Show only each day's banner story

Open-weight models take 30.9% of token traffic in Databricks SI Index

With an identical 89.1% result, the average price of a solved task differs 32-fold between GPT-6 Luna and Claude Opus 5.5 ($0.05 vs $1.62), so the choice between them is determined by the cost of a task, not the price of a token.

Databricks has released a public SI index built from aggregated data from more than 10 thousand companies in Unity Gateway. The share of open-weight models in customers' token traffic grew from 7.3% in July to 30.9% in November, while the share of companies relying on a single provider fell from 64% to 28% by November 2. The index combines Unity Gateway data with Databricks' internal tests for coding agents. In the coding-agent benchmark, GPT-6 Luna and Claude Opus 5.5 both solved 89.1% of tasks, but the average cost per task was $0.05 and $1.62 respectively. According to Superpower Daily's breakdown, Gemini 3.8 Flash spent 140 steps and $11.97 on a single task.

The scenario Google recorded shows that once a patch is out, an agent can write and run an exploit on its own, so updates need to be installed as fast as possible.

Google Threat Intelligence Group records first case of a patch being turned into an exploit automatically, with no human involved

The scenario Google recorded shows that once a patch is out, an agent can write and run an exploit on its own, so updates need to be installed as fast as possible.

According to The Register, in its quarterly GTIG AI Threat Tracker report for Q4 2026, Google Threat Intelligence Group said it had for the first time observed fully automated vulnerability discovery and exploitation with no human in the loop. On an open staging server it found a multi-agent framework called exploit_agent connected to several frontier and local models: it continuously compares the code of popular web applications before and after a patch, finds the vulnerability, writes and runs an exploit; the same server held data stolen from victims and scripts for reconnaissance, brute-forcing, installing web shells and removing competitors' web shells. Google did not attribute the activity to a specific group but noted Russian-language comments in the configuration and, in the victims' telemetry, an interest in French websites in media, entertainment and education; the report also records a rise in zero-day exploitation from 8 a month in 2025 to 11 a month in 2026 and 22 in August 2026.

From November 16, medical organisations can run agents in their existing software through the OpenAI API and ChatGPT for Healthcare/Clinicians; clinics and insurers such as CVS Health are already using them for routine work, and PHI requires a BAA.

OpenAI launches computer-use agents in ChatGPT for Healthcare for clinics and insurers

From November 16, medical organisations can run agents in their existing software through the OpenAI API and ChatGPT for Healthcare/Clinicians; clinics and insurers such as CVS Health are already using them for routine work, and PHI requires a BAA.

According to OpenAI, on November 16 OpenAI announced an expansion of ChatGPT for Healthcare: computer-use agents carry out multi-step tasks in existing medical software — reading the screen, clicking, typing and switching between applications; access permissions, audit logs and more than 20 plugins have been added, and clinical search with source citations has been updated. By OpenAI's estimates, 99% of 325 reviewed tasks for GPT-6 Astra were judged clinically safe, and 93% of 646 physician ratings called the work "very useful" or "extremely useful"; on HealthAdminBench, Astra scored 94.3% of rubric points and GPT-6 Luna 91.8%, at 8.5 cents per task. Deployment partners include Abridge, Ambience, Cigna Healthcare and CVS Health; CVS Health uses the agents to submit documents, process referrals and handle tens of thousands of consultations a day. The capabilities are available from November 16 through the OpenAI API, ChatGPT for Healthcare and ChatGPT for Clinicians; patient data is not used to train models, and a BAA is required to work with PHI.

Once the rule takes effect, a data centre in any country will need a US licence if it can provide computing to companies from D:5, including China and Russia, above the 12 million TPP threshold.

US BIS issues interim final rule on licences for remote access to advanced SI chips

Once the rule takes effect, a data centre in any country will need a US licence if it can provide computing to companies from D:5, including China and Russia, above the 12 million TPP threshold.

According to Bloomberg, on November 16, 2026, the US Commerce Department's Bureau of Industry and Security (BIS) issued an interim final rule requiring licences for the export, re-export or transfer of advanced "super intelligence" (SI) chips and servers to data centres in any country if they can provide computing to companies from D:5 (China, Russia, Iran, North Korea and 18 other jurisdictions) above 12 million TPP (roughly a maximum of 750 Nvidia H100/H200 chips); existing operators must obtain licences within 90 days. Separate licences are introduced for Malaysia, Singapore, Indonesia, Thailand and Brazil as countries at heightened risk of diversion, except for shipments to US companies; commercial recommendation systems are carved out, and the Biden-era AI Diffusion Rule is rescinded. According to BIS, the rule was written at the request of the Treasury Department and the National Security Council, takes effect upon publication in the Federal Register and provides 60 days for comments.

The Qwen4-Max flagship is available only in Alibaba's cloud, so it cannot be run locally; the company gave no timeline for smaller open models.

Alibaba opens access to Qwen4-Max: a 2.4 trillion parameter flagship with closed weights

The Qwen4-Max flagship is available only in Alibaba's cloud, so it cannot be run locally; the company gave no timeline for smaller open models.

On November 16 Alibaba officially unveiled Qwen4-Max, the most powerful model in the Qwen family: 2.4 trillion total parameters, 71 billion activated per token, native vision and a 1 million token context; the model is built on a new MoE architecture, Qwen4. According to Alibaba, the average cost per task in Work360 falls from $4.43 to $1.67, and the API costs $1.50 per 1 million input and $4.50 per 1 million output tokens; access is open through the QwenCloud API, Qwen Studio, QwenWork and Qoder. Qwen4-Max's weights are closed, and smaller open models in the Qwen4 family were promised later without a date; a limited preview began on November 11. Alibaba also says it filters training data against offensive cyber capabilities and has strengthened defences against CBRN risks.

The open-source world model, together with the training framework, gives robotics teams a ready-made base for training their own machines on Chinese Ascend accelerators instead of Western chips.

Huawei Cloud open-sources the CloudRobo-WM1.2 world model for robots

The open-source world model, together with the training framework, gives robotics teams a ready-made base for training their own machines on Chinese Ascend accelerators instead of Western chips.

On November 16, at its online CloudRobo conference, Huawei Cloud released and open-sourced the CloudRobo-WM1.2 world model for embodied SI, trained on Chinese Ascend chips. The company claims five key advances: a PPA multimodal representation architecture, DualAct action generation, adaptive pretraining on heterogeneous data, cross-embodiment motion modelling and four combinable inference modes; it claims first place on the RoboTwin 2.0 and LiberoPro 2.0 benchmarks. Alongside the model, the training framework and Ascend optimisation practices are open-sourced, and Huawei Cloud says 80% of Chinese companies in embodied SI use Ascend Cloud. The same day, a pre-sale was announced with Fulongma Group for a new sweeping robot built on CloudRobo-WM1.2, to be presented in Shanghai in December.

