Monday, 16 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.
- Bessent says the next chief should not come from inside
- Outside
- Jensen Huang resigned
- Thursday
- first earnings report under the interim chief
- Tuesday
- Bessent says Trump drew a line above that chip for China
- H200
Bessent wants Nvidia's next chief hired from outside
Treasury Secretary Scott Bessent said on Sunday that Nvidia needs a dramatic shakeup and that the board should hire its next chief executive from outside the company. He questioned whether the firm had pressed chip sales he called seemingly illicit in order to lift revenue. Nvidia reports earnings on Tuesday, the first print since Jensen Huang resigned on Thursday.
Why it's the story of the dayThe Treasury secretary is questioning the sales culture two days before Nvidia's first earnings without Huang.
Startup of the day
A young company people are talking about today.
Edelbyte comes out of stealth with $70 million for data the agents can query
Edelbyte, founded in 2025 by former Microsoft Kusto engineers, said it raised $70 million across a seed and a Series A led by Sequoia and Quiet Capital, with Greenoaks and Swish Ventures. The Tel Aviv and San Francisco company sells search and analytics for AI applications, in a public cloud, a private cloud or on premises. Chief executive Avner Aharoni said the systems companies use were built for a world before AI.
Why it's worth a lookThe round is for the layer agents read, which is where a lot of enterprise data still cannot be queried.
Newswire ↗Models & labs
New models, the labs behind them, and the rules they answer to.
Aion Instruct replaces Phi Silica on Windows PCs
Microsoft said on 13 November that Aion Instruct, a 2.3 billion parameter on-device model with a 32,000-token window, will replace Phi Silica, which was 3.8 billion parameters with a 3,500-token window. An Insider rollout on Copilot+ PCs is planned for 20 November, with retail targeted for late January, on Nvidia RTX and AMD RDNA4 GPUs. Invisible watermarking is added for text-provenance rules in some regions, and custom LoRAs do not carry over unchanged.
Why it mattersThe model that runs on a Windows PC without a cloud call is changing size, context and how its text is marked.
Microsoft ↗Haru-4B listens and speaks at the same time
Higurashi released Haru-4B on 13 November, a full-duplex speech model built on Qwen3.5-4B. It takes a step every 133 milliseconds, and the company measured about 65 milliseconds end to end on a MacBook. Pretraining used 40 million hours of real speech across more than 160 languages, with the release focused on nine, and the weights are on Hugging Face under Apache 2.0.
Why it mattersA phone-style conversation can sit on a laptop if the model never waits for the other person to finish.
Higurashi ↗Lara Mini is trained from scratch in 45 languages
Translated said on 13 November that Lara Mini was trained from scratch on 6 trillion tokens in 45 languages, including all 24 official languages of the European Union. In the company's own comparison of 36 open base models with up to 2 billion active parameters, it scored 57.0 overall, 3.1 points ahead of the next model, with 600 million active parameters and 7.2 billion in total. It is not in the Lara Translate app yet.
Why it mattersA translation app's next offline model is being judged on languages other than English, not only on the English tests.
Lara Translate ↗Salesforce opens a speech model with 4 billion active weights
Salesforce posted SOV-Transcribe-1 on 12 November, an Apache 2.0 speech model with 4 billion active parameters and 26 billion in total. As of that day the company said it was first among open-weight non-streaming speech-to-text models on Artificial Analysis, with a 2.28 percent word error rate, against 2.81 percent for Nvidia Nemotron 3.5 Transcribe and 2.77 percent for Mistral Voxtral Small. On VoiceCodeBench it said 76.3 percent of recordings passed.
Why it mattersA company can keep call transcripts on its own machines without taking the larger open speech model.
Hugging Face ↗Agnes 2.5 Pro Theta posts weights for local chat
Agnes AI published Agnes 2.5 Pro Theta on Hugging Face on 15 November, a multimodal mixture-of-experts checkpoint with about 187.8 billion parameters and a 262,144-token context, in BF16 for Transformers. The model card's own comparison, which it calls a supplied snapshot, lists 92.9 on GPQA Diamond, 90.5 on LiveCodeBench v6 and 84.7 on AA-LCR. The license is Apache 2.0.
Why it mattersAnother very large open checkpoint is available to run locally, with the publisher's own scores attached.
Hugging Face ↗Aether-15B fine-tunes Qwen3 for Indonesian and Japanese
K-Ai posted Aether-15B on 14 November, a fine-tune of Qwen3-14B-Base with 14.77 billion parameters, aimed at Indonesian, English and Japanese, with extra weight on mathematics and coding. The card says the adapter was merged into the base weights. The license is Apache 2.0.
Why it mattersA mid-size open model is being aimed at languages that the biggest English releases still treat as a side case.
Hugging Face ↗Tess 4.5 tunes Inkling 2 Small for coding agents
Continuity Research posted Tess-4.5-594B-A17B on 13 November, an Apache 2.0 fine-tune of Thinking Machines' Inkling 2 Small. The card describes 594 billion parameters with 17 billion active, image input and a 1 million token context, and it reports 88.2 percent on SWE-bench Pro v2 Hard with 9 percent fewer output tokens. A separate LoRA is posted beside the merged weights.
Why it mattersThe open weights from last week already have a coding recipe on top of them.
Hugging Face ↗Discrete codes try to stand in for a vision encoder
dNaViT asks how far a finite set of discrete visual codes can go. It is a vision transformer that turns a frozen encoder's features into discrete primitives, more like language tokens, using a light residual on top of that encoder. The paper is dated 9 November.
Why it mattersIf pictures can be written as tokens, the same training stack used for text can be pointed at them.
arXiv ↗Fewer bits can beat averaging the answer order
The paper asks whether storing a small local model's weights in fewer bits makes it more sensitive to the order of answer choices, and which use of a fixed budget is cheaper. It tests Qwen3-0.6B and Qwen3-1.7B with simulated round-to-nearest quantization at 8, 6, 4 and 3 bits, against fp16, on ARC-Easy, ARC-Challenge, MMLU and HellaSwag. The paper is dated 10 November.
Why it mattersOn a laptop, the memory you spend on precision and the compute you spend reordering the choices are the same budget.
arXiv ↗$230.9M
Cerebras said third-quarter core revenue was $230.9 million, up 70 percent from a year earlier.
GAAP revenue was $210.8 million. The company guided to $310 million to $330 million of core revenue for the fourth quarter. SEC ↗
Agents & coding
AI that acts on its own, and the tools developers build with.
