Friday, 20 November 2026
The last 24 hours' hottest news in AI, sorted by field, each story with a line on why it matters.
Story of the day
The single most consequential story of the day.

- T.Capital investment
- €100M
- Mistral AI round
- Series D
- Assets managed by T.Capital
- >€2B
- Altair eRedCap deploys (separate item)
- 2027
Deutsche Telekom invests €100M in Mistral AI for ‘sovereign AI’ push
Deutsche Telekom’s T.Capital Tech Fund put €100 million into Mistral AI’s Series D as a first step toward a strategic partnership on sovereign AI in Germany. The plan pairs Telekom’s infrastructure and customers with Mistral’s models and software.
Why it's the story of the dayA major telco is backing European-built AI on local infrastructure for regulated sectors.
Startup of the day
A young company people are talking about today.
Disseqt AI raises $8M for enterprise AI agent governance
Disseqt AI raised $8 million led by Gresham House Ventures, with angels including Rahul Dravid. The startup tests and governs enterprise AI agents, enforcing policy guardrails and audit trails across deployments.
Why it's worth a lookCompanies need independent tools to prove their AI agents are safe and compliant before going live.
AI funding news (Google News) ↗Where the industry gathered
The conferences, keynotes and festivals where people met in person.

AI Summit 2027: A Conversation Before Geneva
Swiss AI experts gathered in the Bay Area to gather views ahead of the 2027 Geneva AI Summit.
- Hugging Face's Irene Solaiman shared her view on AI's opportunities, risks and policy choices.
- Attendees joined one of six roundtables, from AI safety to AI for science, to feed into the Geneva talks.

APEC CEO Summit
It brought together business leaders from Asia-Pacific economies during the APEC summit week.
- Xi Jinping called on Asia-Pacific economies to deepen coordination on AI and the digital economy.
- Xi pointed to unicorn companies and large AI models as signs of progress.
Models & labs
New models, the labs behind them, and the rules they answer to.
Deutsche Telekom puts €100M into Mistral AI for sovereign AI
Deutsche Telekom’s T.Capital invested €100M in Mistral AI’s Series D to kick off a strategic partnership. The aim is trusted, locally run AI for German businesses and public institutions.
Why it mattersIt backs European AI models deployed on locally controlled infrastructure.
TelecomLead ↗Amazon details AI tools to build math proofs without ‘hallucinations’
AWS released AutoCorrode tools that link AI to the Isabelle proof assistant via MCP for checked proofs. Using Isabelle Assistant, Claude Opus 4.6 produced readable, verified proofs for classic results.
Why it mattersTying models to formal checkers can curb made-up answers in technical fields.
amazon.science ↗OpenAI launches ChatGPT for Sales with 30+ partner plugins
ChatGPT for Sales adds a Sales plugin with ready-made workflows, from picking accounts to drafting follow-ups. It also connects to more than 30 new and expanded partner plugins, including Salesforce, HubSpot and Gong.
Why it mattersSales teams can run much of their daily work from one AI assistant linked to the tools they already use.
OpenAI ↗OpenAI widens its free ChatGPT programme for researchers to 200,000
The ChatGPT for Academic Researchers programme will give up to 200,000 researchers free access. Applications for the next round open on 23 November, with 50,000 new grants awarded first come, first served.
Why it mattersScientists in many countries can use top AI tools at no cost.
OpenAI ↗Amazon SageMaker adds full fine-tuning for custom AI models
SageMaker AI Serverless Model Customization now supports full fine-tuning, which updates all of a model's settings. Amazon's own tests show it beats the lighter LoRA method on complex tasks, with no servers to set up.
Why it mattersBusinesses can tailor AI to their own data without managing hardware.
AWS ↗Gemini on Mac can now run multi-step tasks on your computer
Computer use in the Gemini app for Mac works across local files, desktop apps and websites in a separate window. You describe the task, and Gemini clicks and types while you approve each key action.
Why it mattersRoutine clicking, typing and form-filling can be handed off, with you still approving each step.
Google ↗Microsoft and Salesforce join forces on Copilotforce
Salesforce customer data comes into Microsoft Copilot, Teams and Fabric, with a sales plugin offering 38+ prebuilt skills in public preview. Salesforce keeps control of access and business actions.
Why it mattersSellers can act on customer data without switching between apps.
Microsoft ↗879,000
The US approved more than 879,000 Nvidia and AMD AI chips for Chinese buyers.
Those buyers asked for 3.4 million chips, so roughly one in four requests was approved. CASRAI ↗
Agents & coding
AI that acts on its own, and the tools developers build with.
Devin Bots launch to run projects and lead multiple agents
Cognition introduced Devin Bots, persistent agents with memory and tools that manage teams of Devins. They investigate, delegate, and evaluate to own outcomes like releases or repo maintenance.