With no binding protocol in place, the rules for using autonomous weapons are set by individual militaries' national programmes, while international frameworks remain voluntary.

UN review conference on autonomous weapons opens in Geneva as Guterres calls for a binding treaty

With no binding protocol in place, the rules for using autonomous weapons are set by individual militaries' national programmes, while international frameworks remain voluntary.

On November 16 the seventh review conference of the 128 states parties to the UN Convention on Certain Conventional Weapons (CCW) opened in Geneva, where the question is whether to launch negotiations on a legally binding instrument on lethal autonomous weapons systems (LAWS). A decision is expected by Friday, November 20, and the consensus rule allows any single country to block negotiations. In a video message, UN Secretary-General António Guterres called for "moving towards negotiations on a legally binding instrument" with "clear prohibitions and restrictions" and for preventing the use and proliferation of weapons he called "ruthless"; UN High Commissioner for Human Rights Volker Türk also called for urgently agreeing binding rules. A proposal to begin negotiations was put forward by seven countries — Austria, Belgium, Brazil, Ireland, Mexico, New Zealand and Panama; the conference chair proposed a two-year mandate to develop a binding protocol, while the UK, Germany and France propose not to pre-determine the binding nature of the final document. According to Reuters, the US, Russia and Israel oppose binding norms, considering existing international humanitarian law sufficient: the US proposes a five-year process with a non-binding document, while the International Committee of the Red Cross and non-governmental organisations insist the work be completed within a year. Konstantin Vorontsov, deputy director of the Department for Non-Proliferation and Arms Control at the Russian Foreign Ministry, said on November 16 that Russia sees no need for restrictive or prohibitive measures beyond existing international law, including international humanitarian law, and considers additional bans premature. In October the UN General Assembly passed a resolution in favour of negotiations: 163 votes in favour, with the US, Russia, Belarus, Israel and North Korea against, and nine countries abstaining, including China, Iran and Saudi Arabia. According to Bernama, experts see no evidence of fully autonomous weapons being used in conflicts; according to Reuters, semi-autonomous systems with SI have already been used in Ukraine, Sudan and the Middle East.

Nine countries have turned SI supply chain protection into joint working groups: they will have to agree policy frameworks, supply security and "safe harbor" regimes for critical goods, software and infrastructure.

US, Japan, South Korea and six other countries sign San Francisco Declaration on protecting SI supply chains

Nine countries have turned SI supply chain protection into joint working groups: they will have to agree policy frameworks, supply security and "safe harbor" regimes for critical goods, software and infrastructure.

The US, South Korea, Japan, Australia, Israel, the UK, India, Qatar and the UAE have signed the San Francisco Declaration on the Secure Foundations of Enduring Prosperity under the Pax Silica programme; the signing was announced on 16 November by US Deputy Secretary of State for Economic Affairs Jacob Helberg. The document provides for the protection of critical goods, software, knowledge and components of SI supply chains, as well as physical and digital infrastructure — ports, transport, communications and logistics; the signatories plan joint working groups on policy frameworks, SI adoption in key sectors, supply security and "safe harbor" regimes; according to KBS, a working group and an economic hub for SI investment are envisaged. For Seoul, Second Vice Foreign Minister Kim Jina signed at a meeting in San Francisco; Japan's Foreign Ministry confirmed the signing; Germany, Sweden, the Netherlands and the European Commission are expected to join later, while Chinese Foreign Ministry spokesperson Mao Ning said that in the SI field "ideological lines should not be drawn or closed exclusive circles created".

After the deal, Fei-Fei Li reports directly to Lisa Su and leads research at AMD, while the World Labs team will continue to research models and, together with AMD engineers, adapt them to the company's future SI infrastructure.

AMD completes World Labs acquisition; Fei-Fei Li becomes executive vice president and chief scientist at AMD

After the deal, Fei-Fei Li reports directly to Lisa Su and leads research at AMD, while the World Labs team will continue to research models and, together with AMD engineers, adapt them to the company's future SI infrastructure.

AMD announced on 16 November that it had completed its acquisition of World Labs — a company founded by SI researcher Fei-Fei Li that develops models for creating and simulating interactive 3D environments, as well as robot training and simulation technology. Fei-Fei Li has taken the post of executive vice president and chief scientist at AMD and reports directly to chair and CEO Lisa Su. According to Lisa Su, the World Labs team will continue fundamental model research and will work with AMD engineers on applying the models' new capabilities in future SI infrastructure for the open ecosystem; the deal value was not disclosed in the announcement.

For now this is an announcement: the SMART CXL cards must pass customer qualification by the end of 2026, and servers with Photonic Fabric must sample in the second half of 2027.

Marvell and Penguin Solutions unveil HBM-class photonic memory appliance for memory sharing across racks

For now this is an announcement: the SMART CXL cards must pass customer qualification by the end of 2026, and servers with Photonic Fabric must sample in the second half of 2027.

On 16 November at SC26 in Chicago, Marvell and Penguin Solutions announced a strategic collaboration and unveiled the MemoryAI Photonic Appliance — according to Penguin, the first HBM-class photonic memory appliance. It allows CPUs and GPUs to share disaggregated memory across racks via Marvell's Photonic Fabric and CXL 3.0/3.1; claimed figures are up to 57.6 Tbit/s of aggregate bandwidth, up to 16 TB of DDR5 RDIMM and 576 GB of HBM3E per appliance, plus HBM Direct, HBM-DDR5 Cached and DDR5 Direct modes. Penguin SMART CXL cards on Marvell Structera X controllers will enable reuse of DDR4 from decommissioned servers. The cards are expected to be qualified with customers by the end of 2026, and joint servers with Photonic Fabric are expected to sample in the second half of 2027.

NeuroXess disclosed the results: 84% of continuous tasks on the first pass, latency under 50 ms and over 95% small-object grasping — but only with a single clinical participant at its own facility.

NeuroXess says its wireless BCI completed 84% of continuous tasks on the first pass in a simulated home environment

NeuroXess disclosed the results: 84% of continuous tasks on the first pass, latency under 50 ms and over 95% small-object grasping — but only with a single clinical participant at its own facility.

On 16 November 2026 NeuroXess disclosed the results of testing its fully implantable, fully wireless and fully functional BCI at its own simulated-living test facility. According to the company, a patient with high tetraplegia after a car accident, using a BCI + SI model + computer vision setup, performed scenarios of mobility, home, communication and work: latency under 50 ms, about 100% accuracy when switching devices, 84% of continuous tasks completed on the first pass, and over 95% success grabbing small objects with an exoskeleton glove. Transferring skills from home scenarios to office ones cut the average learning time for a new task from 10.5 to 2 hours; the facility is designed for 20 people and, by the company's estimate, will yield about 18,000 hours of interactive data a year.

