Saturday, 7 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.
- parameters in total
- 594B
- parameters active
- 17B
- Artificial Analysis score
- 31
- weights planned next week
- Apache 2.0
Thinking Machines ships Inkling 2 Small
Thinking Machines released Inkling 2 Small on 5 November, a sparse model with 594 billion parameters and 17 billion active. It scores 31 on the Artificial Analysis index, is live on Tinker, and Apache 2.0 weights are planned for next week.
Why it's the story of the dayA young lab's second model is already aimed at long coding jobs, with the weights promised within days.
Startup of the day
A young company people are talking about today.
Fal raises $1.3 billion at a $16 billion valuation
TechCrunch said fal, which hosts image, video and audio models, raised $1.3 billion led by Bessemer, with Nvidia, Sound Ventures and T. Rowe Price. The round values the company at $16 billion, double the $8 billion figure from April.
Why it's worth a lookThe company that serves generative video just priced itself like a major lab.
TechCrunch ↗Models & labs
New models, the labs behind them, and the rules they answer to.
Thinking Machines ships Inkling 2 Small
Thinking Machines released Inkling 2 Small on 5 November, a sparse model with 594 billion parameters and 17 billion active. It scores 31 on the Artificial Analysis index and is on Tinker now, with Apache 2.0 weights planned next week.
Why it mattersA new lab's second model is already in reach, and the weights are promised within days.
Thinking Machines Lab ↗Thinking Machines nears a $5 billion round
Business Insider reported the lab is raising $5 billion at a $40 billion pre-money valuation, with Accel leading. It previously raised a $2 billion seed. The company declined to comment, and the round is not signed.
Why it mattersThe price is below a $50 billion talk that did not happen, and investors are still in.
Business Insider ↗Meta ships specialized models for Muse
Meta said on 6 November it is rolling out models trained with Muse Wonder to Muse and Meta AI. On its Personal Superintelligence Index the new Muse model scores 47 percent, against 38 percent for the previous one.
Why it mattersThe consumer assistant is being updated with a model tuned for personal tasks, not only chat.
Meta AI Research ↗Applied Compute raises $350 million
Applied Compute said it reached a $100 million revenue run rate in under 18 months and raised $350 million at a $3.25 billion valuation, led by Elad Gil. Its inference service now serves more than 1 trillion tokens a day.
Why it mattersA training shop is being priced on revenue it says it already has, not on a demo.
Applied Compute ↗Anthropic open-sources a proof of an answer
Anthropic released FlashProve on 2 November, an Apache 2.0 prototype that proves a model produced a given answer. On Llama 4 Maverick, one H100 produces proofs at about 110 tokens a second.
Why it mattersA user could check that a reply really came from the model, not from a swapped server.
Anthropic ↗vLLM serves Kimi K3.1 Pro on day one
vLLM said on 4 November it can serve Kimi K3.1 Pro, a 2.8-trillion-parameter model with 104 billion active and a million-token context. Day-one support covers Nvidia, AMD, Google TPUs and Alibaba's Zhenwu chips.
Why it mattersA huge new model can be served on more than one kind of chip the week it appears.
vLLM ↗42AI raises $62.5 million for a world model
TechCrunch said 42AI left stealth with a $62.5 million seed for a world model built to predict how people act. The money came from the founders and Infinity Group over three years, not all of it from new outside investors.
Why it mattersA seed round this size is for a model of human behavior, not a chatbot wrapper.
TechCrunch ↗Trajectory raises $45 million to train models
TechCrunch said Trajectory raised a $45 million Series A at a $300 million valuation, led by Sequoia, with Nvidia, Bessemer and Bain Capital Ventures. The 18-person company wants non-specialists to keep training a model on the job.
Why it mattersThe round is for the step after a model ships: teaching it a specific task.
TechCrunch ↗Anthropic shows how Claude tidies its notes
Anthropic on 6 November described dreaming, a step that reorganizes an agent's notes between sessions. The post compares how different Claude models file what they learn on a long task.
Why it mattersA long-running agent is only as good as the notes it keeps when the chat is cleared.
Anthropic ↗$17 billion
Google says that is its investment in a first data centre in Wyoming.
The plan includes a 702-megawatt fuel-cell plant with carbon capture, and the halls will be air-cooled. Google ↗
Agents & coding
AI that acts on its own, and the tools developers build with.
Cursor agents can publish a live app
Cursor said on 5 November that a cloud agent can scaffold an app, start Postgres if it is needed, put sign-in behind Cursor, and hand back a link. Nothing goes live until checks pass and a person publishes it. It is on paid plans.
Why it mattersThe coding agent can now ship the internal tool, not only the pull request.