Why it mattersAgents are moving from single tasks to managing ongoing engineering goals.
devin.ai ↗Cognition's Devin takes on data platform migrations for Databricks
Devin converts legacy data-warehouse code, tests it and checks results against the new copy before switching over. It can run hundreds of tasks in parallel, and one Fortune 500 energy company moved a 142-object platform in eight days.
Why it mattersData moves that used to tie up engineers for months may now take weeks.
Devin ↗Devin Security adds a feed of new software vulnerabilities
Devin Security Threats tracks newly disclosed flaws, including Java4Shell and NGINX2Root, and scans a company's code for them in one click. Devin can then open pull requests with fixes for teams to review.
Why it mattersCompanies get a faster way to check whether new flaws affect their own software.
Devin ↗Perplexity launches Orchestrator API with automatic failover
Orchestrator lets developers reach models from Anthropic, OpenAI, Google and others through one API. If a provider fails, it retries the same model elsewhere, succeeding more often than any single provider, Perplexity says.
Why it mattersDevelopers can switch providers when one fails without changing which model they use.
Unite.AI ↗Unsiloed AI raises $7.5M to make documents readable by AI agents
The San Francisco startup sells an API that turns PDFs, images and spreadsheets into data AI agents can read and check. Felicis led its $7.5 million seed round.
Why it mattersFinance, law and healthcare run on documents, so this could open them up to AI agents.
Wowtale ↗Cline-Bench tests coding agents on real user requests
The new benchmark has 47 coding tasks, each with a working starting repository and tests built to stop agents gaming scores. Gemini 4 Argon led with a 51.1% success rate on the first attempt.
Why it mattersFairer tests make it easier to compare which coding agents actually do the work.
Cline ↗UK regulators set out expectations for personal AI agents
Four UK regulators, the FCA, CMA, ICO and Ofcom, wrote to ministers on 16 November about personal AI agents. Firms must engage early and meet their legal obligations before launch.
Why it mattersThe regulators set expectations that firms engage early and meet obligations before launch.
Global Regulation Tomorrow ↗AI agents can mimic human choices, flaws included, study finds
Northeastern researchers showed that AI models reproduce human decision-making more closely when they account for how people really choose. The Nature Computational Science study was published on 16 November.
Why it mattersSimulating people with AI could let planners test changes like fares or evacuation routes before trying them for real.
Northeastern Global News ↗We built Devin to own software engineering projects end to end. Migrations are a great example where agents need to understand the existing code, make the changes, and test that everything works,
Video generation
The models and tools that make video.
Higgsfield MCP moves clients to a new sign-in method by 2027
Clients that register through DCR must switch to CIMD, which uses a web address the developer controls, before 1 January 2027. Clients built by others need no changes on the Higgsfield side.
Why it mattersDevelopers linking AI tools to Higgsfield have a deadline to update their sign-in setup.
Higgsfield ↗Higgsfield launches Short Drama Studio for script-to-series video
The tool splits a script into episodes, builds one consistent cast and generates the video in three stages. It can run step by step or fully automatically, though the automatic route costs more.
Why it mattersWriters could turn a script into a short serial drama without a film crew.
Higgsfield ↗Pika Enact lets creators direct AI video live on set
Pika Enact is a research preview built on Pika H3 Flex that generates video in sub-second chunks. Creators can drag actors, speak to them and move the camera while the scene keeps playing.
Why it mattersDirecting AI video becomes interactive, closer to how films are shot.
Pika ↗Hedra steers a world model to follow multi-machine instructions
Hedra's mINT method finds the parts of MiniMax H3 that control behaviour and adjusts them without retraining. Success on multi-machine instructions rose from 15.2% to 77.7%.
Why it mattersFixing a model's behaviour without costly retraining could make robot simulations more reliable.
Hedra ↗HeyGen's Avatar 6 makes 2K talking videos of unlimited length
Avatar 6 creates talking videos from a single photo at 2880×1620 with no fixed time limit, with clips of 25 to 28 minutes. In HeyGen's own test, 84.4% of side-by-side answers judged it indistinguishable from real footage.
Why it mattersLifelike AI presenters are becoming long enough for full talks and podcasts.
HeyGen ↗Kuaishou's Q3 adjusted profit falls 84% as Kling AI spending rises
Adjusted net profit fell 84.2% to CNY 790 million, and revenue dropped 11.1%. Research spending rose 41.9%, driven by investment in Kling AI, and Kuaishou expects pressure to continue into the fourth quarter.
Why it mattersThe company behind Kling is spending heavily on AI video while its core business slows.
BigGo Finance ↗A2E AI launches an all-in-one AI video generator
A2E AI launched a browser-based video generator with 225 video and image models, including Kling 4.0 and Veo, under one account. Users pay per run with credits instead of a monthly subscription.
Why it mattersCreators can try several top video models without paying for a separate subscription to each.
GlobeNewswire ↗US lawmakers question Microsoft's reported $500M ByteDance deal
A reported $500 million Azure deal could give ByteDance access to OpenAI models through Microsoft. The House China committee has asked Microsoft for a briefing, and its chairman says Microsoft 'should not sell AI tools to ByteDance'.