According to Xint.io, the cost of finding this vulnerability fell from $350 to $4.24 in six months, and frontier models have become accessible to "mediocre hackers", but framing the task still requires expertise.

Xint.io: GPT 6 Sol found CVE-2026-102321 in Google V8 in 13 minutes for $4.24

According to Xint.io, the cost of finding this vulnerability fell from $350 to $4.24 in six months, and frontier models have become accessible to "mediocre hackers", but framing the task still requires expertise.

Xint.io tested seven models released since February 2026 on finding a type confusion vulnerability in Google V8 (CVE-2026-102321) that leads to RCE and was absent from training data because it was disclosed through Google VRP. In February the best models did not find it in 25 hours and $350; in July GPT 5.6 Sol managed it in 44 minutes and $22; in September GPT 6 Sol did it in 13 minutes and $4.24 on the first prompted query, Claude Opus 5 — on the first attempt for $43.40, while DeepSeek and Z.ai did not solve the task. Xint.io attributes this to falling token prices — on average about 47% per quarter since 2023 and 15% in the past month — and concludes that frontier models are accessible to "mediocre hackers", although expertise in framing the task remains important.

Sources 1

Capacity for SI workloads is increasingly built and sold not by the clouds themselves but by infrastructure funds: AWS remains a tenant for decades, while the capital for construction comes from the REIT market.

Blackstone Digital Infrastructure Trust buys 75.1% of two AWS joint ventures for $5 billion

Capacity for SI workloads is increasingly built and sold not by the clouds themselves but by infrastructure funds: AWS remains a tenant for decades, while the capital for construction comes from the REIT market.

Blackstone Digital Infrastructure Trust (NYSE: BXDC) announced on 16 November that it had signed a definitive agreement to buy 75.1% of two joint ventures with Amazon Web Services for roughly $5 billion at an initial capitalization rate of 7.0%. The JVs receive 13 newly built "powered shells" (data center shells with power brought in) across two AWS campuses: nine buildings in northern Indiana (the Greater Chicago market) and four in Mississippi, with a combined IT capacity of 647 MW. AWS retains 24.9% and leases the entire portfolio on a triple-net basis with an initial 12-year term and extensions up to 32 years. This is BXDC's first investment since its May listing; funding is $1.9 billion from IPO proceeds and about $2.5 billion in non-recourse mortgage loans, with closing expected in Q4 2026. According to Yahoo Finance, BXDC shares rose nearly 6% to $17.67 on the news.

Autonomous cars are reaching passengers through a familiar app and at no markup — the robotaxi is ceasing to be a separate service with its own app and becoming one of the options for an ordinary ride.

Zoox robotaxis start taking orders through the Uber app in Las Vegas

Autonomous cars are reaching passengers through a familiar app and at no markup — the robotaxi is ceasing to be a separate service with its own app and becoming one of the options for an ordinary ride.

According to Bloomberg, on 16 November Uber and Amazon-owned Zoox announced that fully autonomous Zoox robotaxis have appeared in the Uber app in Las Vegas — Zoox's first outing on a third-party platform. Passengers ordering UberX, Comfort, Electric, Comfort Electric or Wait & Save can get a Zoox at no extra charge; they cannot select the vehicle directly, but can enable a preference for autonomous rides in the settings, and Uber, unlike with other partners, will not be servicing these orders, TechCrunch notes. A multi-year partnership with MGM Resorts International as the official rideshare and robotaxi partner for its Las Vegas properties was announced at the same time. The parties plan a Los Angeles launch in 2027; Zoox will start its own rides there by the end of 2026; Zoox began paid rides in Las Vegas in August 2026, and the Nevada permit allows up to 1,000 robotaxis in Clark County for Motional and Zoox.

XPeng has fixed the IRON timeline: 1 January 2027 — production start, Q1 2027 — XPeng stores, Q2 2027 — external commercial customers; the previous target of large-scale production by the end of 2026 has been replaced.

XPeng: series production of the IRON robot to begin January 1, 2027, external customers in Q2

XPeng has fixed the IRON timeline: 1 January 2027 — production start, Q1 2027 — XPeng stores, Q2 2027 — external commercial customers; the previous target of large-scale production by the end of 2026 has been replaced.

On 16 November, on a conference call about Q3 2026 results, XPeng chairman and CEO He Xiaopeng said that series production of the IRON humanoid robot begins on 1 January 2027. The first batch in Q1 2027 will go to select XPeng stores for trial operation, and in Q2 2027 IRON will be officially offered to external commercial customers. The new version will be shown at XPeng Tech Day and the Guangzhou auto show; among the stated capabilities is recognizing multiple people and remembering their identities.

The claimed 20-fold drop in task cost is measured by Tencent's internal benchmark and has not yet been confirmed by independent tests.

Tencent Hunyuan releases Hy4 flash: 12 billion activated parameters, 1M-token context and 0.5 yuan per million input tokens

The claimed 20-fold drop in task cost is measured by Tencent's internal benchmark and has not yet been confirmed by independent tests.

On 16 November Tencent Hunyuan introduced Hy4 flash — a model with 224 billion total parameters, 64 billion of which are Engram parameters, with 12 billion activated per token, multimodality and a 1 million token context. API pricing is 0.5 yuan per 1 million input tokens and 2 yuan per 1 million output tokens, or 0.02 yuan when a request hits the cache; decoding runs at over 150 tokens/s. On the internal WorkIndex benchmark, the cost of solving a task is 1/20 of Hy4, 1/13 of Hy4 preview and 1/8 of DeepSeek-V4.1-Flash-1017 in maximum mode at the "daytime" rate. The model is available in Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis; open weights are promised later without a date.

Sources 1

Arm holds first place in TOP500 and the entire top four of Green500: speed and energy efficiency now belong to the same architecture.

Arm: its architecture took the top four Green500 places and first place in TOP500

Arm holds first place in TOP500 and the entire top four of Green500: speed and energy efficiency now belong to the same architecture.