Cursor ↗Mistral previews hosted agents that keep running
Mistral's 5 November notes put Managed Agents in public preview. The same Python agent can run on your machine, on your own server, or on Mistral with agents deploy. A task can run for hours or days and resume after a crash.
Why it mattersA lab is offering the plumbing that lets an agent keep working after the laptop closes.
Mistral ↗Airbnb sets a travel agent for 2027
Airbnb said the travel agent it is building will roll out in the first half of 2027, with maps, photos and itineraries rather than text alone. In the third quarter about 40 percent of support issues were resolved without a person.
Why it mattersA travel giant is putting a date on an agent that would sit inside the booking, not only the search.
Airbnb ↗GitHub Copilot lands in Xcode's assistant
GitHub said on 4 November that Copilot is in public preview inside Xcode's coding assistant, starting with Xcode 27.3 beta. It needs macOS Tahoe 26.6 or later, and every Copilot plan is supported.
Why it mattersApple's editor can now call Copilot without a separate plugin as the only path.
GitHub ↗IBM's mainframe assistant can take an action
IBM said watsonx Assistant for Z 4.0 is generally available. It recommends an action or runs it after a person approves, and it can turn unstructured notes into material the assistant retrieves. Inference can stay on premises.
Why it mattersA mainframe team can ask for an operational step, and still has to approve it.
IBM ↗AWS answers the web for an agent in one call
AWS announced Web Answer on Bedrock AgentCore. One tool call is meant to return an answer from the web for an agent, instead of the agent writing its own search loop. It is generally available.
Why it mattersAn agent can be given a researched answer without a custom search stack.
AWS ↗Honeycomb opens agents that watch production
Honeycomb put personal and team agents in early access on 3 November. A personal agent acts only when told. A team agent can run on a schedule or an alert and, within the rights it was given, file a ticket or page someone.
Why it mattersThe monitoring tool is starting to do the follow-up, not only draw the chart.
Honeycomb ↗Nooks takes $500 million and launches Node
Nooks said General Catalyst's Customer Value Fund committed $500 million of non-dilutive capital, tied to its growth. It also opened early access for Node, an always-on assistant for salespeople, with general availability aimed at early 2027.
Why it mattersA sales-agent company is funding the rollout with capital it does not call a priced round.
PR Newswire ↗Meta adds Spaces beside the Muse agent
TechCrunch said Meta is adding Spaces, custom setups in the Muse sidebar that a person describes, such as a tracker or a plan. Connected apps can keep updating in the background.
Why it mattersThe assistant is being asked to hold a small app, not only answer in a thread.
TechCrunch ↗There are 12 million people who sell for a living in the US alone and all their jobs are about to change.
Video generation
The models and tools that make video.
Runway serves a live character on H100s
Runway published how it keeps Characters, a live video model, at 24 frames a second on H100 GPUs. The post is about serving one session at a time, for as long as the conversation runs, rather than a batch of clips.
Why it mattersA video model sold as a live character has to hold a frame rate, not only win a demo.
Runway ↗Microsoft Foundry previews fal's video models
Microsoft said Kling, LTX and MiniMax video models are rolling out in public preview in Foundry through fal, as a first-party Azure service. The first set includes text-to-video and image-to-video options.
Why it mattersA developer can call outside video models from the same cloud bill as the rest of Azure.
Microsoft ↗Fal gives away FLUX 3 Max clips this weekend
Fal is offering free FLUX 3 Max videos from Friday 6 November at 9am Pacific until Monday 9 November at noon. The cap is up to five a day, 15 seconds at 720p. After that, 720p is $0.0425 a second.
Why it mattersA top video model is briefly free, then billed by the second with no subscription.
fal ↗A lab speeds up Pika's H3 Flex on a GB200
Hao AI Lab at UC San Diego posted on 3 November a faster way to run Pika H3 Flex, including on a GB200. The post is about cutting the time to generate a clip on Nvidia's newest rack chip.
Why it mattersA video model that is slow to sample is being retuned for the chip labs actually have.
Hao AI Lab ↗Luma's workflows run inside ChatGPT
Luma released an MCP connector on 3 November so its image and video workflows can be started from ChatGPT and other chat apps. A person asks in the chat they already have open.
Why it mattersThe video tool is showing up inside the assistant, not only on its own site.
Luma ↗Sand.ai raises $200 million for video models
Sand.ai said it closed a $200 million Series A+, led by CoStone Capital and Wangsu, at a valuation above $1 billion. The Beijing lab plans a full release of its MAGI-2 video model.
Why it mattersAnother video-model company has crossed a $1 billion price within a year of shipping.
DealStreetAsia ↗ElevenLabs adds reviewers to its ad engine
ElevenLabs said people now review ads made by Ads Engine before they go back to the advertiser. The engine adapts existing ads, and the review is the new gate on that output.
