Sunday, 15 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.
- every Gemini app user, on web and mobile
- US
- cannot move, manage or invest money
- Read-only
- optional link for banks and cards
- Plaid
- the Wallet connection can be turned off
- Opt in
Gemini can read a US wallet, and cannot spend from it
Google said on Friday that every Gemini app user in the United States can connect Google Wallet, on the web and on Android and iOS. Tagging @Google Wallet lets someone ask about spending, subscriptions and a holiday budget. Linking banks and cards through Plaid is optional. If they opt in, Gemini can only read the accounts. It cannot move, manage or invest the money, and the connection can be switched off.
Why it's the story of the dayA budget answer can start from real charges, and the assistant still cannot pay them.
Startup of the day
A young company people are talking about today.
Luminal raises $25 million to compile models onto the chips a company already has
Standard Capital said on 12 November that it led a $25 million Series A in Luminal. Dalton Caldwell wrote that Luminal builds a compiler that searches for an efficient way to run a model on existing hardware, using equality saturation to compare equivalent versions of a computation. Founders Joe Fioti, Matthew Gunton and Jake Stevens are named. Customers can use a hosted inference service or license the software to run on their own machines. The post says the work is meant for GPUs and specialized AI chips, not only Nvidia.
Why it's worth a lookThe round is for the software between a model and the chip, which is where a lot of bought hardware sits idle.
Standard Capital ↗Models & labs
New models, the labs behind them, and the rules they answer to.
Gemini can read a US wallet, and cannot spend from it
Google said on 13 November that every Gemini app user in the United States can connect Google Wallet, on the web and on Android and iOS. Tagging @Google Wallet lets someone ask about spending, subscriptions and a holiday budget, and they can link banks and cards through Plaid. If they opt in, under Connected Apps in Personal Intelligence, Gemini can only read the accounts. It cannot move, manage or invest the money, and the connection can be turned off. The same post adds Memory Summary, a page of saved preferences and goals that can be edited.
Why it mattersA budget answer can start from real charges, and the assistant still cannot pay them.
Google ↗DeepSeek opens search, files and voice on its API
DeepSeek's changelog for 13 November says deepseek-flash can search the web or images, read a URL, take PDFs, Office files and images, and return a refusal marker when a safety check is on. Search, Fetch, OCR, ASR and TTS are also separate APIs. A public service has to turn the safety check on and meet content-review, model-registration and labeling rules. The company open-sourced a demo chat client that can import DeepSeek chat history.
Why it mattersThe app's search and file tools are now something a developer can call, not only a button in the product.
DeepSeek ↗Fish Audio's Live 3 listens while it speaks
Fish Audio introduced Live 3 on 10 November, a full-duplex voice model trained on real conversations so it can acknowledge, wait, answer or stop. The company said it can decide to respond in as little as 186 milliseconds. It speaks English, and it is on the Fish Agent Platform, through the API, and for on-premises enterprise deployment.
Why it mattersA voice agent can leave room for a pause in a phone number instead of treating every silence as the end of the turn.
Fish Audio ↗Transformers 5.23 loads LFM3, Dots 3 Note and IQuest-Coder
Transformers 5.23.0, tagged 13 November, adds Liquid AI's LFM3, a hybrid of gated short convolutions and grouped-query attention, and Dots 3 Note Preview, a mixture-of-experts model that takes text, images, video and audio. IQuest-Coder-V1 is loaded through the Llama classes in 7B, 14B and 40B sizes. The looped IQuest checkpoints are not supported. InternLM3-8B-Instruct is also loaded as Llama, with its own config.
Why it mattersA new batch of architectures can be called through the same library, including one that takes video and audio in one model.
GitHub ↗An experimental JEV head scores a one-token label
The same Transformers release describes JEV as a decision head that scores single-token labels with a linear layer on the last token of the prompt. JevForDecision reads head.safetensors and judge_config.json shipped beside the weights. The note says the head and the decision pipeline are experimental and may change.
Why it mattersA classifier can return a scored label without asking a chat model to write the answer out.
GitHub ↗MiniMax H3-Image draws a 4K frame in under 10 seconds
MiniMax released H3-Image on 11 November, an open-weight model for text-to-image, reference images and editing. The post says a 4096 by 4096 image can be generated in under 10 seconds, with up to 16 reference images. Weights are on Hugging Face, and the model is also on the MiniMax API, MiniMax Code and Hailuo AI.
Why it mattersA 4K edit with many reference pictures is being offered as an open-weight model, not only as a closed app.
MiniMax ↗Harbor turns a Mac into a private model server
MacRumors reported on 10 November that Apple's Claris subsidiary is testing Harbor, which runs open-weight models on an Apple silicon Mac. The product text says data never leaves the hardware and nothing reaches the cloud. Several Macs can be linked over Thunderbolt 5 to pool unified memory. Harbor is in early access and is not publicly available, and MacRumors said an App Store listing has been removed.
Why it mattersA team can try a local server for open weights on machines it already owns, before the product is on the store.