Why it mattersOpenAI bars most Chinese firms, but Microsoft can sell its models without those limits, which raises a loophole question.
Firstpost ↗Video creators today pay for several subscriptions just to try the newest models. With A2E AI, they choose the model that fits the shot, see what it costs before they press Run, and pay only for what they generate,
Creator tools
The apps and platforms people make video with.
ElevenLabs acquires Japanese voice AI lab Kotoba
ElevenLabs is buying Kotoba, a Japanese voice AI lab founded in 2023 by Noriyuki Kojima and Jungo Kasai. The aim is to bring voice AI that fits Japanese language and social norms to AI agents in Japan.
Why it mattersVoice assistants that handle Japanese manners and speech patterns could reach more people in Japan.
elevenlabs.io ↗Canva adds a switch to turn off its generative AI features
Canva now has a toggle that disables tools such as Magic Expand, Magic Write and Dream Lab. Background remover and Magic Grab stay on, and a separate setting shows human-made content first in the library.
Why it mattersUsers who want no AI in their designs can now switch most of it off.
The Verge AI ↗Cursor says a security breach reached its internal systems
Cursor says an attacker used a key exposed at Fireworks to copy old coding agent traces, then reached some internal tools. At most 5,000 users may have received a malicious copy of Anthropic's Claude Code extension.
Why it mattersDevelopers who use AI coding tools may need to check which credentials they shared with vendors.
Cursor ↗Hume releases a model that predicts how people rate audio
Hume Audio Decision estimates how a group of listeners would judge a voice clip, such as its naturalness, without a transcript. Hume says it scored highest across 11 models on four public human-preference benchmarks.
Why it mattersTeams can test how human a voice sounds in about a quarter of a second, instead of running a slow listening test.
hume.ai ↗Cerence sues ElevenLabs over five voice AI patents
Cerence filed a patent complaint in Delaware federal court against ElevenLabs over five speech technology patents. Cerence says it warned ElevenLabs in August and ElevenLabs kept selling its products.
Why it mattersThe case could affect how voice AI tools are built and licensed in the United States.
Wisevoter ↗Pony.ai robotaxi revenue rises 165% in third quarter
Pony.ai reported record third-quarter revenue of RMB 298.8 million, up 75.2%. Robotaxi revenue reached RMB 119.1 million, and all three Gen-7 models are now in fully driverless commercial operation.
Why it mattersDriverless taxis are moving from test runs into paid service in China and overseas.
Gasgoo ↗Suno launches Visuals for lyric videos and album art
Suno's new Visuals tool makes AI images and clips of up to 15 seconds for songs in a user's library, on web, iOS and Android. Visual credits are sold separately from song credits, and filters block depictions of specific artists and copyrighted works.
Why it mattersMusic makers can now build a full visual package around a song inside one app, with limits meant to protect artists.
Unite.AI ↗ElevenLabs opens second Chroma Awards season
ElevenLabs opened entries for the second Chroma Awards, open to creators worldwide in film, music videos, advertising and games. Submissions close on 7 January 2027, and the prize pool is over $1M in giveaways and $450,000 in cash and travel perks.
Why it mattersIt is a free route for independent creators to get noticed without going through big studios or festivals.
Wisevoter ↗Does Firefly's 'commercially safe' label cover partner AI models?
At its MAX event, Adobe added models from ByteDance, Google, Kling, OpenAI and ElevenLabs alongside Firefly. Firefly's commercial-use promise covers Adobe's own models, not necessarily these partner models.
Why it mattersDesigners and agencies need to check each model's own terms before using what it makes.
Creative Bloq ↗Higgsfield MCP ends DCR client sign-in on January 1, 2027
Developers whose MCP client registers through DCR must move to CIMD before that date. People who use a client built by someone else need to change nothing on Higgsfield's side.
Check for a POST to the registration endpoint and move the client to CIMD before then. Higgsfield ↗
Media & entertainment
AI in the films, shows, micro-dramas, music and ads people watch, the studios and platforms making them, and the rules for AI-made content.
Spotify removes AI ‘slop’ while planning licensed AI remixes
Spotify said it pulled over 50M AI tracks from recommendations, yet is preparing an AI remix and cover tool with UMG for early 2027. Paying users could make AI remixes with revenue sharing.
Why it mattersPlatforms are curating AI music while building ways to sell it.
WIRED Middle East ↗Margot Robbie's Neighbours scenes to become a microdrama
RoseBerry Media is using an AI workflow called RedSnapper to turn Neighbours episodes into vertical mini-episodes about Donna. Fremantle says a human editor approves each step and no new scenes or synthetic voices are made.
Why it mattersIt shows how old TV libraries can be reused for phone-first series, and how performers feel about it.