The top ten of the November TOP500 is unchanged since June: the 68th edition of the ranking was announced on November 16 in Chicago at the SC26 conference. China's LineShine kept first place with 2.198 exaflops on the HPL test. It is followed by El Capitan (1.809), Frontier (1.353), Aurora (1.012) and JUPITER Booster (1.000). Four of the ten fastest systems are built on Arm. LineShine uses custom Arm-compatible LingKun LX2 cores, and JUPITER Booster in Jülich uses NVIDIA Grace Hopper with Arm Neoverse. Fugaku is built on Fujitsu A64FX processors, and Alps also uses Grace Hopper. All four top Green500 places are KAIROS (73.28 GFlops/W), ROMEO-2025 (70.91), Levante GPU extension (69.43) and Isambard-AI phase 1 (68.84). Each of these systems uses NVIDIA GH200 on the Arm Neoverse platform. According to Arm, five such systems are in the top 10 for energy efficiency and 14 in the top 30. Only 27 systems on the list are new, 5.4%, the lowest turnover in the ranking's history. The entry threshold at 500th place rose 7.3% to 2.857 petaflops. The combined performance of all 500 systems grew 4.4% to 19.57 exaflops.

The deal gives 17 LVMH brands a ChatGPT Enterprise rollout and placement of their material in ChatGPT with links to the originals, on non-exclusive terms.

LVMH media division signs OpenAI deal for ChatGPT Enterprise and surfacing of 17 brands' content

The deal gives 17 LVMH brands a ChatGPT Enterprise rollout and placement of their material in ChatGPT with links to the originals, on non-exclusive terms.

According to OpenAI, on November 16 in Paris LVMH's media division and OpenAI announced a strategic agreement: OpenAI tools and ChatGPT Enterprise are being deployed in the newsrooms of 17 brands — Les Echos, Le Parisien, Paris Match, L'Opinion, Challenges, L'Agefi and others — with an audience of 27 million readers a month and roughly 700 journalists among 1,800 employees. In ChatGPT, the publications' material will appear as summaries and excerpts with source attribution and direct links to the original articles. Financial terms were not disclosed; the agreement is non-exclusive and is the second for the French media group after Le Monde.

Baidu's SI business has for the first time become more than half of General Business, but total revenue is still 1% below last year's level.

Baidu: AI Cloud Infra revenue up 94%, total revenue down 1%

Baidu's SI business has for the first time become more than half of General Business, but total revenue is still 1% below last year's level.

Baidu published its Q3 2026 results on 16 November. AI Cloud Infra revenue grew 94% year on year to RMB8.2 billion, GPU Cloud grew 191% year on year and 43% quarter on quarter, and revenue from external customers' tokens on the Qianfan platform rose 1,159%. Total revenue fell 1% to RMB30.8 billion ($4.60 billion), online marketing dropped 22% to RMB11.9 billion; Baidu Core AI-powered Business grew 37% to RMB13.2 billion and for the first time accounted for 53% of General Business revenue. Baidu's net income was RMB1.4 billion ($208 million), operating income was RMB626 million at a 2% margin; the company returned $373 million through share repurchases and expects a first dividend by the end of 2026 subject to board approval.

For now this is an application to the InvestAI competition: the consortium has been announced, but no funding decision has been announced.

Nokia, Nscale, Telia, Verda and Skeleton announce AI Gigafactory Northern consortium

For now this is an application to the InvestAI competition: the consortium has been announced, but no funding decision has been announced.

On November 16 Nokia, Nscale, Skeleton, Telia (with Telia Cygate) and Verda announced the creation of the AI Gigafactory Northern consortium for a distributed network of SI gigafactories in Finland, Estonia, Latvia and Sweden. Sites have been designated in Oulu, Iru and Valmiera; the fourth site in Sweden has not yet been chosen. The project is entered in the European competition for large SI gigafactories under InvestAI (€30 billion) and is meant to combine compute, sovereign cloud and SI platforms, high-speed connectivity and low-carbon energy. Roles are divided up: Nokia — secure networks, Nscale — cloud platform and SI infrastructure, Telia — connectivity and operations, Verda — training and inference infrastructure, Skeleton — power systems.

Southeast Asian countries now simultaneously have a Chinese SI cooperation program for 2027 and new US licenses for SI chip supplies.

China and ASEAN declare 2027 the Year of Superintelligence Cooperation

Southeast Asian countries now simultaneously have a Chinese SI cooperation program for 2027 and new US licenses for SI chip supplies.

On November 16, 2026, at the 29th China–ASEAN Summit in Manila, the sides declared 2027 the Year of Superintelligence Cooperation. Chinese Premier Li Qiang said China is ready to work with ASEAN to ensure the success of that year and called "AI plus" cooperation an engine of regional economic development. On the same day, the US Commerce Department introduced licenses for SI chip supplies to Malaysia, Singapore, Indonesia and Thailand.

The regulator has recorded that banks already see attempts by SI agents to act in retail accounts without confirmable authorisation, and called customer consent verification and technical limits on agents necessary.

Fed governor Michael Barr warns SI agents could speed bank runs up to minutes

The regulator has recorded that banks already see attempts by SI agents to act in retail accounts without confirmable authorisation, and called customer consent verification and technical limits on agents necessary.

According to Bloomberg, on November 16, 2026, at a conference in Milan, Fed governor Michael Barr warned that autonomous SI agents could automatically move deposits in response to market signals or rate differentials and accelerate bank runs to a matter of minutes, outpacing traditional supervisory procedures. According to him, large banks are already recording attempts by agents to act in retail accounts whose authorisation by the customer cannot be confirmed. Barr also cited the risks of herd behaviour, market manipulation and cyber threats, referring to the Fed's Financial Stability Report, in which 87% of respondents called SI a substantial risk to the financial system.

Plural now has €1 billion to invest in European startups in energy, SI infrastructure and defence; about two-thirds of the total is earmarked for an early-stage fund making 25–35 investments.

Plural raises €1 billion fund for European startups in energy, SI infrastructure and defence

Plural now has €1 billion to invest in European startups in energy, SI infrastructure and defence; about two-thirds of the total is earmarked for an early-stage fund making 25–35 investments.

London-based venture firm Plural announced on November 16 that it had raised €1 billion ($1.1 billion) — the largest fund in its history, almost four times its first fund from 2022 (€260 million) and twice its second (€500 million, 2024). About two-thirds will go to a third early-stage fund (25–35 investments), with the rest going to an opportunities fund for existing portfolio companies; the focus is energy, SI infrastructure including chips, defence and health. Partner Taavet Hinrikus told Reuters the firm will look at SI infrastructure, including chip manufacturing, robotics and defence; the portfolio includes Helsing, Proxima Fusion, OLIX, The Exploration Company and Hypersonica. British Business Bank is contributing €75 million — its first investment in Plural.

The AI-Quryltai SI environment was used for the first time to prepare an official bill — the children's digital safety draft with more than 120 amendments has been sent for the government's opinion.

Kurultai vice-speaker: children's digital safety bill first drafted in the AI-Quryltai SI environment

The AI-Quryltai SI environment was used for the first time to prepare an official bill — the children's digital safety draft with more than 120 amendments has been sent for the government's opinion.