Why it mattersA generated ad does not leave the building until a person has looked at it.
ElevenLabs ↗Disney is testing AI lyric videos
The Los Angeles Times reported that Disney, which has fought AI firms on copyright, is now testing AI lyric videos. The piece is about a trial with generated visuals set to songs the studio already owns.
Why it mattersA studio that sued over training is trying generated video on its own catalogue.
Los Angeles Times ↗Tool of the day
Something you can try in the next five minutes.
Cursor Apps
Cursor's cloud agents can now build a web app, add Postgres if it is needed, put it behind Cursor sign-in, and publish a link. Checks run before anything goes live, and a person has to press publish.
Creator tools
The apps and platforms people make video with.
Adobe gives startups a free year of Firefly
Adobe on 2 November opened a year of Firefly Pro Plus, Acrobat Studio with Express Premium, and Semrush at no charge. A startup qualifies if it was founded in the last five years and has raised under $15 million.
Why it mattersThe creative suite is being given away to young companies, with a cap on how much they have raised.
Adobe ↗Suno can start a song from a MIDI file
Suno said on 4 November you can drop a MIDI file into Create and ask it to follow the notes or develop them. MIDI holds pitch and timing, not a recording, so it is not the same as remixing an audio file.
Why it mattersA composer can hand the model the notes, instead of describing the tune in words.
Suno ↗ElevenLabs ships a new sound-effects model
ElevenLabs released SFX v3 for developers on 2 November. Callers set the model id to eleven_text_to_sound_v3. The lab says it follows prompts with several instructions more closely than SFX v2, and handles vocals and music better.
Why it mattersSound effects in an app can be requested from a newer model without leaving the same API.
ElevenLabs ↗Kittl runs inside ChatGPT and Claude
Kittl said on 3 November that its design tools now connect to ChatGPT and Claude. Paid plans get 300 free images after connecting. The connector also works from other apps that can add a remote MCP server.
Why it mattersMockups and product shots can be asked for in the chat, using the design tool's own account.
Kittl ↗Canva ships Affinity for iPad
MacRumors said Canva released Affinity for iPad on 2 November. Sketch, vector, pixel and layout share one file that also opens on the desktop. Core tools are free. A Canva account is required.
Why it mattersThe design app Canva bought is now on a tablet, and the main tools are not behind a subscription.
MacRumors ↗Premiere's assistant now spends Firefly credits
Adobe said the AI assistant inside Premiere now uses Firefly credits from the plan on the account. A credits mark shows on each reply, and more credits can be bought inside the editor when they run low.
Why it mattersThe video editor's assistant is no longer a separate allowance. It draws on the same credit pool.
Adobe ↗DigitalOcean puts its agents on the iPhone
DigitalOcean introduced Cuttlefish, an iPhone app for its managed agents, which stay in public preview. A high-risk tool call can require Face ID or the passcode. Sessions run in the company's microVMs.
Why it mattersSomeone can approve an agent from a phone, instead of leaving a laptop open.
DigitalOcean ↗Redis Iris adds the pieces an agent stores
Redis said four Iris pieces for agents are generally available: a router, playbooks, memory and a context gateway. Two more, Lens and Agent Mesh, are in beta. The starting offer includes 5,000 free messages.
Why it mattersThe database is selling the memory and the routing, not only a place to cache a reply.
Redis ↗xAI's Data Bot takes warehouse questions
xAI said on 2 November that Data Bot, a teammate inside Grok Bot, queries the warehouse and can post to Slack. Shared memory changes need a manager to approve. The lab estimates it handled about 70 hours of work for its data-science team in a week.
Why it mattersA lab is using its own agent on internal data, and publishing what that replaced.
xAI ↗EU AI rules start to apply in Northern Ireland
From 10 November, a limited part of the EU AI Act applies in Northern Ireland under the Windsor Framework, including rules for prohibited and high-risk systems.
If you ship an AI product into Northern Ireland, check whether those duties cover it. UK Government ↗
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.
Weta's chief says studios now demand an AI plan
Variety reported that the chief of Weta FX, the house behind Avatar, says Hollywood has flipped on generative AI. The headline quotes the line studios now send: if you do not have an AI strategy, do not bother.
Why it mattersA vendor that was told to keep AI out is now being told to show a plan.
Variety ↗WME steps back into the OpenAI copyright case
Deadline reported that WME has again joined the copyright fight that includes the New York Times and OpenAI. The agency is weighing in on how models use creative work.
Why it mattersThe largest talent agency is back in the case over what models were trained on.