AOL ↗More than 80% of Americans want an AI music kill switch, UMG says
Universal Music Group says its newest survey finds over 80% of US consumers want a switch that screens out all AI-generated music. A separate Deezer survey found 41% want AI music filtered out entirely.
Why it mattersListeners want to know what is AI-made, and many want the choice to avoid it.
Digital Music News ↗Will Packer launches June, a microdrama studio for Black audiences
Producer Will Packer and his son Dom are launching June to make and buy vertical video series aimed at Black audiences. NBCUniversal and HarbourView Equity Partners back the new company.
Why it mattersMajor Hollywood names are now investing in microdramas, short vertical series made for phones.
The Hollywood Reporter ↗T-Series and Hungama launch FastTV microdrama platform
Indian music label T-Series and Hungama Digital Media Entertainment have launched FastTV, a microdrama platform. The partners committed INR3 billion ($32 million) and own the venture equally.
Why it mattersIndia's biggest music labels are betting heavily on short vertical drama for phones.
Variety ↗Tencent Music says AI music listening is negligible and falling
On its earnings call, Tencent Music said AI-generated content makes up a negligible and declining share of listening on its platforms. It credited content filtering and legal action against copyright-infringing AI music.
Why it mattersA major streaming service reports that most listeners still choose human-made music.
Digital Music News ↗Microdrama writers at DramaWave unionise with WGA West
Writers at DramaWave, a major microdrama platform, have formed a union with the Writers Guild of America West. They plan to bargain over pay, job security, credits and guardrails on AI in writing.
Why it mattersWriters of short phone dramas are seeking the same protections as film and TV writers, including on AI.
Deadline ↗DramaWave writers win union recognition from the WGA
The 27 full-time writers at the microdrama app DramaWave have won recognition from the Writers Guild of America. It is the first full-time writing staff at a microdrama company to win such recognition, and they plan to bargain over pay, job security, and AI use.
Why it mattersWriting jobs behind one-minute episodes are becoming a union issue as the category grows fast.
K-EnterTech Hub ↗ShareChat's revenue passes ₹1,000 Cr as microdramas grow
Mohalla Tech, parent of ShareChat and Moj, reported FY26 revenue of ₹1,016.27 Cr, up 40.5%, helped by subscriptions from Moj and QuickTV. Its net loss widened to ₹1,978.40 Cr, largely from share-based payments.
Why it mattersThe microdrama app QuickTV is now a growth engine, but the company has yet to prove it can turn a profit.
Inc42 ↗Consumer research indicates […] that the vast majority of consumers demand transparency. They want the content identified. They want the content labeled. An increasing percentage are looking for what we call a kill switch,
Chips & compute
Chips, data centres and the power behind them.
Cerebras debuts disaggregated inference for ultrafast tokens
Cerebras now pairs GPUs for prefill with its wafer-scale systems for decode, claiming 5× more tokens per megawatt at the same speeds. It targets growing agent workloads and power limits.
Why it mattersSplitting inference stages could boost capacity without more power.
cerebras.ai ↗House releases US AI chip licence data on China
House Foreign Affairs Chairman Brian Mast released export licence records showing Chinese buyers requested 3.4 million Nvidia and AMD AI chips. Regulators approved more than 879,000, including sales to firms the Pentagon lists as Chinese military companies.
Why it mattersIt shows how export controls on advanced chips work in practice and which buyers got them.
CASRAI ↗Permitting bill would speed data centres, including on public lands
Opponents tried to add an amendment banning data centres on public lands, but a key Senate vote stalled. Backers of the BAAJA permitting bill say it is essential to the AI data centre boom.
Why it mattersThe bill could decide how fast new data centres can be built, and where.
Center for Western Priorities ↗Figure AI's ties to Nvidia put its compute deal at risk
Nvidia invested in humanoid robot maker Figure AI and is also a shareholder in Nscale, the cloud firm Figure relies on. Figure has committed $3.5 billion to $6 billion on Nvidia Vera Rubin chips.
Why it mattersRobot makers depend heavily on Nvidia money and chips, so shifts at the chipmaker reach them too.
Mike Kalil ↗Study finds data centres are more often built in communities of colour
Researchers at George Washington University and American University examined 81,000 US census tracts and 5,291 data centres. Majority-Black neighbourhoods had 2.2 times the odds of hosting one, compared with white neighbourhoods in the same state.
Why it mattersWhere data centres go affects local residents' power bills, land and pollution.
404 Media ↗A data centre slowdown could help AI developers, argue two analysts
Mark MacCarthy and David Klaus argue the market fall this month was driven mainly by doubts about AI revenue, not new regulation. They say local pushback could stop an oversupply of data centres.
Why it mattersOpinion on whether the data centre boom is building too much capacity too fast.
Tech Policy Press ↗House committee releases AMD and Nvidia China chip export files
The House Foreign Affairs Committee published confidential licence applications from AMD and Nvidia to export advanced AI chips to China. AMD sought approval to ship 600,000 chips and called the risk of diversion 'low' in every application.