Sierra's Poppy lets a personal agent talk to the shop
Sierra published Personal Agent Protocol 1.0 on 13 November. Shopify has turned it on for every merchant, so each store publishes a Poppy file, and Muse looks for Poppy first on every site it visits. More than 100 design partners are named, including Delta, Intuit, Salesforce, Twilio, Wayfair and Zoom. Mercury can show balances and start payments that still follow the account's approval rules, and Turkish Airlines can quote a fare the customer approves before paying.
Why it mattersThe assistant in the chat and the shop's own agent now have a shared way to meet, instead of the assistant scraping the page.
Sierra ↗New Bedrock agent sessions start from a snapshot
AWS said on 12 November that new Amazon Bedrock AgentCore runtimes use a next-generation runtime by default. Each session restores from a snapshot of the initialized agent, can scale up to 8 vCPU and 32 GB of memory, and scales idle agents to zero, still inside its own Firecracker microVM. On a 2 GB image, AWS said the 75th percentile start is under 2 seconds, down from 29.7 seconds.
Why it mattersAn agent that sits idle between turns no longer has to pay for a warm machine, or wait half a minute to wake.
AWS ↗Qwen Code 0.27.4 adds a managed session and chat alerts
Qwen Code 0.27.4, tagged 15 November, lets a managed agent session use granted local directories and an interactive web terminal without storing the credentials or paths. Outputs can go to Feishu and WeCom webhooks, with long replies split into ordered segments instead of being cut off. Auto mode asks for an explicit go-ahead on risky actions. The note says there are no known breaking changes.
Why it mattersA coding agent can be handed a folder and a chat channel without keeping the login around after the session.
GitHub ↗LangChain sends xAI chats through the LangSmith gateway
langchain-xai 1.4.0, tagged 15 November, routes ChatXAI through the LangSmith gateway.
Why it mattersFour chat providers can now be pointed at one gateway, which is where the traces and the bill land.
GitHub ↗SingVL scores a choice in one forward pass
Inclusion AI posted SingVL S1, a 4 billion parameter decision model on Qwen3.5, on 11 November. It answers named questions over text or JSON and up to four images, scoring the options in a single forward pass. On the card's Public Decision Index 0.3 it lists 69.76 on text, against 60.70 for TypeSafe Jev 1.14, and 73.94 on vision.
Why it mattersA small model that only has to pick from a list can skip writing out a chain of thought.
Hugging Face ↗Meta Encoder embeds, retrieves and decides
Meta posted Meta Encoder v1 on Hugging Face on 12 November. Callers pass content and a task, and the wrapper writes the prompt. Encode returns a 6,656-dimensional vector. Retrieve ranks a corpus. Decide answers questions, including over images or video. The card says one 80 GB GPU holds the model in bf16, and that it needs Transformers 5.15 or newer.
Why it mattersSearch, ranking and a yes-or-no can share one encoder instead of three separate models.
Hugging Face ↗WhiteMatter raises $20 million to put a guardian on each agent
WhiteMatter, formerly QuilrAI, said it raised a $20 million Series A led by Blumberg Capital, with Crew Capital and Sprout & Oak. The San Francisco and Austin company assigns a guardian agent to each deployed agent. It said more than 2.25 million agents are under management, revenue rose fivefold in six months, and customers include Sentara Health, Alter Domus, New American Funding and Concentrix.
Why it mattersThe agent that does the work and the agent that watches it are being sold as a pair.
WhiteMatter ↗A routing test covers twelve languages and mixed ones
PolyRoute is a benchmark for typed-decision models, which answer in one forward pass, outside English. It has 322 support messages in 12 languages and 9 writing systems, plus 156 code-mixed or romanised messages, including Hinglish, Spanglish and Arabizi. The decision is a three-way department route. The paper is dated 10 November.
Why it mattersA model that only has to pick a department can look solved in English and still fail when the message mixes languages.
arXiv ↗The notes an agent keeps can be trained, not just written
Reinforcement learning from a single score throws away the error logs and other detail from a trial. The paper keeps the Agentic Context Engineering harness fixed and trains one model to act, reflect and curate a task-local playbook. Candidate handoffs are scored, so the notes passed to the next attempt are learned rather than only prompted.
Why it mattersWhat an agent writes down for its next try can be part of the training signal, not only a prompt trick.
arXiv ↗Tool of the day
Something you can try in the next five minutes.
Kitten TTS 2 Mini
An 82 million parameter voice model, 16 languages and 118 voices, that runs on CPU from an 82.7 MB ONNX file.
pip install kittenml
Video generation
The models and tools that make video.
LongCat-Video-Avatar 1.6 keeps a face moving past ten minutes
Meituan's LongCat repo said on 15 November that Avatar 1.6 is an open-source upgrade for audio-driven human video. It aims for more natural lips and more expressive face, head and body motion than 1.5, with less error building up on long clips. The note says a post-training mix of reinforcement learning, distillation and long rollouts allows videos past 10 minutes. Weights are on Hugging Face under the MIT license.
Why it mattersA talking-head clip that used to drift after a minute is being asked to hold still for a length closer to a scene.
GitHub ↗Duo 1.2 weights are up for full-duplex video and sound
Inclusion AI posted Duo 1.2 causal-base on Hugging Face on 12 November, calling it a full-duplex world model for real-time audio and video. The card says this variant improves generation quality over the earlier Duo weights. Code is on GitHub. The license is the MiniMax H3 community license, not a standard open-source grant.
Why it mattersA model that has to watch and answer at the same time is now a set of weights, under a license you have to read.
Hugging Face ↗Higgsfield puts $1 million on short films made in its app
Higgsfield's contest page, dated 12 November, offers $1,000,000 across 20 awards: $200,000, $150,000 and $100,000 for the top three, two audience prizes of $50,000, and 15 category awards of $30,000. Films run 3 to 7 minutes and must be made in a contest project from 12 November through 26 November at 19:00 UTC. Every entry, including ones marked private, becomes public at the deadline, with prompts and generation history.
Why it mattersThe prize money comes with a rule that the prompts and the generation history are published too.
Higgsfield ↗A benchmark asks a video model to render a world the code already built
WorldDraft separates building a 4D scene in code from drawing it. The test is whether a video model keeps the motion and the camera path it was given, and whether the look described in text stays put over time. The benchmark has 432 paired cases across 48 executable world families, in six kinds of motion. It crosses how rough the draft is with whether the requested look agrees with the draft or fights it.
Why it mattersIf the scene is already specified, the video model is being graded as a renderer, not as the author of the physics.
arXiv ↗A harness uses a language model's physics to repair a clip
PhysLoop is a harness that guides a language model while a video model generates, selects and corrects clips, so the language model's physical knowledge is distilled into the training videos. The authors say it can also repair a clip at inference time. They release 100,000 videos of indoor and outdoor scenes, along with the generation setup.