Kurultai Vice-Speaker Dinara Zakiyeva said at a Government Hour on November 16, 2026 that the draft law "On Digital and Information Security of Children" became the first bill developed in the AI-Quryltai intelligent environment. According to her, a specially trained legal model combines a database of more than 800,000 regulatory legal acts and was used for the first time in drafting the text; more than 120 amendments were prepared, including mandatory labeling and verification of digital products, content filtering and "parental control" on platforms. All SI proposals, she asserted, were reviewed and refined by deputies together with invited lawyers and experts; the document has been sent for the government's opinion and will be further developed by a working group.

The lawsuit by five Canadian media companies against OpenAI will be heard on the merits in Ontario, not in a US court.

Ontario appeal court dismisses OpenAI appeal: Canadian media lawsuit stays in Canada

The lawsuit by five Canadian media companies against OpenAI will be heard on the merits in Ontario, not in a US court.

On November 16 the Ontario Court of Appeal dismissed OpenAI's appeal and upheld the 2025 first-instance ruling on jurisdiction over the lawsuit brought by a coalition of Canadian media outlets. Filed in November 2024, the suit brings together La Presse canadienne, Torstar, The Globe and Mail, Postmedia and CBC/Radio-Canada; the outlets accuse OpenAI of copyright infringement and of "permitting and inducing" third parties to infringe rights in their content. The court said OpenAI had not shown a conflict with US law and that Ontario is the "most appropriate, adequate and convenient" forum, while arguments about whether infringement actually occurred go to the merits of the dispute rather than to jurisdiction.

Most of the zero-days VicOne found are in development and simulation tools, and one of the vulnerabilities was used in a demonstration to block a robot's remote stop.

VicOne: SI-assisted research finds 28 zero-day vulnerabilities in robotics software in under two days

Most of the zero-days VicOne found are in development and simulation tools, and one of the vulnerabilities was used in a demonstration to block a robot's remote stop.

On November 16, IEEE Spectrum published sponsored content from VicOne, in which the company said that with the help of SI-assisted research it found 28 zero-day vulnerabilities in less than two days in software and tools of the Physical AI ecosystem, including projects from Google, NVIDIA and Meta. According to initial validation, all 28 were classified as zero-days, some as yet without CVE IDs; 75% of the findings were in development, training, simulation and validation tools — the stages before a robot is deployed. In a separate demonstration, researchers used one of the vulnerabilities to block a robot's remote stop command.

CAICT's public position moves some "red lines" for SI agents from a voluntary declaration into the category of potentially mandatory requirements.

CAICT says some "red lines" for SI agents may become mandatory

CAICT's public position moves some "red lines" for SI agents from a voluntary declaration into the category of potentially mandatory requirements.

On November 16 at 15:35 Beijing time, the outlet Bǎiké Cáijīng published an interview with Zhang Xie, director of the industrial development and compliance management department at the intellectual property innovation centre of the China Academy of Information and Communications Technology (CAICT). According to her, CAICT is studying common methods and a compliance assessment framework for SI agents across four dimensions — access rights management, behavioural boundaries, personal data protection and content safety — and is developing standards in parallel. Some of the "red lines" — preventing agents from operating outside the authority granted to them, rejecting "algorithmic exploitation" and guaranteeing ultimate user control — could become mandatory over time. On November 4, CAICT and 20 organisations, including Huawei, Tencent, Ant, Alibaba and Douyin, signed the "Declaration on Self-Regulation of SI Agent Development".

Developers get automatic switching across 80+ providers without changing code, but their own provider keys no longer work for language models.

Hugging Face moves language model calls to the OpenRouter router

Developers get automatic switching across 80+ providers without changing code, but their own provider keys no longer work for language models.

Hugging Face has replaced its own Inference Providers router with OpenRouter's system. Requests to language models are now routed among more than 80 providers based on live metrics of price, speed and reliability, and on failure are retried with another provider. Code and the Hugging Face token stay the same; OpenAI Chat Completions/Responses API and Anthropic Messages API are supported, including Claude Code and Codex. Payment for language calls now goes only through the Hugging Face account or organization, and users' own provider keys no longer work for these models. Image, video and audio generation is still handled directly through Inference Providers. According to OpenRouter's measurements, the new routing cut capacity errors by 46%, improved time to first token by 16.9% and increased throughput by 16.3%. The share of open-weight models in tokens on OpenRouter grew from 26.8% in October 2025 to 71.7% by the end of September 2026. The weekly volume of such requests grew more than 60-fold.

Users in Europe can now plan trips and check place availability in Google Maps, while agentic booking with a call to the venue is initially US-only.

Google launches Ask Maps with Gemini in more than 30 European countries plus a trip planner

Users in Europe can now plan trips and check place availability in Google Maps, while agentic booking with a call to the venue is initially US-only.

On November 16 Google announced that Ask Maps — the conversational layer built on Gemini in Google Maps — is rolling out to more than 30 European countries, including Spain, Germany and the UK, with coverage reaching more than 200 countries and territories. Day-by-day trip planning has begun rolling out on Android and iOS: the route is built with numbered stops, showing traffic, opening hours, details of the next stop, weather and parking; the plan can be changed in conversation and exported to Google Docs on desktop. Search across saved lists checks availability, while agentic booking with a call to the venue and along-route suggestions launch first in the US. The service uses data on more than 300 million places and reviews from more than 500 million contributors.

Small businesses in Singapore have a browser tool for quickly drafting a product video, but 94% of those polled by YouGov for Meta report a lack of guidance on SI.

Meta unveils Dream Render, a browser-based service for drafting a product video from a description or photo

Small businesses in Singapore have a browser tool for quickly drafting a product video, but 94% of those polled by YouGov for Meta report a lack of guidance on SI.

On November 16, 2026, at Meta Upskill Day 2026 in Singapore, Meta unveiled Dream Render, a browser-based SI app (dreamrender.com) that turns a short description or product photo into a draft product video in minutes, with editing of shots, voiceover and captions and SI labelling; sign-in is via a Facebook page or a professional Instagram account. At the same event Meta released a YouGov online survey it commissioned: 3,983 small business owners in 9 Asia-Pacific markets were polled from August 6 to September 2, 2026, including 501 in Singapore. Among the Singapore respondents, 94% agree that small businesses need more guidance on SI, 96% plan to use SI for marketing and content in the next 12 months, and 82% of those already using SI reported a rise in productivity.

At the upcoming Suneung, candidates will have to undergo glasses inspections and hand over smart devices: warnings about cancellation of results and criminal liability will appear at entrances, and retaking the exam is proposed to be banned for three to ten years.