Why it mattersChip exports to China are under close scrutiny, and AMD says it complied with US laws.
Hunterbrook ↗Hunterbrook questions AMD's financing of its AI customers
Hunterbrook, which holds a short position in AMD, says the chipmaker guarantees leases and buys stakes in customers, much as Nvidia does. It also gave its two largest AI customers rights to buy up to 16% of the company for a penny a share.
Why it mattersIf AI customers cut spending, AMD could face weak chip sales while still owing large commitments.
Hunterbrook ↗Nvidia publishes Linux tools for display memory on its chips
Nvidia opened a GitHub repository of small Linux utilities for its platforms, starting with one that changes the display memory reserved by firmware. Changes take effect on the next boot.
Why it mattersIt is a small, technical release mainly for developers and system builders running Nvidia hardware on Linux.
GitHub ↗exercised his authorities under the Export Control Reform Act to publicly reveal export control license application information regarding NVIDIA and AMD’s AI chip licenses for the People’s Republic of China.
Policy & safety
Laws, lawsuits and safety incidents.
Bipartisan House caucus forms to craft AI safety laws
Over two dozen lawmakers launched the AI Safety and Innovation Caucus, citing agent incidents at Hugging Face and Anthropic. Co-chairs aim for targeted safety legislation and rapid responses.
Why it mattersCongress is organizing to write rules for autonomous AI risks.
SiliconANGLE ↗California proposes mandatory AI safety standards and shutdowns
Gov. Gavin Newsom proposed state-set AI safety standards, independent testing and emergency shutdown powers. They would also let the state cap recursive self-improvement, where AI designs its own successor.
Why it mattersCalifornia is moving to set AI rules itself while federal oversight is still absent.
Lake County Record-Bee ↗Trump and Xi agree AI needs safety rules
Trump said he and Xi Jinping agree AI needs safety rules and that the two will work together on reasonable ones. China's embassy in Washington echoed the call for cooperation and keeping AI under human control.
Why it mattersThe two biggest AI powers signalling shared safety goals is a rare point of agreement amid a chip rivalry.
The Times of India ↗Beyer outlines AI safety rules; Warner targets AI financing
Rep. Don Beyer outlined bipartisan rules for top AI builders: safety frameworks, incident reports and outside review. Sen. Mark Warner introduced a bill on how public firms finance and oversee AI.
Why it mattersCongress is working on AI from two directions, though Beyer's plan is not yet written as law.
The Alexandria Brief ↗Chinese AI safety group briefs OpenAI staff in San Francisco
Beijing-based Concordia AI briefed OpenAI staff on China's safety approach and on slowing advanced AI together. The US and nine allies also signed the San Francisco Declaration on securing AI supply chains.
Why it mattersSafety talks between US and Chinese researchers show the topic moving into diplomacy.
SFist ↗Trahan revises FRONTIER Act with penalties for near misses
The revised FRONTIER Act would give independent evaluators embedded access to AI companies and give federal officials enforcement powers. Penalties would cover near misses, and the full legislative text is still to come.
Why it mattersIt would punish AI close calls before harm happens, not just actual damage.
Superpower Daily ↗New York creates a unit to enforce AI safety laws
New York's Attorney General has set up a dedicated unit to enforce the state's AI safety laws. The source gives no further detail on its powers or staff.
Why it mattersEnforcement is now a named job in New York, so AI firms face an office that checks compliance.
CDO Magazine ↗Watchdog: top AI chatbots helped plan US attacks in tests
A Government Accountability Office report found ChatGPT, Claude, Gemini and Grok all produced proposals for a terrorist attack when investigators asked. Incoming Senate Homeland Security chair Maggie Hassan plans a hearing on AI risks in January.
Why it mattersGovernment testers were able to get past safety limits, so pressure for AI rules is growing.
Baton Rouge Business Report (AP) ↗AI safety rules may leave businesses liable for agent accidents
Anthropic and OpenAI safety rules take effect this month. If an AI agent causes a factory accident, the company that runs it may carry the blame, not the developer.
Why it mattersBusinesses deploying AI agents should check who carries the risk before the agent goes live.
Bloomberg Law News ↗The federal government’s reckless failure to create any form of meaningful AI oversight or accountability should alarm every American,
Robotics
Robots, self-driving and AI in the physical world.
GlobalFoundries to test Atlas humanoid after scaling Spot
Boston Dynamics and GlobalFoundries will expand Spot across fabs in the US, Europe and Asia. GF will be first chipmaker to test Atlas in an early adopter program starting in 2027.
Why it mattersHumanoids are entering semiconductor fabs after proven mobile robots.
bostondynamics.com ↗New open robot model predicts how long a task has left
RynnValue-8B watches a robot's task video and predicts the seconds left until it is done. It is released under Apache 2.0 in Hugging Face's LeRobot library, but it can only run predictions, not be retrained.
Why it mattersScoring how close a robot is to finishing a job is a building block for robots that learn from practice.