Why it mattersThe model that writes the critique is being used to build the next training set, not only to score a demo.
arXiv ↗An open recipe turns an image generator into a vision system
Open Vision Banana publishes the data, weights and choices behind a system that treats vision tasks as conditional image edits. Starting from 4 billion and 9 billion parameter open generators, and 2.1 million to 3.8 million public training pairs, the authors say it beats Vision Banana's reported Cityscapes zero-shot score, 73.9 mean intersection over union against 69.9.
Why it mattersThe closed recipe that treated segmentation as an image edit now has a public counterpart.
arXiv ↗A paper says a pretty clip is not yet a world model
WorldFoundry argues that a generated video can look right and still fail as a world model if it has no persistent state, no physical dynamics, or no consistent answer when something in the scene is changed. The authors define a world model as a system that keeps a history-conditioned state, updates it when an action happens, and rolls out what follows, including for actions that were not taken. The paper is dated 7 November.
Why it mattersThe test being proposed is whether the scene is still there after you look away, not whether the frame is sharp.
arXiv ↗A driving video model also draws the view from above
BirdDreams generates a driver's-eye video and a simple bird's-eye video together, so the layout of the road and the other cars is part of the prediction rather than an input. The overhead view is trained from scene annotations and drawn as plain shapes at a fixed scale. Perspective video keeps the signals and the look of objects that the overhead sketch leaves out.
Why it mattersA model that has to invent where the other cars go can no longer be handed their future paths.
arXiv ↗A benchmark counts how many tries a world model needs
The paper asks how many times a world model must see a new game mechanic before it can predict the effect. MechEval labels 69 recurring mechanics across 20 games from ARC-AGI-3, read from the games' own programs. Each prediction is scored against that label.
Why it mattersA world model that looks fluent on a familiar game may still be guessing the rule it has only seen once.
arXiv ↗Nvidia reports on Tuesday
CNBC noted that Bessent's remarks landed two days before Nvidia's first quarterly earnings since Jensen Huang's resignation. Colette Kress is interim chief executive.
The report is the first set of figures under the new leadership. CNBC ↗
Creator tools
The apps and platforms people make video with.
Phosphene 5.0.24 speaks ten languages
Phosphene 5.0.24, released 14 November, adds Spanish, German, French, Italian, Portuguese and five further languages to the Mac app, alongside English. The notes say the language pack is free. Phosphene is the local video, image and music app installed through Pinokio.
Why it mattersA local video studio on a Mac can now follow the language of the person editing, not only English menus.
GitHub ↗Phosphene 5.0.25 uses less memory at the end of a long clip
Phosphene 5.0.25, released 14 November, says long LTX clips use about a third less memory at the end of the render. The notes call it a single change, and say it is free.
Why it mattersThe last step of a long local render was the one running out of memory.
GitHub ↗Phosphene's storyboard grid marks the shot that failed
Phosphene 5.0.26, released 15 November, shows in the storyboard grid which shot has a problem. The notes say that before this, Render all drafts did not point at the shot that failed. It is a one-change release.
Why it mattersA batch of local drafts can say which frame broke, instead of failing as a pile.
GitHub ↗LiteLLM signs releases without a long-lived key
LiteLLM 1.111.0-dev.4, tagged 15 November, is signed with Sigstore keyless signing through GitHub's identity token. The notes say cosign verify with the old cosign.pub key does not work for this release, and that earlier releases still verify with that key. A partly published dev.3 tag points people at dev.4.
Why it mattersChecking that an inference proxy's image is the one the project built no longer depends on a key file in the repo.
GitHub ↗LanceDB 0.42 beta.15 exposes full-text operators
LanceDB 0.42.0-beta.15, tagged 15 November, lets Python see full-text operators and match counts, and adds filters when listing server-side jobs. It also shows progress and image details when a function is registered.
Why it mattersA search over a table can ask how a match was made, not only which rows came back.
GitHub ↗Ultralytics OCR reads Korean, Arabic and more scripts
Ultralytics 8.4.212, tagged 13 November, lets OCR pick a PP-OCRv5 model for Korean, Cyrillic, Arabic, Devanagari, Thai, Greek, Tamil or Telugu. Those models run through ONNX Runtime. The same release notes changes to Platform discovery and training reliability.
Why it mattersA detection library's text reader is no longer limited to the Latin script it started with.
GitHub ↗Wrangler simulates Artifacts bindings on the local machine
Wrangler 4.162.0, tagged 13 November, uses the local simulator for Artifacts bindings that do not set remote. The notes call it a behavior change: during development those bindings now default to the simulator.
Why it mattersA binding that used to reach the network in dev now stays on the laptop unless you ask for remote.
GitHub ↗Luma's agent sorts a board of up to 10,000 assets
Luma said on 14 November that its agent can group and label a board, a frame or a selection, and can tell a reference photo from a generated variation. It works through a large board in sections, up to 10,000 assets. Nothing moves until the person confirms, and a run can be reversed without losing later edits.
Why it mattersThe pile of generated frames can be filed by the same system that made them, and undone if the filing is wrong.
Luma ↗One table is scored on filling in, predicting and generating
The One Table Benchmark runs generation, imputation and prediction on the same 56 datasets, with 46 registered methods and 20 capability axes, under shared splits. The authors train five backbones under every non-empty mix of those objectives. Rank comparisons tie imputation to prediction.
Why it mattersA model that is good at guessing a missing cell is not automatically good at inventing a new row.
arXiv ↗I want to keep coming back to what was the culture at Nvidia, and were they pressing these seemingly illicit sales to goose their revenues and their earnings?
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.
The first AI Music Awards ran late and lip-synced
Variety's account of the 14 November ceremony at the Mayan Theater in Los Angeles says tracks had to be on a Verified Registry as fully AI-generated, AI-assisted, or a human-and-AI hybrid with a cloned lead voice. Xania Monet, the AI persona of Telisha Jones, won artist and song of the year. The show started about 30 minutes late, hosts shared two microphones, and a Deepgram introducer often failed. A 14-year-old lip-synced the vocal for one song.
Why it mattersAn awards show for machine-made music still needed a person on stage when the vocal was played.
Variety ↗Adobe shows research tools that edit across a whole scene
Adobe's MAX Sneaks post, dated 11 November, previews Scene Sweep, which carries one edit across characters, locations, cuts and camera angles, and Live Set, which changes the subject, the place or the camera while a scene plays. Screen Test builds a fictional character's look, voice and personality. On Cue times an action from a drawn signal or from video or music. Some of the previews are in the Adobe Labs playground.