Deadline ↗USA Today ties a revenue miss to AI search
On the 5 November call, chief executive Mike Reed said search traffic was lower than a year earlier in the second and third quarters. Indirect digital ads, about $100 million a year, fell 32 percent.
Why it mattersA big publisher is putting a number on the traffic it says AI answers took.
The Motley Fool ↗ElevenLabs is the first AI member of DDEX
DDEX said ElevenLabs has joined as a full member, the first AI company in a group of more than 150. The group's draft standard for saying how a generative track was made is aimed at publication in early 2027.
Why it mattersA music-data body is letting an AI company help write the label on a generated song.
DDEX ↗ElevenMusic tracks can post to SoundCloud
ElevenLabs said a person can connect SoundCloud and send an original ElevenMusic track there, with a description, genre and tags. SoundCloud Artist Pro members are offered three months of ElevenMusic free.
Why it mattersA generated track can go to a public catalogue without a separate upload step.
ElevenLabs ↗A São Paulo court stops Meta using news
A São Paulo court granted newspaper groups an injunction that suspends Meta's new commercial terms and bars it from collecting member outlets' content without permission, including for training AI systems. The ruling is dated 3 November.
Why it mattersA platform's terms, which would have covered news used to train models, are on hold in Brazil.
LatAm Journalism Review ↗Newstex licenses news into Cashmere
Newstex said it will supply content from more than 1,600 publications to Cashmere, which already licenses academic and professional books for AI apps. Newstex says it delivers more than 1.5 million original articles a month.
Why it mattersAnother news catalogue is being sold into AI products under a licence, not scraped.
EIN Presswire ↗Weta signs Taiwan as an AI production base
Deadline reported that Weta FX and TAICCA signed a partnership to position Taiwan as an AI hub for production. The deal is meant to lift local visual-effects work toward international film and television.
Why it mattersThe Avatar house is planting an Asia base while studios ask vendors for an AI plan.
Deadline ↗We hit a point in summer where we were not able to serve our customers and had to rate-limit them.
Chips & compute
Chips, data centres and the power behind them.
Google puts $17 billion into a Wyoming data centre
Google said on 5 November it will invest $17 billion in its first Wyoming data centre. The plan includes a 702-megawatt fuel-cell plant with carbon capture, and the halls will be air-cooled.
Why it mattersThe next Google campus arrives with its own power plan, and a promise to pay for the electricity.
Google ↗Shield AI agrees to buy Echodyne for $5.3 billion
Shield said on 5 November it will buy radar maker Echodyne for $5.3 billion. It is also raising $2.5 billion, including a $1 billion Series H at a $28.8 billion valuation.
Why it mattersAn autonomy company is buying the radar and raising the cash in the same announcement.
Shield AI ↗Microsoft is narrowing its rented GPU clouds
CNBC reported on 6 November that Microsoft is deciding which outside GPU-cloud providers it will keep. Partners it has used include CoreWeave, Nebius, Lambda, Nscale and Oracle. People familiar with the talks said an Oracle deal was not renewed.
Why it mattersThe biggest renter of other people's GPUs is cutting the list of landlords.
CNBC ↗Moonshot looked for Nvidia chips in Europe
CNBC reported that Moonshot has talked with European clouds about renting Nvidia Blackwell clusters of 8,000 GPUs or more. One proposal would have charged the Chinese lab more upfront than a buyer outside China.
Why it mattersA Chinese model lab is shopping for chips it cannot easily buy at home.
CNBC ↗New data centres are being sited on their own power
TrendForce said on 4 November that three-quarters of October's added capacity in its tracker, 3,111 megawatts, sits on campuses built next to their own supply. Amazon's cloud unit was a large share of the month's additions.
Why it mattersThe scarce input is no longer only the chip. It is a power plant next door.
TrendForce ↗Micron slows a Taiwan factory for AI memory
The Taipei Times reported that work has largely stopped on Micron's Taoyuan fab, which had been aimed at high-bandwidth memory. Micron said Taiwan remains part of its footprint and that operations continue, and it did not confirm a halt.
Why it mattersA memory maker's next AI-chip factory is in doubt in the country that makes most of its DRAM.
Taipei Times ↗Restricted memory is shipped inside a cheap processor
Tom's Hardware reported that high-bandwidth memory barred for China is being packaged with cheap, unsanctioned processors so the shipment can leave as a legal export. The loophole is the bundle, not a hidden crate.
Why it mattersExport rules written for a chip are being met by stapling it to a part nobody restricted.
Tom's Hardware ↗The White House says near-term data centres stay
CNBC reported that Kevin Hassett said data centres already scheduled over the next couple of years are not affected by this year's local ballot measures. He said the country is large enough that those projects can still be built.
Why it mattersAfter a wave of local votes, the White House is telling industry the next builds are still on.