Hugging Face ↗Bristol's smart walking frame uses AI to guide older users
PhD researcher Sergio Sierra Marin co-developed the SAW, a walking frame with sensors, AI and an on-screen companion. It guides users with lights and vibrating handles and brakes before they sit down.
Why it mattersAssistive robots like this could help older people and those with disabilities live independently for longer.
University of Bristol ↗Pony AI reports 2,768 robotaxis as Q3 revenue jumps 75%
Pony AI's robotaxi fleet reached 2,768 vehicles by 30 September, and its third-quarter revenue rose 75.2% to US$44.6 million. Its net loss widened to US$105.6 million, against US$61.3 million a year earlier.
Why it mattersPony AI's numbers are being used as a yardstick for how big Tesla's robotaxi fleet would need to be.
shattered.io ↗CEVA and SIT open a robotics lab for warehouse work in Singapore
CEVA Logistics and the Singapore Institute of Technology opened the ILAS Lab, a five-year project on robots and AI for warehouse tasks such as assembly, labelling and packing. Solutions are expected at CEVA's Blue Hub within one to two years.
Why it mattersWarehouse automation for small, changing batches could make logistics jobs less repetitive.
Container News ↗GM plans 20% more factory automation by 2028
General Motors plans to raise automation across its North American factories by 20% by 2028, focusing on autonomous mobile robots it is building itself at Kokomo, Indiana. Suppliers bidding for contracts must now show how they will automate.
Why it mattersAutomakers are turning to robots to offset higher labour costs ahead of the 2028 union talks.
International Finance ↗Free simulator lets developers test the Asimov 1 robot
A LeRobot release adds a MuJoCo simulation of Menlo's Asimov 1 robot, with its own control protocol and three camera feeds. Developers can walk it with LeRobot's controller before touching the real machine.
Why it mattersRealistic simulators let robot builders test software safely before using hardware.
Hugging Face ↗Free simulator runs many SO-101 robot arms in parallel
The LeRobot tabletop environment runs a batch of simulated SO-101 arms on NVIDIA graphics cards, with the task of putting four cubes in a green box. It is built on MuJoCo Warp and can be used to evaluate robot policies.
Why it mattersFaster simulation lets robot researchers test many attempts quickly before real-world trials.
Hugging Face ↗Webinar explores AI in supply-chain robotics
A Robot Report webinar, scheduled for 18 November, covered how warehouse robots, AI and predictive analytics are changing supply chains. Speakers included the CEOs of inVia Robotics and Gideon and the CTO of Ambi Robotics.
Why it mattersLabour shortages are pushing warehouses toward robots and AI, making supply chains an early test for physical AI.
The Robot Report ↗Microsoft should not sell AI tools to ByteDance,
Science & health
AI in biology, medicine, climate and maths.
FutureHouse launches ‘superintelligent’ agents for scientists
FutureHouse released a platform with agents for literature search, deep reviews, novelty checks, and chemistry planning. It says agents beat frontier models and PhD researchers on some retrieval tasks.
Why it mattersResearch agents promise faster, more accurate scientific reviews.
futurehouse.org ↗Chinese AI lab MegaRobo raises nearly $200M
MegaRobo, a Chinese company combining AI and lab automation, announced a raise of nearly $200 million. The source gives no investor or valuation details.
Why it mattersAutomated labs are attracting big money as AI tries to speed up drug and materials research.
Endpoints News ↗Isomorphic Labs and MSD team up on cancer drug candidates
Isomorphic Labs will design small-molecule candidates for an oncology target with its IsoDDE engine. MSD will run trials and sales, paying milestones and royalties.
Why it mattersPharma giants are paying AI drug designers for molecules, testing whether AI speeds up early discovery.
Securities.io ↗Thyroscope and TORC to track eye disease with smartphone AI
South Korea's Thyroscope and The Orbital Radiology Center signed an MoU to monitor thyroid eye disease using smartphone AI, orbital imaging and clinical data. They will build a shared research database and a longitudinal study.
Why it mattersHome monitoring could spot changes between clinic visits for a condition that can cause vision loss.
BioSpectrum Asia ↗UNC pharmacy school uses AI to speed up drug discovery
UNC's Eshelman School of Pharmacy team used an AI-guided method to find compounds against a key tuberculosis protein in six months. The collaborators boosted one set of compounds' enzyme potency more than 200-fold in a few iterations.
Why it mattersAcademic labs are reaching early drug candidates in months, which could speed work on hard diseases like TB.
UNC Chapel Hill ↗AI models drop climate facts when users push back
UNESCO and UCL tested 15 AI models: they answer 90% of climate questions correctly. But eight in 10 switched answers at the first hint of user disagreement, and are weaker in less common languages.
Why it mattersUsing chatbots to shape climate policy is risky if they simply agree with whoever is asking.