Why it mattersAn edit that used to be redone on every shot is being tried as one instruction for the scene.
Adobe ↗Shortical premieres a series starring an actor's AI double
TheWrap reported on 13 November that Inevitable, a fully AI-generated microdrama, premiered that day on the Shortical app, with 24 of 50 episodes and the rest weekly. It stars an AI version of actor Aki Avni, who helped make the double. Director Ofir Lobel said the future is AI with actors, not instead of them. Sharon Ben Tovim wrote it.
Why it mattersA performer is in the credits of a series that uses his likeness, and the episodes are still arriving.
TheWrap ↗Nvidia's venture arm asks out of a Suno deposition
Music Business Worldwide reported that NVentures, a November 2025 Series C investor in Suno, filed a motion on 10 November in the Northern District of California to cancel or limit a deposition sought by Universal and Sony. It said it has produced 6,324 documents and has no board seat. The page says responses are due 24 November. Nvidia Corporation filed its own motion the same day.
Why it mattersA shareholder is arguing that writing a cheque does not make it a witness to how the model was trained.
Music Business Worldwide ↗Mobiquity ties an AI campaign tool to a music distributor
Mobiquity said on 11 November that a 12-month exclusive with Intercept Music followed a LOCASH campaign it credits with more than 8 million streams. Spotify monthly listeners rose from about 729,000 to 1.2 million, and Pandora from about 200,000 to more than 700,000. Intercept's network is more than 1,000 artists. The release illustrates $3 million to $6 million a year if a quarter of them pay, and says that figure is not a forecast.
Why it mattersA distributor is being offered a monthly fee for an agent that buys the audience, with the sample size still one campaign.
GlobeNewswire ↗Film sellers at AFM say YouTube is carrying the indie slate
Media Play News covered a 12 November panel at the American Film Market. Linda Nelson of Indie Rights said YouTube passed Amazon earlier this year as a revenue source and that 98 percent of her channel's views come from nonsubscribers. Luiz Bannitz said AI makes dubbing cheaper but still misses language and culture, so he prefers people for higher-end films. Romeo Duchene said a 1996 Lionsgate title has made more than $100,000 through that YouTube distribution.
Why it mattersThe catalogue business is describing YouTube as the till, and still hiring people when the dub has to sound right.
Media Play News ↗A Los Angeles Times column maps AI jobs on a film set
A 15 November Los Angeles Times piece describes Innovative Dreams and Wonder Project, with Luma and AWS, showing a workflow at MBS Media Campus that they call final-pixel and production-ready. It notes Netflix's purchase of Ben Affleck's InterPositive, and quotes Netflix's Elizabeth Stone that the tools should help storytellers rather than replace them. It also lists new job titles, including prompt director and model rights manager.
Why it mattersThe credits on a production are gaining roles whose job is to say what the model may use.
Los Angeles Times ↗Sonify pairs a generated picture with a music buyout
Authentic Music Business Intelligence said Sonify's 13 November note includes Video Lab, which pairs AI-generated visuals with original music under SoniBuyout certificates. Video Lab was already running. A buyout can cover a later recut, depending on the agreement. Separate licenses exist for venues and DJs.
Why it mattersA track and the picture that goes with it are being sold as one cleared package.
Authentic Music ↗A benchmark checks motion graphics against the file, not the eye
MotionGraphicsBench is 100 expert-written cases for animating a static SVG from a text instruction. The pictures are already supplied, so the test is the motion, not whether the model can draw. Each case is checked against the motion design rather than only by looking at the frames. The paper is dated 7 November.
Why it mattersA moving graphic can be marked wrong because the file does not match, even when the clip looks fine.
arXiv ↗Chips & compute
Chips, data centres and the power behind them.
TSMC sets aside $500 million for the next chip startup
Taiwan News reported on 13 November that TSMC is setting up a NT$16.45 billion, about $500 million, fund for chip startups, after Kevin Zhang told Nikkei Asia the company wants to support the future's Nvidia as well as today's customers. Zhang said the fund could also back suppliers and intellectual-property vendors, which TSMC has not invested in before. The same piece says TSMC has moved to invest up to NT$32.87 billion in Anthropic and OpenAI.
Why it mattersThe foundry that builds other companies' chips is putting its own money into the startups that might design the next ones.
Taiwan News ↗Cerebras posts $230.9 million of core revenue
Cerebras said on 12 November that third-quarter core revenue was $230.9 million, up 70 percent from a year earlier, and GAAP revenue was $210.8 million, up 55 percent. It guided to $310 million to $330 million of core revenue for the fourth quarter and said it expects to triple revenue in 2027. Remaining performance obligations were $26.1 billion, and 700 gross megawatts of data-center capacity are under contract for delivery by the end of 2027. The CS-5 system is targeted for early summer 2027.
Why it mattersA second source of very fast inference is telling investors the backlog is already measured in years of power.
SEC ↗DeepX finishes the design of an on-robot language chip
The Elec reported that DEEPX has finished the DX-M2 design, to be made on Samsung's second-generation 2-nanometre process, with mask tape-out in early November. First silicon is planned for March 2027, samples in the second quarter and initial shipments in the fourth quarter of 2027. The chip is aimed at language and vision-language models at about 5 watts, for robots and industrial machines.
Why it mattersA robot's language model is being drawn onto a chip that is supposed to run on a few watts, not a rack.
The Elec ↗Tendrils raises $30 million to redo the CPU
Tendrils said it raised $30 million led by Benchmark, with Visionaries Club, the UK Sovereign AI Fund and others, and that Jack Altman is joining the board. The company says it is building a processor with thousands of parallel cores, without shared mutable memory or a linear instruction stream, and an open-source language called Vine next to the chip. The post does not name a fab or a tape-out date.
Why it mattersThe bet is that the part of the computer that is not a GPU still has to change if agents are going to run on it.
Tendrils ↗Epoch puts Huawei's Horinger cluster at about 7 percent of Colossus 2
Epoch AI estimates Horinger is running about 105,000 Ascend 910C chips and 8,000 Ascend 910B chips, which it equates to 81,000 H100s, with a 90 percent interval from 36,000 to 141,000. It describes a cooling load of 286 megawatts of IT power across four buildings, and one western building not yet in operation. Against SpaceX's Colossus 2 in Tennessee, Epoch says Horinger has about 7 percent of the compute.
Why it mattersA published estimate now puts a number on how far a flagship Chinese cluster sits behind the largest US one.