CNBC ↗Texas's next governor proposes a data-centre pause
CNBC reported that Gina Hinojosa, the Texas governor-elect, has proposed a statewide pause on new data-centre permits. Projects already under construction would reportedly be exempt.
Why it mattersThe state with the most new server halls is being asked to stop issuing permits.
CNBC ↗Policy & safety
Laws, lawsuits and safety incidents.
Justice Department opens a super-intelligence unit
The Justice Department on 5 November announced a Super Intelligence Task Force led by Stanley Woodward and Stephen Weiss. It will coordinate cases on misuse by states, criminal groups and fraud.
Why it mattersThe department is giving AI enforcement a single desk, under the White House's new name for the technology.
US Department of Justice ↗House chairs ask Jensen Huang for a briefing
CNBC reported that three House chairs asked Jensen Huang to brief them on contacts with a Megaspeed executive. The invitation is for 18 November, the day after Nvidia reports earnings.
Why it mattersChip export questions are moving from letters to a requested hearing, with a date on it.
CNBC ↗The EU AI Act starts to apply in Northern Ireland
The UK said a specific, limited slice of the EU AI Act applies in Northern Ireland from 10 November under the Windsor Framework. It covers prohibited, high-risk and limited-risk systems, and it tells suppliers to check whether they are in scope.
Why it mattersA product sold into Northern Ireland can pick up EU AI duties that the rest of the UK does not have.
UK Government ↗Chicken Soup sues over Grok training on its books
A copyright suit filed 4 November in Northern California accuses SpaceX Super Intelligence, the former xAI entities, of copying Chicken Soup for the Soul books to train Grok. The complaint seeks a jury and is not a class action.
Why it mattersA mass-market book brand is testing the training fight against Grok, one publisher at a time.
Copyright Now ↗West Virginia sues Snap over My AI
West Virginia sued Snap on 5 November under the state's consumer-protection law. A large part of the complaint is about the My AI chatbot, which the state says was rushed to users of all ages and encouraged children to treat it as a friend.
Why it mattersA state is using an ordinary consumer law against a chatbot built into an app teenagers already have.
Lawsuit Intelligencer ↗Delaware subpoenas OpenAI after a security incident
Spotlight Delaware reported that the state subpoenaed OpenAI on safety and governance, its first compulsory step since the restructuring. The request seeks board materials on model and agent problems.
Why it mattersThe state that charters the company is asking for the board file, not only a blog post.
Spotlight Delaware ↗Malaysia questions Anthropic on an influence campaign
Malay Mail said the government told parliament it is using the Online Safety Act on claims of an AI influence campaign, and has asked Anthropic for a technical explanation. A separate AI governance bill is being drafted, with no date set.
Why it mattersA government is treating a reported model abuse as an online-safety case, and writing a bill beside it.
Malay Mail ↗New York floats a wider AI oversight bill
StateScoop reported on a bill called RAISE Act 2.0 that would expand New York's oversight of frontier AI developers. It was introduced in late October and is a follow-on to the state's first safety statute.
Why it mattersThe state that already passed an AI safety law is drafting a second, broader one.
StateScoop ↗SI development will continue to be powered by American innovators.
Robotics
Robots, self-driving and AI in the physical world.
Boston Dynamics unveils Stretch Gen 2
Boston Dynamics this week showed Stretch Gen 2, a trailer-unloading robot rated at up to 1,000 cases an hour. It cuts most of the internal cabling from the first generation, and a full drive swap takes about an hour instead of a day.
Why it mattersThe unloading robot is being rebuilt around the failures that showed up after 100,000 hours on docks.
Engineering.com ↗Waymo opens more of Los Angeles and beyond
The Los Angeles Times said Waymo on 3 November added Torrance, El Segundo and Manhattan Beach, inside a 210-square-mile service area. The same day it extended San Diego, Nashville, Orlando and San Francisco. Airport pickups at LAX are still off.
Why it mattersThe robotaxi map grew in several cities on one day, and the airport is still a gap.
Los Angeles Times ↗Waymo says it will pause when the snow is too much
Waymo said that if a forecast or live weather exceeds its safety rules, it will shrink the service area or pause for hours or days. The winter fleet gets snow tires, and cars share road and weather data with each other.
Why it mattersThe robotaxi company is publishing when it will take itself off the road.
Waymo ↗Zipline raises $1.5 billion for drone delivery
Zipline said on 5 November it raised $1.5 billion at a $21.5 billion post-money valuation, led by Paradigm, WCM and Layer Global. It has flown 150 million commercial autonomous miles without a safety incident. Walmart is adding 300 sites.
Why it mattersThe largest delivery-drone network is raising like a software company, on a safety record it publishes.