UNESCO ↗Sino Biological launches AI-ready target protein collections
Sino Biological launched two collections for AI-driven discovery: AI-Trusted, with over 960 proteins, and AI-Qualified, with over 1,700 that have binding data. The company offers more than 11,000 target proteins in total.
Why it mattersResearchers need reliable lab proteins to test whether AI-designed molecules actually work.
BioPharma APAC ↗AI models ace the English and maths sections of Korea's CSAT
In a reporter's test, a paid Claude model answered all 74 English and maths questions correctly in under five minutes. Free Perplexity got 71 right, while free ChatGPT scored 32 of 46 in maths.
Why it mattersExams built to test human learning are being cleared by free AI tools in minutes.
Korea Times ↗AI weather forecasting is working without physics equations
AI models learn weather patterns from large datasets instead of solving the physics equations that traditional forecasts use. Google DeepMind's GraphCast, unveiled in 2023, produced a global 10-day forecast in under a minute and beat top physics-based models on several measures.
Why it mattersFaster, cheaper forecasts could reach more places, though scientists still debate how well data-only models handle complex Earth systems.
The Korea Daily ↗AI systems are remarkably good at explaining scientific knowledge, but this can give us a false sense of their capabilities. Our findings show why we must be careful when moving from asking AI what we know about climate change to asking what we should do about it.
Open source
Open-weight models and community releases.
OVHcloud unveils Olala mini 7A1B open MoE model
OVHcloud released Olala mini 7A1B, a 7B-parameter mixture-of-experts with 1B active per token under Apache 2.0. Trained on European supercomputers, it targets CPU-friendly speed and broad language support.
Why it mattersAn open, small MoE aimed at running fast without a GPU.
explainx.ai ↗Reply.io releases open-source harness for AI sales outreach
Jason Harness turns general AI agents such as Claude Code or OpenAI Codex into a sales outreach workspace, with nine roles and campaign records kept on the user's computer. A person must approve every message before it is sent.
Why it mattersAI can draft sales emails now, but a human still has to check them before they go out.
Unite.AI ↗SenseTime releases open-weight Loop-9B model for free
SenseTime released the weights and code for Loop-9B, a model of about nine billion parameters that reads text and images. It runs some of its layers twice for deeper computation and handles up to 512,000 tokens of context.
Why it mattersA model that needs about 15 GiB of memory to run is within reach of many developers.
TechNode ↗Ai2 and Taylor Geospatial team up on open satellite AI
The two groups announced a partnership at SatSummit in St. Louis to make geospatial AI easier to compare, run and build on. Their plans include free benchmarks, the OlmoEarth models and a global map of farm field boundaries.
Why it mattersFree tools for reading satellite images could help conservation, farming and aid groups without large budgets.
Ai2 ↗Palo Alto Networks buys Applied Compute for $4.5B
Palo Alto Networks agreed to buy Applied Compute, a firm that trains and fine-tunes open-weight models for businesses, for $4.5 billion. The deal was disclosed on its earnings call and is meant to help it build AI that supervises other AI.
Why it mattersCybersecurity firms are moving into training their own AI models, which shows how central AI has become to digital defence.
Channel Dive ↗Haru-4B is a free model that talks and listens at once
Higurashi AI released Haru-4B, a research base model that listens and speaks at the same time, as in a phone call. It supports nine languages and can copy a voice from a short reference clip.
Why it mattersModels that talk and listen at once could make voice assistants feel more natural.
Hugging Face ↗US weighs forcing firms to disclose Chinese AI use
The Treasury and the White House are discussing a rule that would make US companies publish which Chinese AI models they use, such as Qwen, Kimi and DeepSeek. Reuters reported it citing three people, and no rule exists yet.
Why it mattersIt would discourage reliance on Chinese models without outright bans, which would affect firms like Airbnb.
Pasquale Pillitteri ↗Uncensored AI models are easy to find on code-hosting sites
Researchers analysed 22,096 uncensored large language models from five hosting platforms, including Hugging Face and ModelScope. Models with safety filters removed are easy to download and are often used inside apps without users knowing.
Why it mattersPeople using AI apps may not know that the model underneath has had its safeguards removed.
The Register ↗Sarvam defends hosting foreign AI models on Token Factory
Sarvam says its Token Factory platform runs on GPUs it bought privately, not the subsidised compute it used under India's IndiaAI Mission, which ended in March 2026. The platform offers six Chinese models alongside its own.
Why it mattersThe row tests how far India's sovereign AI push can go while relying on foreign models.
Moneycontrol.com ↗Tool of the day
Something you can try in the next five minutes.
Jason Harness
Jason Harness turns general-purpose AI agents such as Claude Code or OpenAI Codex into a workspace for sales outreach. It keeps campaign state on your computer and requires a person to approve each message before it goes out.
Quick hits
Smaller stories worth a glance.