Epoch AI ↗Thinking Machines trained Inkling 2 Small with a 4-bit forward pass
Thinking Machines said on 13 November that expert-layer forward passes for Inkling 2 Small used NVFP4, with MXFP8 on the backward pass. The lab called it the largest model it knows of that was trained end to end with a native 4-bit forward pass, from pretraining through the final reinforcement-learning rollout. On the largest jobs it reported 8 percent faster pretraining and 20 percent faster reinforcement learning than full MXFP8, measured by the clock.
Why it mattersThe training run, not only the served model, is being squeezed into the 4-bit format the newest Nvidia chips prefer.
Thinking Machines ↗CNBC says Nscale's updated prospectus shows a wider loss
CNBC reported on 12 November that Nscale's amended IPO filing shows a third-quarter net loss of $1.69 billion on $307.5 million of revenue, up 1,344 percent from a year earlier. The loss widened from $1.02 billion in the second quarter, when revenue was $140.6 million. The company has not said how much it plans to raise. CNBC said Anthropic, in late September, exercised a right to more computing power at an Nscale site in West Virginia.
Why it mattersA company renting out Nvidia chips is heading toward a listing with a loss that grew as the revenue did.
CNBC ↗ExLlamaV3 1.6.7 pipelines the prefill
ExLlamaV3 1.6.7, tagged 13 November, says it fixes split-device quantization, improves checkpointing, pipelines the prefill, and fixes a disk tier shared by several instances. It also tightens expert routing and sparse attention indexing.
Why it mattersA local runtime used for large mixture-of-experts models is changing how the first pass over a long prompt is scheduled.
GitHub ↗A day-long test changes which infra agent looks best
Infrathon scores language agents on long infrastructure-tuning jobs. Six models were run 150 times with 24-hour budgets, more than 10,000 GPU-hours in total. Rankings changed with the length of the test, and 135 of the 150 runs still improved in the second half of the budget. The authors say shorter tests miss both the gains and the change in who is ahead.
Why it mattersAn agent that looks best after an hour of tuning a cluster may not be the one ahead at the end of the day.
arXiv ↗Cerebras’ technology remains the undisputed champion of speed.
Policy & safety
Laws, lawsuits and safety incidents.
Bessent wants Nvidia's next chief hired from outside
Treasury Secretary Scott Bessent said on Fox News on Sunday that Nvidia needs a dramatic shakeup and that the board should look outside the company for its next chief executive. He asked whether the culture had pressed seemingly illicit chip sales to lift revenue, and said that if it were a bank that far out of compliance, the comptroller would already be removing executives. Nvidia reports earnings on Tuesday, its first since Jensen Huang resigned on Thursday. Colette Kress is interim chief executive.
Why it mattersThe Treasury secretary is questioning the sales culture two days before the first earnings report without Huang.
CNBC ↗Megaspeed says Alice Huang left in August 2025
Megaspeed told CNBC on Sunday that Alice Huang officially left on 30 August 2025 and has no current role, and that it does not comment on her personal matters. The company said recent reports about its business and supply chain are not supported by verifiable evidence, and that it has hired outside lawyers for an internal investigation. It said its controls are built to meet US export rules and local rules in each country where it operates.
Why it mattersThe company at the center of the chip-diversion claims is putting a departure date on the record and opening its own inquiry.
CNBC ↗The US says North Korean deepfakes took more than $400 million
The Justice Department said on 12 November that it seized cloud infrastructure used in FaceStealer, which it says stole more than $400 million in cryptocurrency over 18 months using secretly recorded video and deepfakes. The FBI seized more than 900 victim webcam videos. Separately, a court authorized seizure or restraint of more than $1.2 billion taken from Binance on or about 29 October, out of more than $2.7 billion frozen. Targets include crypto workers and, increasingly, open-source maintainers.
Why it mattersA theft that starts as a video call is now a named case, with the recordings in evidence.
US Department of Justice ↗Microsoft says the global outage was not caused by AI
CBC, citing Thomson Reuters, reported on Saturday that Microsoft said Teams, Outlook and other Microsoft 365 services had fully recovered after an outage the broadcaster described as more than 24 hours. Microsoft blamed a technical issue from a network capacity upgrade, said it was not aware of malicious activity, and ruled out AI as the cause. Reports started climbing on Thursday and hit governments, hospitals and businesses.
Why it mattersA day-long failure of the tools people use to work was blamed on a capacity change, and explicitly not on an agent.
CBC News ↗The FAA says its AI advisor did not ground the planes
The FAA said a nationwide ground stop began about 15:30 ET on 13 November and traffic resumed at 20:15. Legacy equipment became unstable when data hit the network at about 1,000 connection attempts every 4 to 5 seconds, during router work by L3Harris at sites in Amarillo and El Dorado. The agency said there is no indication of a cyberattack, and that the SMART system did not cause it because SMART does not feed the air-traffic system. About 3,500 airborne aircraft landed safely.
Why it mattersAn AI tool that advises controllers was in the rumor, and the agency says the failure was a router.
Federal Aviation Administration ↗Italy fines Meta €805 million over phone photos
Italy's data protection authority fined Meta Platforms Ireland €805 million for Facebook on Android reading photos, videos and metadata on Italian phones without a specific consent. The feature, tips from the gallery while browsing, has been available since 2023 and was on by default. The authority said this case is under the ePrivacy rules and did not cover AI training, which Meta denies. About 20 million Android users in Italy are affected. Meta has 60 days to require a separate consent and to default the feature off.
Why it mattersA fine this large is about the photos already on the phone, and the authority has set the AI-training question aside.
Garante Privacy ↗Anthropic tells labs how to test whether a model did the right thing
Anthropic published guidance on 13 November for safety evaluations, which it defines as tests of whether a model did the right thing, not of a raw capability. On the API, it said recent models' harmless response rates on single-turn tests sit between 94 and 98 percent. It recommends at least 100 items for a complex multi-turn test and 1,000 for a simpler single-turn test, and says a romance-scam comparison of Claude Opus 5 and Opus 5.5 was not statistically significant.
Why it mattersA lab is publishing the sample sizes at which its own safety scores stop meaning much.
Anthropic ↗CSM-1 is a safety model that is supposed to refuse attack tools
OpenModel Labs posted CSM-1-Preview on 14 November, a fine-tune meant for detection, incident response, hardening, secure coding and static malware analysis. The card says it is designed to decline requests for ready-to-use attack tooling, that the LoRA was merged into bf16 weights, and that users should check CVE identifiers because the model can be wrong.
Why it mattersA model sold for defense work is being explicit about the requests it is supposed to turn down.
Hugging Face ↗Washington and Singapore put Pax Silica on a meeting note
The State Department said on 13 November that Deputy Secretary Christopher Landau met Singapore's minister for digital development, Josephine Teo. They discussed trusted digital infrastructure, super intelligence, cybersecurity and secure technology supply chains, including through Pax Silica, ahead of Singapore's 2027 ASEAN chairmanship. The readout does not publish a list of controlled technologies.