Zipline ↗Panasonic shows a robot that corrects by touch
Panasonic said T-Rex pairs vision, running at about 5 hertz, with fingertip corrections at about 20 hertz. On 12 real contact-heavy tasks it reports 65 percent average success. The results are set for CoRL in Austin, 9 to 12 November.
Why it mattersA robot policy is being sold on what the fingers feel, not only on the camera.
Panasonic ↗XPENG open-sources a robot perception model
XPENG said researchers with the company released GroundingPI, a 4-billion-parameter vision model, on 4 November. It scored 73.68 on their perception tests, against 71.54 for GPT-6 Astra. Weights are public.
Why it mattersA carmaker is publishing the eyes of a robot, and scoring them against a frontier model.
XPENG ↗NEURA, E.ON and enviaM plan robots for the grid
NEURA said on 5 November it will train robot skills with E.ON and enviaM for inspections and substation checks. A pilot partner expects first field experience in 2027.
Why it mattersA humanoid maker is aiming at power sites, where the customer keeps the data and the skills.
NEURA Robotics ↗Xiaomi shows humanoids on an eight-hour factory shift
Electrek said Xiaomi posted video of full-size humanoids sorting parts at its car factory, labeled autonomous. Xiaomi says the robots covered that station's production day. It did not publish a success rate or how often a person stepped in.
Why it mattersA phone maker is showing robots on its own line, without the numbers a buyer would ask for.
Electrek ↗Verne releases an open-weight robot policy
Verne Robotics published V0.5, an open-weight model that turns camera views and a plain instruction into robot motion. The weights are on Hugging Face, and the lab says not to run the base model unattended.
Why it mattersA factory can try a published robot policy before it pays for a closed one.
Verne Robotics ↗Science & health
AI in biology, medicine, climate and maths.
FDA clears an AI estimate of sleep time
Nihon Kohden said the FDA cleared software that estimates total sleep time from a pulse-oximeter signal. It is for adults, it is not a diagnosis by itself, and it is the company's first cleared machine-learning sleep tool.
Why it mattersA home sleep test can use a signal it already records, instead of adding wires.
Nihon Kohden ↗FDA clears a wearable heart-rhythm model
PulseAI said the FDA cleared its model for reading heart rhythms from a wearable. Recordings go to the cloud and a structured result comes back for a clinician. The company cites a study of 296 people against the Apple Watch's own algorithm.
Why it mattersA watch's ECG can be read by a second model that has its own clearance.
Health Technology Net ↗GE adds lung and valve AI to a handheld scanner
GE HealthCare added Lung Guidance and an aortic-stenosis check to the Vscan Air handheld ultrasound. Both are FDA-cleared and CE-marked. Existing customers on the innovation package are expected to get the update by the end of November.
Why it mattersA pocket scanner is getting two AI reads that used to wait for a larger machine.
24x7 Magazine ↗FDA classifies an AI reader for Pap slides
The FDA on 5 November classified a digital cervical-cytology system that uses an AI algorithm as class II. It is a prescription device that marks regions on a digital Pap image for a person to read. It is not exempt from premarket review.
Why it mattersThe agency has a regulatory box for an algorithm that points at cells, and a person still signs the read.
Federal Register ↗FDA grants a breakthrough tag to a skin-biopsy AI
CLP Magazine said the FDA gave Breakthrough Device Designation to Primaa's Cleo Skin, which reads digitized skin-biopsy slides. The tag is not a licence to sell. The software already has a European IVD certificate from March.
Why it mattersA pathology model has a faster US review path, and still cannot be marketed on that tag alone.
CLP Magazine ↗Planet lets you search the Earth in plain language
Planet on 5 November opened an app that searches ten years of its imagery from a sentence, and shows change on a map. Anyone can use it. The company says a summer beta had hundreds of users, and access is free with limits it can change.
Why it mattersSatellite search no longer starts with a drawing tool and a specialist.
Planet ↗DeepMind posts chemistry skills for Antigravity
Google DeepMind published a walkthrough of chemistry agents inside Antigravity, its coding agent. The post shows the agent searching public compound libraries and proposing molecules, using tools built on open chemistry software.
Why it mattersA general coding agent is being pointed at a lab task that used to need a specialist stack.
Google DeepMind ↗Stanford tests medical AI on smaller images
Stanford said Med-REDUCE tried lower-resolution images and smaller models on dermatology, pathology and radiology tasks. Images went from 512 to 64 pixels on a side. Distilled models matched or beat the larger ones in 39 of 42 comparisons.
Why it mattersA hospital model may not need the full image, or the full model, for the tasks in this test.
Stanford ↗A hospital trial records more side effects with a chatbot
ASCO's news service said an interim look at a chemotherapy trial found physicians who could use ChatGPT recorded more toxicities, 5.4 versus 2.8, with no gain for patients on a quality-of-life score. Enrollment stopped after that look.