South Korea to inspect every CSAT candidate's glasses and study cutting public Wi-Fi at test sites

At the upcoming Suneung, candidates will have to undergo glasses inspections and hand over smart devices: warnings about cancellation of results and criminal liability will appear at entrances, and retaking the exam is proposed to be banned for three to ten years.

On November 16, South Korean Prime Minister Han Seong-sook held an interagency meeting in Seoul and announced the creation of a government-wide emergency response team ahead of the Suneung (CSAT) exam, which will be held on November 19 for around 550,000 candidates; she demanded that all smart devices be handed in, that their presence be checked during the exam, and instructed the Ministry of Science and ICT to study shutting off public Wi-Fi at all test sites. Education Minister Choi Kyo-jin told a parliamentary hearing that at all 1,312 sites every candidate's glasses would be inspected in every section, with at least one specially trained staff member per site, a hotline running between the centre, the regions and the ministry, and warnings at entrances about cancellation of results and criminal liability; he rejected postponing the exam. The ministry is working with parliament to extend the ban on retaking the exam from one year to 3–10 years and is considering radio detectors for next year; further measures were promised for November 17. The trigger was the November 15 admission by lawyer Han Won-seok that he cheated on the tax consultants' exam using a microcamera and generative SI; he handed the microcamera, laptop and smartphone over to police.

A mandatory weekly course on SI and digital safety will be introduced for all students in grades 5–11; in parallel, the Ministry of Digital Development plans to tighten age identification at platform registration, leaving the choice of mechanism to parents.

Russia's Ministry of Digital Development announces mandatory SI and digital safety course for grades 5–11

A mandatory weekly course on SI and digital safety will be introduced for all students in grades 5–11; in parallel, the Ministry of Digital Development plans to tighten age identification at platform registration, leaving the choice of mechanism to parents.

Starting this year, schools are launching a mandatory extracurricular course on superintelligence and digital safety — one lesson a week for students in grades 5–11, Digital Development Minister Maksut Shadayev said at the International Ombudsman Forum in Moscow (November 16–18). According to him, the decision was taken by the president; the course is meant to give schoolchildren the skills to use digital services safely. At the same time, the Ministry intends to strengthen age identification and parental control when registering on platforms, leaving the choice of the specific mechanism to parents.

OpenAI's attempt to buy Goodfire did not end in a deal: Goodfire remains an independent seller of the technology to several SI developers.

The Information: OpenAI offered over $5 billion for Goodfire this summer, but the deal fell through

OpenAI's attempt to buy Goodfire did not end in a deal: Goodfire remains an independent seller of the technology to several SI developers.

According to The Information, the outlet reported on November 16 that in the summer of 2026 OpenAI, including CEO Sam Altman, held several rounds of talks on buying Goodfire, a mechanistic interpretability startup in which Anthropic has invested. One of OpenAI's top executives offered more than $5 billion or an alternative commercial deal — several times Goodfire's February private valuation of $1.25 billion. The deal did not happen: Goodfire's management judged that selling the technology independently to several SI developers would be more profitable. The talks were prompted by a rise in security incidents, including the Hugging Face breach linked to OpenAI's technology; Goodfire's products, including Silico, allow observing a model's internal computations instead of reading chain-of-thought.

What is on record now is the EU's stated willingness to discuss global SI governance with the US Congress and plans to push the issue through the G7, the G20 and the UN; there are still no agreed international rules.

FT: EU revives push for global SI governance after US midterms

What is on record now is the EU's stated willingness to discuss global SI governance with the US Congress and plans to push the issue through the G7, the G20 and the UN; there are still no agreed international rules.

According to the Financial Times, the outlet reported on November 16, 2026 that after the US midterm elections, in which Democrats won control of both chambers of Congress, the European Union revived efforts toward an international agreement on SI governance. European Commission President Ursula von der Leyen said cooperation with Congress would be valuable and that there is an opportunity to discuss advancing global SI governance, though the goal is "difficult and takes time." European Commissioner Henna Virkkunen said the EU will push the issue through the G7, the G20 and the UN, and is gathering incident data from OpenAI and Anthropic and may open investigations under the AI Act.

Corma raised a $125 million Series A at a $1 billion valuation and increased server capacity 24-fold to meet growing demand from Fortune 100/500 and Fortune Global 500 companies.

Corma announces $125 million Series A at a $1 billion valuation

Corma raised a $125 million Series A at a $1 billion valuation and increased server capacity 24-fold to meet growing demand from Fortune 100/500 and Fortune Global 500 companies.

Corma (Tel Aviv and San Francisco), which develops its own foundation models for defensive cybersecurity, announced on November 16 a $125 million Series A at a $1 billion valuation; the round was led by NEA, with participation from Sequoia Capital, Khosla Ventures and Coatue. Total capital raised reached $185 million three months after emerging from stealth with a $60 million seed, and amid growing demand from Fortune 100/500 and Fortune Global 500 companies the company increased server capacity 24-fold. ARR, by Globes' estimate, is in the tens of millions of dollars; the CEO is Alon Pluda.

SK hynix is taking on part of the suppliers' risk: the program provides an advance of up to 50% of development costs and payment for the work even if goals are not met, directly lowering the barrier for Korean suppliers of materials and equipment.

SK hynix announces 1.4 trillion won ($1.05 billion) five-year program to support Korean suppliers of materials, components and equipment

SK hynix is taking on part of the suppliers' risk: the program provides an advance of up to 50% of development costs and payment for the work even if goals are not met, directly lowering the barrier for Korean suppliers of materials and equipment.

On November 16 SK hynix announced a program worth 1.4 trillion won ($1.05 billion) over five years to support South Korean suppliers of materials, components and equipment — from R&D and testing to mass production and sales. The company also disclosed the Trinity Fab test center worth 874.6 billion won at its Yongin campus (SK hynix — 443.2 billion, the central government — 373.6 billion, local authorities — 57.7 billion); opening is scheduled for May 2027, with a 3,305 sq m clean room inside for evaluating materials on 12-inch wafers. The program provides an advance of up to 50% of initial development costs and compensation for the work even if goals are not met, and, together with financial institutions, 2 trillion won in financing under a Korea Trade Insurance Corporation guarantee. On November 16 SK hynix showed journalists 51 production tools in Trinity Fab and, according to East Asia Brief, plans to support about 300 suppliers by 2027.

An unverified citation of a nonexistent court ruling led to the cassation appeal being returned; the court restored the deadline only after the defects were remedied and held that checking citations remains the responsibility of the parties to the proceedings.

Ukraine's Supreme Court returns lawyer's cassation appeal that cited four nonexistent court rulings

An unverified citation of a nonexistent court ruling led to the cassation appeal being returned; the court restored the deadline only after the defects were remedied and held that checking citations remains the responsibility of the parties to the proceedings.