Why it mattersA supply-chain pact that has been discussed in the press is now named in an official bilateral note.
US Department of State ↗If this were a bank and they hadn't complied like this, the OCC would be in there right now, and a lot of the executives would be removed.
Robotics
Robots, self-driving and AI in the physical world.
OpenAI breaks up its standalone robotics team
Fortune reported on 13 November that OpenAI confirmed it disbanded the standalone robotics team and reassigned the researchers. Research now reports to chief research officer Mark Chen, and applied robotics and hardware report to president Greg Brockman. A spokesperson said the change is meant to speed the work. Fortune says OpenAI's account is that general models, including GPT-6 Astra, can be applied to robots directly.
Why it mattersThe lab is betting that a general model can drive a robot without a separate robotics group in the middle.
Fortune ↗Tesla opens Cybercab pre-orders, inside a fence and for a fee
TechCrunch reported on 14 November that Tesla is letting select customers pre-order a Cybercab for personal use at $29,990, plus $299 a month to operate it. Marques Brownlee said delivery has to be accepted in one of a handful of Texas cities, operation is geofenced, and the estimate is December 2026. Elon Musk said it will be available nationwide as regulatory approval arrives, and called the monthly fee too high.
Why it mattersThe car with no wheel is for sale to a few people, and it only drives where the company has already switched the fence on.
TechCrunch ↗Robbyant open-sources a touch model for robot hands
Robotics and Automation News reported on 15 November that Robbyant's LingBot-Touch, from Ant Group, was pretrained on nearly 5,000 hours of real tactile data across magnetic, visual and piezoresistive sensors. On 26 real-world tests with two hands, four action models gained about 14 to 26 percentage points of success when the touch representation was added. The light version used in those tests has about 5.3 million parameters. Weights and code are public.
Why it mattersA hand that can already see is being given a shared way to feel, small enough to sit next to the policy.
Robotics and Automation News ↗FAA expands the SMART advisor to all 22 en-route centers
The FAA said SMART, after a September start in the Washington area, will be used in the airspace of all 22 en-route centers. It pulls together 200 data streams, including weather, flight paths, traffic and staffing. Specialists in Washington check the data and send recommendations to the command center in Warrenton, Virginia. Since 21 September the agency counts 180 flights rerouted, about 10 minutes and almost 80 gallons of fuel saved per flight.
Why it mattersThe tool that suggests a reroute is going national, and a person still has to accept the suggestion.
Federal Aviation Administration ↗Nidus raises $30 million to teach factory robots a wire harness
Nidus said on 10 November that it raised $30 million backed by Kleiner Perkins and Costanoa, with Ludlow Ventures, SF1, Absolute Partners, Tau and others. The New York company builds models that let existing industrial robots do precision assembly, starting with the wire harnesses in aircraft, missiles and vehicles. It said revenue passed seven figures in the first four months. Chief executive Justin Sealy said each customer's data stays with that customer.
Why it mattersThe robot already bolted to the factory floor is being given a model for the fiddly job people still do by hand.
Nidus ↗A plain harness lifts a robot arm that GPT-6 Astra drives
JustRUA takes apart the harnesses used when GPT-6 Astra drives a robot arm, and retests them in one framework. A lean harness that keeps the full episode and stops each motion early accounts for much of the gain. Scene geometry helps most, even when depth is predicted from a color camera. The authors say this simple harness, with no learned policy, beats the official L3 setup it starts from.
Why it mattersThe jump in robot scores may be the scaffold around the model, not a new model.
arXiv ↗One policy trains on human, robot and simulated hands
EgoX lines up demonstrations from people, real robots and simulation by matching what each can see and do, then pretrains a shared transformer with a separate tokenizer per sense and a separate expert per action. A training-only objective predicts future sight and touch. The model is fine-tuned on the robot it will actually run.
Why it mattersFootage from a person's headset can be used only if the model knows that person does not have the robot's joints.
arXiv ↗A release pairs 1,589 room scans with 4,230 first-person recordings
Ropedia Xperience 2 argues that robot models need both demonstrations of how people do a task and places where a policy can be tried in a loop. The release includes 1,589 metric 3D scans and 4,230 egocentric recordings. The authors say frontier models that drive robots from pixels are still brittle on long, contact-heavy tasks.
Why it mattersThe missing piece being published is the room and the recording, not another policy network.
arXiv ↗A first exam lets a coding agent pick how it drives the robot
The Embodied First Exam is a test where a coding agent may drive a robot directly, write code that drives it, or train a policy. It has 1,156 tasks in 35 suites on 47 robots, including 9 new suites. The paper is dated 9 November and says existing robotics benchmarks were built for visuomotor policies, not for this choice.
Why it mattersThe agent is being graded on whether it picks a sensible way to move the robot, not only on whether the arm arrives.
arXiv ↗Science & health
AI in biology, medicine, climate and maths.
Charta Health raises $55 million to read a chart before the claim
Charta Health said it raised a Series B led by Insight Partners, with Madrona, Bain Capital Ventures, South Park Commons and QuantumLight. The product reviews a chart before a claim is sent, and a person approves the findings inside the record. Customers named include American Family Care, Grow Therapy and Pediatric Associates. The company says fee-for-service customers see a double-digit drop in denial rates, and a payer suite starts with Medicare Advantage and ACA marketplace plans.
Why it mattersThe same chart read is being sold to the clinic and to the insurer, before the bill goes out.
Charta Health ↗A startup ships an offline model that scores a pathogen
WIRED reported on 11 November that Radical Numerics announced Omnii Defense, an edge model meant to flag natural and AI-designed pathogens on a laptop with a USB sequencer, including with no network. Chief executive Eric Nguyen said biosecurity was far behind on defense. The tool scores pathogenicity, whether a sequence looks machine-made, and taxonomy, then a threat score from zero to one. Access is limited to vetted testers. WIRED says George Church resigned as an adviser last week.
Why it mattersThe check is designed to run next to the sequencer, not in a cloud account that a lab may not want to use.
WIRED ↗A lattice definition of the Standard Model, written with Astra
OpenAI's science blog on 15 November carries a guest post by Nathaniel Craig, a visiting scientist there and a particle theorist at UC Santa Barbara. He describes an October preprint, prepared with GPT-6 Astra, that proposes a candidate lattice definition of the Standard Model. Craig writes that he had the strategy but had not proved it covered every case, and that Astra turned it into a construction and led the detailed proof. He calls the result a candidate, not a finished simulation.