Why it mattersA chatbot in the clinic changed what got written down, and not the outcome the trial was built to move.
ASCO ↗Open source
Open-weight models and community releases.
PFN releases PLaMo 3 Translate weights
Preferred Networks released PLaMo 3 Translate in four sizes, from 0.8 billion to 31 billion parameters. It covers 53 languages. Commercial use is free under a revenue cap.
Why it mattersA translation model is open to download, with a line on the licence for larger companies.
Preferred Networks ↗Allen Institute opens a block world for robot training
The Allen Institute for AI on 6 November introduced ToyBlocks, an open tabletop simulation of an arm and blocks. Demonstrations are generated in simulation rather than by a person teleoperating, and the code is public.
Why it mattersA robot skill can be practised in a world anyone can rerun, before it touches hardware.
Hugging Face ↗A small adapter tunes a model to its coding harness
A Hugging Face post on 6 November describes one small adapter per model and harness, trained inside unmodified pi, opencode and hermes. A checker sits between the harness and the model and scores reward-hacking tool calls at zero.
Why it mattersThe same weights can be nudged to behave inside the agent program that will actually run them.
Hugging Face ↗Nvidia open-sources a faster robot motion planner
Nvidia published cuRobo v0.9.0 on 5 November. The release, titled cuRoboV2-Fast, says the planner is about 10 times faster. It adds torque-aware planning and a GPU path search to the existing open repository.
Why it mattersA robot arm's path can be recomputed on a GPU without a proprietary planner.
GitHub ↗Transformers patches a sandbox escape
Hugging Face published Transformers 5.22.2 on 6 November as a security and distributed-training patch. One change hardens the Jinja sandbox used when a template runs. The previous tag on the list is from 30 October.
Why it mattersA library that loads hundreds of models is closing a hole in how templates execute.
GitHub ↗Unsloth's beta speeds image and video loads
Unsloth 0.1.916-beta, published 6 November, says unified memory offload now works for large models, and image and video loads are up to 40 percent faster. Two matching idle Nvidia GPUs can split a step when the model fits on one card.
Why it mattersFine-tuning a vision model spends less time waiting on the files, on a beta you can install today.
GitHub ↗The open speech leaderboard adds two Englishes
Hugging Face said on 6 November that the Open ASR leaderboard now includes private sets of Egyptian and Nigerian conversational English, 1.34 and 1.55 hours. The sets keep punctuation, casing and the hesitations of real speech.
Why it mattersA speech score built on US audio can now be checked against English spoken elsewhere.
Hugging Face ↗Rerun open-sources the engine its agents run on
Rerun published its app and the engine its always-on agents run on, under the AGPL-3.0 licence, with a Docker guide for self-hosting. The company's cloud service still runs the same agents for people who do not want to host them.
Why it mattersAn always-on agent can be inspected and run on your own machines, under a copyleft licence.
Rerun ↗Cost saving tip
One change that makes the same work cheaper, from today's news.
Use the free window, then pay by the second
Fal is giving away FLUX 3 Max clips until Monday 9 November at noon Pacific, up to five a day, 15 seconds at 720p. After that, 720p is $0.0425 a second, with no subscription.
fal ↗Render the clips you need before Monday noon Pacific, then pay only for extra seconds.
Quick hits
Smaller stories worth a glance.
- Suno posts a fresh song prompt each daySuno's release notes say the app now offers a new prompt every day, on every plan, and saves the songs to a calendar in the library.Suno ↗
- Copilot's dashboard counts agent useGitHub said the Copilot impact dashboard now includes chat and the VS Code agents window, and renames passive users as chat-first users.GitHub ↗
- A Korean lab builds a humanoid from local partsSeoul Economic Daily said KIST's KAPEX humanoid uses more than 90 percent Korean parts, and its actuators produce up to 300 newton-metres.Seoul Economic Daily ↗
- Six humanoids staffed a shop on a livestreamInteresting Engineering said six AGIBOT G2 humanoids stocked shelves, made coffee and cleared rubbish during a livestream on 2 and 3 November.Interesting Engineering ↗
- Glasshand rents a coding agent a real iPhoneGlasshand, a 5 November Product Hunt launch, rents physical iPhones and Macs in its data centre so an agent such as Claude Code or Codex can use a device.Hunted ↗
- llama.cpp publishes build b12179The project tagged prerelease build b12179 on 6 November, the current build after b12138 landed the night before.GitHub ↗
Deals & funding
Who raised money, and who bought whom.
Funding rounds 7
Investors put money into a company in exchange for a share of it.