By a ruling dated 26 March 2026 in case No. 523/9632/24 (a dispute over a father's participation in raising his son), Ukraine's Supreme Court returned a cassation appeal: the lawyer cited four rulings (No. 522/19197/15, No. 761/26645/16-с, No. 753/12866/17, No. 686/15262/17) that were never issued by either the Supreme Court or Ukraine's Supreme Court and concerned other categories of disputes. The court said it had no objection to the use of SI tools in preparing procedural documents, but placed on participants the duty to verify that decisions exist, that case numbers are correct and that legal conclusions match the circumstances of the dispute. After the defects were remedied, the lawyer filed the appeal again, and the court restored the deadline and opened cassation proceedings.

Sources 1

The agreements recorded a broadening agenda: a memorandum on scientific and technical cooperation covering SI and data analysis was signed.

Lee Jae-myung and Marcos agree at ASEAN summit to expand defense, SI and shipbuilding cooperation

The agreements recorded a broadening agenda: a memorandum on scientific and technical cooperation covering SI and data analysis was signed.

On November 16, 2026, on the sidelines of the 49th ASEAN Summit in Manila, South Korean President Lee Jae-myung and Philippine President Ferdinand Marcos Jr. held their third bilateral meeting. According to South Korean presidential spokesperson Kim Sung-wan, the sides agreed to deepen cooperation in the defense industry and infrastructure and to extend it to shipbuilding, SI, nuclear energy, space and disaster response. The Philippine Presidential Office said cooperation in SI and cybersecurity was discussed, and an agreement on grant aid for KOICA and a memorandum on scientific and technical cooperation covering, among other things, SI and data analysis were signed; two-way trade in January–September 2026 totaled $21.3 billion.

A small team can keep on its desk a cluster that runs a 700-billion-parameter model locally, with no cloud and no network switch.

NVIDIA puts four-node DGX Spark kit with 512 GB of memory on sale

A small team can keep on its desk a cluster that runs a 700-billion-parameter model locally, with no cloud and no network switch.

At the SC26 conference, NVIDIA unveiled the DGX Spark HPC Developer Kit and opened sales. The kit is four DGX Spark nodes on GB10 Grace Blackwell, connected directly in a ring via ConnectX-7, with no network switch. The kit provides 512 GB of total memory and 20 TB of NVMe. With NVFP4 quantization, models of up to 700 billion parameters can be deployed on it locally; the example in the blog is DeepSeek V4.1. The kit can be bought from OEM partners — Acer, ASUS, Dell, GIGABYTE, H3C, HP, Lenovo, MSI and xFusion — and on the NVIDIA Marketplace.

A research group of three can get paid access to ChatGPT Plus by applying on November 23: allocation is first-come, first-served rather than by lottery.

OpenAI to expand ChatGPT program for scientists to 200,000 people

A research group of three can get paid access to ChatGPT Plus by applying on November 23: allocation is first-come, first-served rather than by lottery.

According to OpenAI, the company is expanding its ChatGPT for Academic Researchers program: the next round of 50 thousand grants opens on November 23, 2026, and the stated goal is up to 200 thousand researchers. The grant gives ChatGPT Plus to the recipient and up to two co-authors (the summer pilot handed out Pro), so the round could cover up to 150 thousand people. Applications are now reviewed on a first-come, first-served basis instead of a lottery; the pilot involved 20 thousand researchers from 33 countries. At the same time, OpenAI published Economic Research, a study built on 580,186 messages from 1,819 scientists. Of the 250 who responded, 98% expect discoveries in their field to accelerate. Participants estimated that the same work without SI would have taken them 1.9 times longer — this is a self-assessment, not an independent measurement.

The £1 million programme is designed for up to 25,000 young people over 18 months, prioritising NEETs, and ties training to recognised qualifications.

OpenAI and The King's Trust announce £1 million partnership to train up to 25,000 young Britons

The £1 million programme is designed for up to 25,000 young people over 18 months, prioritising NEETs, and ties training to recognised qualifications.

According to OpenAI, on November 16 OpenAI and the British youth charity The King's Trust announced a partnership worth £1 million. Over 18 months the plan is to train up to 25,000 young people across the UK in SI skills, with priority for those not in education, employment or training (NEET). The programme combines hands-on training, employment support and nationally recognised SI qualifications; the initiative is part of a series of events marking The King's Trust's 50th anniversary.

For now Alice AI only lets users view and analyze a brokerage account; trading is not enabled, although about 5% of users already make trades through third-party SI agents.

T-Bank connects T-Investments brokerage accounts to Alice AI via MCP for portfolio analysis only

For now Alice AI only lets users view and analyze a brokerage account; trading is not enabled, although about 5% of users already make trades through third-party SI agents.

T-Bank said T-Investments clients can analyze their portfolios through the T-Investments plugin in the open Alice AI catalog from Yandex, with accounts connected via the MCP protocol. At the first stage only informational and analytical functions are available — the value of the investments, each asset's contribution to the result, market conditions; trading operations are promised later. According to the bank, a third of clients who have connected SI agents use the service on almost every trading day, totaling more than 14 million requests a month, while about 5% make real trades through agents. In August 2026 T-Investments had already allowed accounts to be connected to external SI agents, including ChatGPT and Claude.

European defense forces are buying ANAFI UKR under 1–3 year programs, so Parrot expects micro-drone revenue of more than €20 million as early as Q4 2026.

Parrot announces over €10 million in orders for ANAFI UKR reconnaissance micro-drones, with deliveries in Q4 2026

European defense forces are buying ANAFI UKR under 1–3 year programs, so Parrot expects micro-drone revenue of more than €20 million as early as Q4 2026.

On November 16 Parrot announced that new and additional orders for ANAFI UKR reconnaissance micro-drones worth more than €10 million in total are being fulfilled in Q4 2026. These are procurement programs for European defense forces received in 2026 with terms of 1 to 3 years: some are previously undisclosed contracts with European defense customers, others are additional orders through NSPA. According to Parrot's description, the under-1 kg device is designed for reconnaissance, surveillance and target acquisition, and uses onboard SI and a camera when GPS is jammed; the recipient countries were not named. The company expects micro-drone revenue above €20 million in Q4 2026, said Parrot senior vice president and chief commercial officer Chris Roberts.

The shortage of computing capacity has already translated into quotas for DeepSeek's free users, while the paid, pay-as-you-go API has not been affected by these restrictions.

Users report new DeepSeek limits: roughly 40,000 characters per request and a daily message quota

The shortage of computing capacity has already translated into quotas for DeepSeek's free users, while the paid, pay-as-you-go API has not been affected by these restrictions.