Why it mattersA fifty-year gap in how the theory is defined is being claimed as a proof a model helped write, still waiting on other physicists.
OpenAI ↗OpenAI opens a blog for scientists using its models
OpenAI said on 15 November that it is starting a science blog for results, methods and accounts from researchers. The post points to an internal system that in September produced a proof aimed at the Navier-Stokes existence problem, and to October results on the quasi-Riemann hypothesis. It also says thousands of agents using GPT-Rosalind, with PerturbAI and Scripps, analyzed data from millions of mouse brain cells in 32 minutes.
Why it mattersThe lab is giving scientific claims a front door of their own, separate from the product blog.
OpenAI ↗Antigravity's teamwork feature is pointed at proofs and spectra
Google Research said on 13 November that Antigravity's teamwork feature, using Gemini 4 Argon, Carbon and Hydrogen, formalized the 2013 Courtade-Kumar conjecture in Lean 4, in 50 million lines, and a formalization of the Weil conjectures in about 1.8 million lines across 28,800 subagent calls. On an NMR structure task, 30 agents reached 95 percent top-1 exact match. Authors of 578 STOC 2027 submissions opted into an automated review.
Why it mattersA coding agent that can call a crowd of subagents is being used on proofs and on chemistry, not only on repositories.
Google Research ↗A vision-language model reads a full brain scan
Neuro-VL is trained on nearly one million neuroimaging studies from routine care, and is meant to read a whole examination and talk about the findings. The authors say that on internal and international sets it outperformed frontier general models and specialized medical models at diagnosis and localization. The paper is dated 11 November.
Why it mattersThe scan a hospital already stores is being treated as a conversation, not only as a label.
arXiv ↗A system picks the next experiment from what it already believes
Sage-Taiyi keeps an explicit account of evidence, assumptions and open questions, and uses that to choose the next target and the method. On Taiyi-Bench, 265 tasks drawn from 50 papers across ten fields, the authors say it ranks first on all eight metrics they report, with an overall score of 54.0 against 47.9 for AI Scientist-v2. The paper is dated 9 November.
Why it mattersThe claim is not a new answer to one problem, but a system that decides which problem to attempt next.
arXiv ↗A model can answer without writing, or spend a set number of tokens
The paper trains one backbone to do two things: answer from a single forward pass, writing no tokens, or run a full chain of thought of T tokens. A small block estimates whether the slow mode is worth a price set per request, and runs it only then. The paper is dated 12 November.
Why it mattersThe choice is no longer how long to think, but whether to think in words at all.
arXiv ↗A benchmark for models that decide in one pass
SO-Arena tests System One models, which answer without a long chain of thought, across 11 domains and five families of skill. Labels are checked against sources the model can see. It also measures whether the decision stays put when the option list is changed. The paper is dated 11 November.
Why it mattersA fast answer can be scored on whether it flips when the choices are reordered, not only on whether it was right once.
arXiv ↗Astra helped turn that strategy into a construction and took the lead in developing the detailed proof.
Open source
Open-weight models and community releases.
ZONOS3-Mini clones a voice from 100 million active parameters
Zyphra released ZONOS3-Mini on 12 November under Apache 2.0, a text-to-speech model with 100 million active parameters and 800 million in total, for 43 languages. It was trained in 2.5 days on 112 H100s. The card lists time to first audio under 570 milliseconds in a browser on a Pixel 9, under 150 milliseconds on an M3 MacBook Air, and 80 milliseconds median on one H100. Runtimes include SGLang, MLX, ggml and WebGPU.
Why it mattersA clonable voice that fits a phone is being posted with the code that runs it, not only the weights.
Hugging Face ↗Kitten TTS 2 Mini speaks 16 languages from 82 million parameters
KittenML posted Kitten TTS 2 Mini on 13 November, an 82 million parameter model with 118 voices at 24 kHz across 16 languages, under Apache 2.0. The ONNX build uses int8 weights and is 82.7 MB, so it runs on CPU without PyTorch. Install is pip install kittenml.
Why it mattersA small speech model with a one-line install is aimed at machines that do not have a GPU.
Hugging Face ↗Diogenes 0.2 keeps facts in a memory beside a 55 million parameter model
Counfield posted an early Diogenes 0.2 checkpoint on 15 November, a 55 million parameter byte-level model trained from scratch together with a content memory and a reader of about 858,000 parameters. The card says this is step 267,000 of a planned 380,000, with no pretrained initialization and no tokenizer. The license is MIT. The design claim is that the language model itself does not store the facts.
Why it mattersA tiny model is being trained on the assumption that remembering and speaking are different parts.
Hugging Face ↗llama.cpp reads Higgs and VibeVoice as speech
llama.cpp build b12426, tagged 15 November, adds mtmd support for Higgs v3 speech-to-text and VibeVoice ASR. The note says it uses a proper mapping and replaces a hack for audio length with a helper that reads media metadata.
Why it mattersTwo more speech models can be run through the same local engine people already use for text.
GitHub ↗llama.cpp reads a smaller Qwen3-VL output head
Build b12425, tagged 15 November, reads the Qwen3-VL output head size from the tensor. The note says this allows Qwen3-ForcedAligner, whose output head of 5,000 timestamp classes is smaller than the vocabulary.
Why it mattersA speech-alignment head that does not match the word list can load instead of being rejected.
GitHub ↗llama.cpp speaks with ZONOS3-Mini
Build b12424, tagged 15 November, adds the ZONOS3-Mini text-to-speech model to mtmd, including a default voice stored in the file, voice cloning, and text normalization.
Why it mattersThe small open voice model can be called from the same local runtime as the chat models.
GitHub ↗TensorFold 1.2.9 shares a decode round across streams
TensorFold 1.2.9, tagged 15 November, says every speedup keeps the replies the same. On an M5 Ultra, Nemotron without an MTP head shares one decode round across streams: two streams about 20 percent faster, four about 75 percent, eight about 2.3 to 2.5 times. It also adds an Ollama-compatible API and fixes a wrong reply and a hang.
Why it mattersSeveral local chats can share one trip through the model instead of each waiting its own turn.
GitHub ↗TensorFold 1.2.8 adds DeepSeek-V4.1 on a Mac
TensorFold 1.2.8, tagged 13 November, adds Recursive's Flint 40B on M5-class Macs, an experimental native path for DeepSeek-V4.1 on Macs, Qwen3 embedding and reranker families on CUDA, and OCR through a POST /v1/ocr route on HIP. A request can be cancelled by id, and bodies, templates and grammars get fixed limits.
Why it mattersA desktop runtime is picking up a frontier open model and a way to stop a request that has run too long.
