 partnerships with educational organisations to train more than a million schoolteachers in India to work with SI. The programme is designed for 12–18 months and is aimed at SI literacy and integrating tools into classroom work. The task is to help educators use SI as a practical assistant in day-to-day teaching. The project's mission is to build basic SI infrastructure for India's multilingual educational environment.

Sources 1

The graph neural network approach flags a refrigerant leak on the DATARK test bench on average 200 minutes after it appears, with F1 = 0.87.

Graph neural network from HSE University and UrFU spots data center cooling leaks in 200 minutes

The graph neural network approach flags a refrigerant leak on the DATARK test bench on average 200 minutes after it appears, with F1 = 0.87.

Researchers from HSE University and UrFU have developed an approach based on graph neural networks to detect refrigerant leaks and heat exchanger contamination in data center cooling systems. The model takes into account the engineering layout of the loop — compressor, evaporator and condenser — and predicts expected values, registering a sustained divergence from the forecast. TASS reported the development on November 17, citing the university's press service. On an experimental test bench of an industrial cooling system from DATARK, the SI detected leaks with F1 = 0.87. On average it found the fault 200 minutes after it appeared, which is better than the algorithms previously used. The model ran on data from 12 sensors and 4 calculated indicators at one-minute intervals; cooling accounts for 30–50% of a data center's power consumption. The results have been published in IEEE Transactions on Industry Applications.

The complaint filed did not change the date: the CSAT remains scheduled for November 19, and the authorities refused to move the exam.

High schooler An Min-jun challenges CSAT date in Korea's Constitutional Court

The complaint filed did not change the date: the CSAT remains scheduled for November 19, and the authorities refused to move the exam.

High school student An Min-jun is challenging in Korea's Constitutional Court the date of the national CSAT exam, set for November 19. On November 17 he filed a complaint and a motion to suspend the date until the court rules. The complaint argues that holding the exam without an effective system of protection against sophisticated forms of cheating violates the right to fair competition. Both the date set by the Korea Institute for Curriculum and Evaluation and the Ministry of Education's decision are being challenged. Education authorities confirmed that the exam will not be moved, including because of the risk of leaking materials that have already been distributed to test sites.

From December 8, ADX will host Lunate's fifth thematic ETF, with 41 companies from the robotics chain and a TER of 0.49%.

Lunate to list humanoid robot ETF on Abu Dhabi exchange on December 8

From December 8, ADX will host Lunate's fifth thematic ETF, with 41 companies from the robotics chain and a TER of 0.49%.

Abu Dhabi investment group Lunate announced on November 17 that it will list the Boreas VettaFi Humanoid Robots UCITS ETF on the Abu Dhabi Securities Exchange under the ticker ROBOTS. Trading begins on December 8; it is its fifth thematic ETF on ADX. The fund tracks the VettaFi Humanoid Robot Select Markets Net Total Return index. The starting basket will include 41 companies from Asia, the US and Europe across four links of the chain: semiconductors, components and sensors, humanoid developers and software. The fee (TER) will be 0.49%; the fund is registered in Dublin as a European UCITS, with parallel trading on Xetra. Sherif Salem of Lunate and Geir Espeskog of Northwind explain the launch by falling robot costs and labour shortages. Because of that shortage, applications may move out of factories into hospitals and homes.

The lab's focus is automating small-batch value-added services, which already account for about half of warehouse hours at Blue Hub.

CEVA Logistics and Singapore Institute of Technology open ILAS lab in Singapore

The lab's focus is automating small-batch value-added services, which already account for about half of warehouse hours at Blue Hub.

The Intelligent Logistics & Autonomous Systems (ILAS) laboratory has begun operations in Singapore: it was opened on 17 November by CEVA Logistics and Singapore Institute of Technology. The ceremony was led by Singapore's Minister of State for Foreign Affairs and Trade and Industry Gan Siow Huang. The initiative is supported by the Singapore Economic Development Board and the National Robotics Programme. The five-year partnership focuses on robotics, SI image recognition and linking robots with warehouse systems. The main goal is value-added services: kitting, labelling and packing. Traditional lines are poorly suited to this segment because of the broad product range and small batch sizes. CEVA Logistics plans to deploy the first solutions at the Blue Hub distribution centre in Singapore within one to two years. About half of warehouse hours at Blue Hub go to such services.

Auchan Online has gained a chatbot that compares products by price, weight and promotions on its own and assembles a basket to fit a recipe, occasion and budget.

Auchan Retail Russia launches Klik SI assistant in Auchan Online

Auchan Online has gained a chatbot that compares products by price, weight and promotions on its own and assembles a basket to fit a recipe, occasion and budget.

The Klik SI assistant has started working in the Auchan Online service in a Russian messenger. Auchan Retail Russia carried out the launch on 17 November 2026. The chatbot selects products taking into account recipes, the occasion for the purchase and a set budget. It compares products, looks for alternatives, filters the assortment by price, weight and promotions, and also analyses the entire basket. Ekaterina Alferova, sales director for the Hypermarkets format and digital sales at Auchan Retail Russia, said the functionality will be gradually expanded.