- Tencent Hy4 lands on Together AI with 1M-token contextTogether AI added Tencent’s Hy4 on Serverless Inference with image input, tool calling, and pricing from $0.83 per 1M input tokens.together.ai ↗
- DW: AI data centers strain power and waterDW reports data centers could double electricity use to about 945 TWh by 2030, with cities like Frankfurt heavily impacted.DW.com ↗
- Sync adds studio-grade visual dubbing to localizationSync introduced visual speech alignment with batch APIs, zero-shot models, and SDKs for multilingual video workflows.sync.so ↗
- The King's Trust and OpenAI launch £1m youth AI skills partnershipOpenAI is donating £1m so the charity can help up to 25,000 young people in the UK build AI skills for work.OpenAI ↗
- Wall St falls after Anthropic’s record IPO pop coolsAnthropic’s IPO briefly topped $2T in value before settling; shares were up about 15% around $95.50.Korea Times ↗
- Lawmakers seek Microsoft briefing on reported ByteDance AI dealA House panel asked Microsoft to brief on a report it may sell OpenAI models via Azure to ByteDance in a $500M deal.Communications Today ↗
- Devin used for enterprise data migrations to DatabricksCognition announced Devin for Databricks Migrations with Databricks, converting and testing code and reconciling sources.The National Law Review ↗
- Nscale’s IPO path hit by US chip rules; ByteDance is 73% of revenueU.K.BigGo Finance ↗
Deals & funding
Who raised money, and who bought whom.
Funding rounds 5
Investors put money into a company in exchange for a share of it.
MegaRobo raises new funding
- Raised
- nearly $200M
MegaRobo valuation not disclosed
Chinese AI and lab-automation company MegaRobo announced a raise of nearly $200 million.
T.Capital Tech Fund backs Mistral AI
- Raised
- €100M
Mistral AI valuation not disclosed
Deutsche Telekom’s investment arm invested to support a strategic partnership on sovereign AI in Germany.
T-Series backs FastTV
- Raised
- $32M
FastTV valuation not disclosed
T-Series and Hungama committed INR3 billion ($32 million) to launch the FastTV microdrama platform, which they own equally.
HDHungama Digital Media Entertainment+0 more
Felicis backs Unsiloed AI
- Raised
- $7.5M
Unsiloed AI valuation not disclosed
The San Francisco startup raised a $7.5 million seed round to build document tools for AI agents.
NBCUniversal backs June
- Raised
- Undisclosed
June valuation not disclosed
Will and Dom Packer launched June, a microdrama studio backed by NBCUniversal and HarbourView Equity Partners.
HEHarbourView Equity Partners+0 more
Acquisitions 2
One company buys another outright.
Palo Alto Networks buys Applied Compute
- Price
- $4.5B
Applied Compute is now part of Palo Alto Networks
Palo Alto Networks agreed to buy AI development firm Applied Compute for $4.5 billion, to add training and engineering talent.
ElevenLabs buys Kotoba
- Price
- Undisclosed
Kotoba is now part of ElevenLabs
ElevenLabs is acquiring Kotoba, the Japanese voice AI lab, to speed up voice AI for agents in Japan.
Partnerships & contracts 5
A large commercial agreement, or a government contract.
Figure AI and Nscale sign a deal
- Value
- $3.5B to $6B
Multi-year agreement
Figure has committed to spend at least $3.5 billion, and up to $6 billion, on compute with Nscale using Nvidia Vera Rubin chips.
Microsoft and ByteDance sign a deal
- Value
- $500M
Multi-year agreement
A reported $500 million Azure deal would let ByteDance use OpenAI models through Microsoft's cloud, drawing questions from US lawmakers.
The King's Trust and OpenAI sign a deal
- Value
- £1m
Multi-year agreement
OpenAI donates £1m to support AI skills training for up to 25,000 young people in the UK.
· OpenAI ↗
Rakuten Mobile and Mastercard sign a deal
- Value
- Undisclosed
Multi-year agreement
Rakuten's Travel eSIM Platform is integrated with Mastercard Agent Pay for agentic travel payments.
· PYMNTS ↗
Microsoft and Salesforce sign a deal
- Value
- Undisclosed
Multi-year agreement
A partnership to bring Salesforce customer data and actions into Microsoft Copilot, Teams and Fabric.
On the calendar
What's coming up, in date order, and how sure we are about each.
- 20NovAI Film Award submissions closeToday
- 21NovInaugural edition in HollywoodTomorrow
- 23NovChatGPT for Academic Researchers applications open3days away
- 30–3Nov–DecContent London 202610days away
- 1DecClosed-door interview with three House committees11days away
- 1DecHearing on Apple's bid to block testimony11days away
- 4DecAI-made film opens in limited release2weeks away
- 4DecGods Don’t Give Gifts premiere2weeks away
- 4DecSingapore AI Safety Fellowship ends2weeks away
Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.















































.gif)








