On November 16, users of the free DeepSeek app and web version reported that conversations with no activity for more than 7 days can no longer be continued, the maximum size of a single message has been cut to roughly 40,000 characters, and that once the daily quota is exhausted a message saying "usage quota exhausted" appears. DeepSeek attributed the restrictions to a shortage of computing capacity but has not issued public rules, did not disclose the size of the quota, and its status page showed normal operation on November 16. According to OpenRouter, in the week of November 9–15 request volume to V4.1-Flash-1017 reached 59.2 trillion tokens, the model ranks first in the world, and at least eight outages have been recorded since October. The restrictions do not affect the API, where separate limits and pay-as-you-go pricing apply.

A memristor has been produced with switching in less than 10 nanoseconds and stability over tens of thousands of cycles, which reduces the spread of characteristics between elements.

MIPT and Kurchatov Institute physicists create parylene memristors with ZnTe nanoparticles and predictable switching

A memristor has been produced with switching in less than 10 nanoseconds and stability over tens of thousands of cycles, which reduces the spread of characteristics between elements.

MIPT's press service said physicists from MIPT and the Kurchatov Institute research center have developed memristors based on the parylene polymer with zinc telluride nanoparticles: the nanoparticles make the formation of conductive paths orderly, so switching becomes predictable and repeatable. The devices switch in less than 10 nanoseconds, consume little energy and retain their properties for tens of thousands of cycles, which reduces the spread of characteristics between individual elements. According to Andrey Emelyanov, an associate professor at MIPT, this brings the creation of economical SI systems for processing visual information closer.

Dreamteam's SI CRM is already replacing classic Salesforce and HubSpot at some customers, not just running pilots.

Dreamteam raises $53M for SI CRM that has replaced Salesforce and HubSpot at some clients

Dreamteam's SI CRM is already replacing classic Salesforce and HubSpot at some customers, not just running pilots.

According to Bloomberg, Dreamteam, a startup founded by Freshworks alumni, has raised $53 million in two rounds: a seed round led by Together Fund and a Series A led by Accel, with Premji Invest participating in both; the size of the individual rounds and the valuation were not disclosed. The product is built around "Company Brain" — automatic collection of context from emails and conversations — and four SI agents that prepare meetings, write follow-ups, monitor the deal pipeline and set up processes. In three months of active sales the company has signed more than 100 customers in 13 countries, in some cases replacing Salesforce (DMS Health) and HubSpot (including at Plex); headcount stands at more than 35 people, with the engineering team in Chennai.

Taisei has gained a ten-year vehicle for investing in SI and robotics worth $150 million, managed by Pegasus Tech Ventures.

Taisei announces ¥24 billion ($150 million) fund to invest in SI and robotics

Taisei has gained a ten-year vehicle for investing in SI and robotics worth $150 million, managed by Pegasus Tech Ventures.

On November 16, Taisei Corp. announced the creation of a corporate venture fund of about ¥24 billion ($150 million) together with US-based Pegasus Tech Ventures, which will become the general partner; the fund's term is 10 years, from November 2026 to November 2036. The target areas are startups in SI, robotics and digital twins, as well as new business models; Taisei participates as a limited partner. Taisei President Yoshiro Aikawa said the fund is needed to strengthen existing businesses and create new opportunities; the company explains its interest in automation by the labour shortage in Japanese construction.

An agent written on any platform can be connected via the OpenTelemetry Protocol, and its traces, environments and evaluations can be seen in LangSmith without rewriting it for LangChain's tools.

LangChain releases rebuilt LangSmith with a unified agent record

An agent written on any platform can be connected via the OpenTelemetry Protocol, and its traces, environments and evaluations can be seen in LangSmith without rewriting it for LangChain's tools.

LangChain has opened global access to a new version of LangSmith — in the cloud, as a BYOC option and on self-hosting. The product is rebuilt around a single "agent record". Traces, development, staging and production environments (dev/staging/prod), quality evaluations, Agent Server and Engine incident analysis are all tied to one agent ID. The agent type — general purpose or coding — sets the default metric set; an agent can be connected from any platform via the OpenTelemetry Protocol. Each record is assigned an owner, ID and type. Voice agents and organization-level dashboards are announced as future features. The company reported its Series C round and Agent Platform a week earlier; availability of the reworked product was announced on November 16.

A team can attach the plugin to an already running coding agent and cut its token bill by roughly a fifth without changing anything in the model itself.

TokenEconomizer plugin cuts coding agent costs by 18–31%

A team can attach the plugin to an already running coding agent and cut its token bill by roughly a fifth without changing anything in the model itself.

Microsoft Research, together with UC Berkeley and Microsoft Security, published the paper TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents and an open plugin that turns logs from completed agent sessions into hints for subsequent sessions. The plugin has three mechanisms: micro-skills specific to a given repository, disabling unused tools, and intervention on repeated errors. The model and the agent stay the same — the plugin connects through the agent's existing interfaces. On 277 SWE-Bench tasks from six repositories and five agent/model combinations, average cost fell by 18.4–30.9% with no loss of quality. On Web-Bench the reduction was 16.0–21.2%.

With no national framework in place, about three-quarters of New Zealand schools have no planned approach to using SI, according to ERO.

New Zealand teachers' union demands national SI rules for schools

With no national framework in place, about three-quarters of New Zealand schools have no planned approach to using SI, according to ERO.

On November 16, The Post/Waikato Times reported that the Post Primary Teachers' Association has demanded a nationwide framework for the use of SI in schools. The trigger was a roadshow by the Ministry of Education, where attendees were advised to use Copilot to work through the new curriculum, and the slides were reportedly generated by SI: with a garbled slogan "kia curious" and a map on which Taranaki ended up at the level of Wellington. PPTA President Chris Abercrombie said schools are left to decide for themselves how to use SI, whereas Australia has a detailed framework in place; acting deputy secretary of the Curriculum Centre Antony Harvey responded that the ministry's guidance requires SI materials to be checked. According to a July ERO report, 93% of school leaders, more than 80% of teachers and three-quarters of students use SI, while about three-quarters of schools have no planned approach.

Users can pick the voice variant for their region so Gemini's spoken answers sound closer to the pronunciation they are used to — Indian English instead of American, for example.

Google rolls out 66 regional Gemini voices on Android and iOS

Users can pick the voice variant for their region so Gemini's spoken answers sound closer to the pronunciation they are used to — Indian English instead of American, for example.

Google has begun rolling out 66 new regional voices in the Gemini app on Android and iOS — separate variants for each language. The set includes English for the US, UK and India, Hindi, Brazilian Portuguese, Indonesian and Korean. The selected voice applies across the whole app: in Gemini Live and when reading answers aloud in chat. It can be chosen in the Gemini Live menu or in settings. The company links the rollout to plans announced at Google I/O.

Sources 1