Accel backs Thinking Machines Lab
- Raised
- $5B
Thinking Machines Lab now valued at $40B
Business Insider reported Thinking Machines is raising $5 billion at a $40 billion pre-money valuation, with Accel leading. The company declined to comment, and the round is not closed.
Disruptive backs Shield AI
- Raised
- $2.5B
Shield AI now valued at $28.8B
Shield AI is raising $2.5 billion, including a $1 billion Series H at a $28.8 billion post-money valuation, $1 billion of structured equity from Blackstone and $500 million of debt led by JPMorgan.
IXInnovation X CapitalBBlackstoneJCJPMorgan Chase+0 more
Paradigm backs Zipline
- Raised
- $1.5B
Zipline now valued at $21.5B
Zipline raised $1.5 billion at a $21.5 billion post-money valuation, led by Paradigm, WCM and Layer Global, to expand autonomous drone delivery.
WIWCM Investment ManagementLGLayer Global+0 more
Elad Gil backs Applied Compute
- Raised
- $350M
Applied Compute now valued at $3.25B
Applied Compute raised $350 million at a $3.25 billion valuation, led by Elad Gil, and said it has a $100 million revenue run rate.
Kleiner PerkinsLCLux CapitalNNVIDIABBlackstone+0 more
CoStone Capital backs Sand.ai
- Raised
- $200M
Sand.ai now valued at >$1B
Sand.ai closed a $200 million Series A+, led by CoStone Capital and Wangsu, at a valuation above $1 billion, for its video models.
WSWangsu Science & Technology+0 more
Infinity Group backs 42AI
- Raised
- $62.5M
42AI valuation not disclosed
TechCrunch said 42AI came out of stealth with a $62.5 million seed for a world model aimed at predicting human behavior. The sum came from the founders and Infinity Group over three years, and was not all new outside capital.
Sequoia backs Trajectory
- Raised
- $45M
Trajectory now valued at $300M
Trajectory raised a $45 million Series A at a $300 million valuation, led by Sequoia, to let non-specialists keep training models.
NvidiaBVBessemer Venture PartnersBCBain Capital Ventures+0 more
Acquisitions 1
One company buys another outright.
Shield AI buys Echodyne
- Price
- $5.3B
Echodyne is now part of Shield AI
Shield AI agreed to acquire radar maker Echodyne for $5.3 billion, with $4.8 billion paid at closing and up to $500 million more if targets are met.
Partnerships & contracts 1
A large commercial agreement, or a government contract.
Nooks sign a deal
- Value
- $500M
Multi-year agreement
General Catalyst's Customer Value Fund committed $500 million of non-dilutive capital to Nooks, which also launched Node, a sales assistant.
Moves
Who changed jobs in AI.
On the calendar
What's coming up, in date order, and how sure we are about each.
- 2–13NovProtoStar climate challengeToday
- 5–7NovThird LA Vertical Drama MarketToday
- 6–7NovProtests call for a data-centre pauseToday
- 7NovTerry McMillan Presents: Paradise with You premieresToday
- 9–20NovCOP31 climate conference in Antalya2days away
- 9–12NovCoRL meets in Austin2days away
- 9–11NovThe Hinton Lectures on AI safety2days away
- 9NovCharmed By a Southern Monster premieres2days away
- 10NovEU AI Act starts to apply in Northern Ireland3days away
- 11NovJules becomes read-only4days away
- 12NovUpdated Claude usage policy takes effect5days away
- 15NovComments close on China's draft tech-export list8days away
- 15NovWorkBuddy and CodeBuddy free period ends8days away
- 18NovWebinar on running a dock with Stretch11days away
- 18NovNvidia invited to brief Congress11days away
- 20NovAI Film Award submissions close13days away
- 20NovConductor cloud service shuts down13days away
- 20NovUpdated OpenAI usage policies take effect13days away
- 21NovInaugural edition in Hollywood2weeks away
- 22Nov'The Challenge Sverige' originally scheduled (now delayed to 2027)2weeks away
- 29NovDaybreak Blue sessions can be used for training3weeks away
- 30–3Nov–DecContent London 20263weeks away
- 30NovRose d’Or Awards3weeks away
- 1DecHearing on Apple's bid to block testimony3weeks away
- 2–3DecWorkshop on autonomous surgical robots4weeks away
- 4DecAI-made film opens in limited release4weeks away
- 4DecGods Don’t Give Gifts premiere4weeks away
- 4DecSingapore AI Safety Fellowship ends4weeks away
- 31DecHy4 half-price offer ends8weeks away
- 31DecJules shuts down8weeks away
- Later in NovGuided Learning pilot in US schools
- Mid-NovXREAL Aura shipments begin
- H1 2027Digit 5 early customer access
Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.