GitHub ↗Zeroing a tiny slice of weights can match a full fine-tune
Deletion Tuning learns a binary mask over the gate projection and sets those weights to zero, changing fewer than 0.05 percent of them. The authors say that matches full fine-tuning across reinforcement learning with verifiable rewards, supervised fine-tuning, and on-policy distillation. The mask is hundreds to thousands of times smaller than the fine-tune checkpoint, and the method forgets less. The paper is dated 7 November.
Why it mattersA capability that looked like it needed a new copy of the model may be a handful of weights turned off.
arXiv ↗Quick hits
Smaller stories worth a glance.
- Phosphene puts Move to another drive where people lookRelease 5.0.22 says the control for keeping models on an external drive is back where the Mac app's menus lead you to look.GitHub ↗
- Phosphene 5.0.23 chases the renders that failedThe notes say the team went through recent failed renders and fixed the causes that were theirs.GitHub ↗
- A new Cloudflare project defaults to the 12 November compatibility datecreate-cloudflare 2.74.12 sets the default compatibility date to 2026-11-12, matching the bundled workerd.GitHub ↗
- LangChain core lets one context pick a different gatewaylangchain-core 1.8.1 adds a per-context override for the LangSmith gateway.GitHub ↗
- llama.cpp's Jinja templates gain four featuresBuild b12422 adds four Jinja features that the template engine had been missing.GitHub ↗
- A LiteLLM candidate still verifies with the pinned keyRelease 1.110.0-rc.3 is signed with the key introduced after the October incident, unlike the keyless 1.111 builds.GitHub ↗
- LiteLLM tells people to skip a half-published tagThe 1.111.0-dev.3 notes say the tag was only partly published and to use 1.111.0-dev.4, built from the same commit.GitHub ↗
- LanceDB beta.14 takes server-issued listing tokensThe beta.14 notes say SQL catalog listings use tokens the server issues, and functions can declare nullable scalars.GitHub ↗
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.
Sequoia backs Edelbyte
- Raised
- $70M
Edelbyte valuation not disclosed
Edelbyte said Sequoia and Quiet Capital led $70 million across a seed and a Series A for search and analytics aimed at AI applications.
QCQuiet CapitalGGreenoaksSVSwish Ventures+0 more
Insight Partners backs Charta Health
- Raised
- $55M
Charta Health valuation not disclosed
Charta Health said Insight Partners led a Series B for software that reviews a clinical chart before a claim is sent.
MMadronaBCBain Capital VenturesSPSouth Park CommonsQQuantumLight+0 more
Kleiner Perkins backs Nidus
- Raised
- $30M
Nidus valuation not disclosed
Nidus said it raised $30 million, backed by Kleiner Perkins and Costanoa, to put models on industrial robots doing precision assembly, starting with wire harnesses.
CCostanoaLVLudlow VenturesSSF1APAbsolute PartnersTTau+0 more
· Nidus ↗
Benchmark backs Tendrils
- Raised
- $30M
Tendrils valuation not disclosed
Tendrils said Benchmark led a $30 million round for a processor with thousands of parallel cores and an open-source language called Vine.
VCVisionaries ClubUSUK Sovereign AI Fund+0 more
Blumberg Capital backs WhiteMatter
- Raised
- $20M
WhiteMatter valuation not disclosed
WhiteMatter said Blumberg Capital led a $20 million Series A for a guardian agent assigned to each deployed agent.
CCCrew CapitalS&Sprout & Oak+0 more
Partnerships & contracts 1
A large commercial agreement, or a government contract.
TSMC sign a deal
- Value
- $500M
Multi-year agreement
Taiwan News reported that TSMC is setting up a fund of about $500 million for chip startups and suppliers, and has moved to invest up to NT$32.87 billion in Anthropic and OpenAI.
On the calendar
What's coming up, in date order, and how sure we are about each.
- 9–20NovCOP31 climate conference in AntalyaToday
- 17NovThird-quarter results and earnings callTomorrow
- 17NovTomorrowX SummitTomorrow
- 18–19NovTrump and Xi meet in Shenzhen2days away
- 18NovWebinar on running a dock with Stretch2days away
- 18NovNvidia invited to brief Congress2days away
- 18Nov‘Dot’ agents become the default in workspaces2days away
- 19NovKinference demo livestream3days away
- 19NovClaude Data turns on for Enterprise3days away
- 20NovAI Film Award submissions close4days away
- 20NovConductor cloud service shuts down4days away
- 20NovUpdated OpenAI usage policies take effect4days away
- 20NovAion Instruct reaches Windows Insider devices4days away
- 21NovInaugural edition in Hollywood5days away
- 22Nov'The Challenge Sverige' originally scheduled (now delayed to 2027)6days away
- 23NovIn-person registration closes for the FDA robot workshop7days away
- 25NovQwen4-Max preview ends9days away
- 26NovHuang's deadline to testify10days away
- 26NovShort-film contest submissions close10days away
- 29NovDaybreak Blue sessions can be used for training13days away
- 30–3Nov–DecContent London 20262weeks away
- 30NovRose d’Or Awards2weeks away
- 30NovGemini 4 Carbon introductory price ends2weeks away
- 30NovKorea AI Advertising Awards entries close2weeks away
- 1DecHearing on Apple's bid to block testimony2weeks away
- 2–3DecWorkshop on autonomous surgical robots2weeks away
- 3DecPotential winners of the film contest2weeks away
- 4DecAI-made film opens in limited release3weeks away
- 4DecGods Don’t Give Gifts premiere3weeks away
- 4DecSingapore AI Safety Fellowship ends3weeks away
- 9DecBarclays Global Technology Conference3weeks away
- 10DecIntelligent eyewear goes on sale3weeks away
- 16DecKorea AI Advertising Awards winners4weeks away
- 16DecRing Rover goes on sale4weeks away
- 22DecKorea AI Advertising Awards ceremony5weeks away
- 31DecHy4 half-price offer ends6weeks away
- 31DecJules shuts down6weeks away
- 1JanFrontier labs must register under the RAISE Act7weeks away
- 27JanRing indoor security drone goes on sale10weeks away
- Later in NovGuided Learning pilot in US schools
- Mid-NovXREAL Aura shipments begin
- DecemberNothing's first AI-native product
- 2027Two Nighthawk r3 processors linked
- H1 2027Digit 5 early customer access
- Nov–Dec 2026Fort Benning robotaxi pilot
- Later this yearAge checks reach Threads
- 2028R200 systems ready to deploy
- through Feb 2028Language Explorer in Real Talk
- early 2027Firefly always-on agents enter beta
- End of NovExpo to detail Agent Sessions
Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.