MacRumors ↗A learned energy solves puzzle benchmarks after a brief training run
The authors train CRAM, a learned version of Hopfield and Tank's energy, on reasoning puzzles. They report that it solves 99.996 percent of Sudoku-Extreme and 100 percent of Maze-Unique after 42 ms and 132 ms of training on one GPU, at least 1,000 times faster than the looped models they compare on those tasks. They also say it matches those models' highest Sudoku-Extreme accuracy with at least 282 times fewer inference FLOPs.
Why it mattersOn these two puzzles, the energy method is doing the job with a very short training run.
arXiv ↗Protein language models break when the amino-acid names are swapped
A 12 November paper relabels the amino-acid alphabet with a global bijection that keeps alignment information, column entropies and pairwise edit distances the same. Pseudolikelihood Potts models and inverse-covariance estimators stay invariant. Protein language models fail the test. The authors say a high contact-map score is not, by itself, evidence that the model learned the same coevolutionary signal as an alignment method.
Why it mattersA contact map can look right even when the model is using the names of the letters, not the structure of the alignment.
arXiv ↗21,333
AMD said an Instinct MI455X on Helios delivered 21,333 tokens per second per GPU on DeepSeek-R1 in an MLPerf Inference offline test.
AMD said that is up to 30 percent higher throughput than Nvidia's latest published Vera Rubin offline result. AMD ↗
Agents & coding
AI that acts on its own, and the tools developers build with.
Agents wrote every line of the new AWS Emulator
Amazon said on 11 November that AWS Emulator, a local stand-in for AWS APIs, was written entirely by autonomous agents and that the team has not written or reviewed the code. Matthew Goldberg described three groups: one writes fidelity tests and checks them against real AWS, one replays the tests and files gaps, and one fixes a single issue and must pass a merge gate. He said the suite has tens of thousands of tests, and that hundreds of agents have been fixing thousands of issues a day. The source is on GitHub under CC0.
Why it mattersA large compatibility layer shipped with the tests, not a human review, as the check on every change.
Amazon ↗Muse will call US businesses for paying subscribers
TechCrunch reported on 12 November that Meta's Muse can place calls to US businesses for paid subscribers, after a test in September. Alexandr Wang pointed to calls people avoid, including insurance, a cable company, the DMV and a restaurant without online booking. Plans start at $20 a month for 500 million Muse tokens a week, or $100 a month for 3 billion tokens. TechCrunch said Meta had put Muse past 10 million users. Wang said a Muse email address will soon reach free users. Calls stay on the paid plans.
Why it mattersThe call is now a paid feature, and the person on the other end may still hang up.
TechCrunch ↗Devin Apps hosts a private tool on Devin's machine
Cognition said on 13 November that Devin Apps are private web apps on Devin's cloud computer. A person describes the tool, or starts from a template, and Devin builds it. Everyone who opens an app must sign in to Devin, and apps are never public. Access can be limited to named people, an organization or an enterprise. The machine sleeps when idle, and billing follows how long it is awake.
Why it mattersAn internal dashboard can live where the coding agent already has the repository and the connections.
Cognition ↗ChatGPT Finances arrives on the Mac
9to5Mac reported on 11 November that ChatGPT Finances is in the Mac desktop app, after mobile and the website. The article said that since October it has been on every plan, including free, and only for users in the United States. A ChatGPT post quoted on the page says it is available to Free, Go, Plus and Pro users, and that people can build custom financial tools. The article says the feature supports weekly reports, stock monitoring and voice, and shows up in the sidebar after the update.
Why it mattersThe same spending tools are now in the desktop app, still limited to the United States.
9to5Mac ↗Google Cloud will certify an agentic architect
Google Cloud posted on 11 November that the Professional Agentic Architect certification is part of its Gemini Agent Ready program. The company said nearly one million developers are using that program. The exam is a proctored multiple-choice test plus hands-on labs. Ipsos research cited on the page says 82 percent of learners using Google Skills say naming a Google Cloud certification on a resume draws more recruiter responses, and 90 percent of cloud professionals say such credentials show the know-how for in-demand roles.
Why it mattersA hiring manager can ask for a named exam instead of taking a course list on faith.
Google Cloud ↗Tabletop turns a shared pile of links into a plan
Google Labs said on 11 November that Tabletop, an experiment, takes links, files and a prompt, or a starter template, and builds things such as a trip, a dinner poll, an expense split or a bracket. An invite lets a group use the space at the same time, which the post compares to a Google Doc. It is rolling out in the United States for ages 18 and older, through a waitlist at labs.google/tabletop.
Why it mattersA group can edit one generated plan together instead of passing a chat transcript around.
Google ↗A memory built for the next coding task, not the last one
CoReMem prices memory by what it is worth on the following task. On ChainSWE, the authors say it raises a 27B actor's resolve rate from 21.0 percent to 25.7 percent, with 17 percent fewer steps and 11 percent fewer cache-aware tokens, including the memory calls, against the untrained actor with no memory.
Why it mattersThe memory is being scored by the next job, which is the one the agent has not seen yet.
arXiv ↗Pine Phone lets an agent place the call
The same 13 November release puts Pine Phone on Pine Computer, so an agent can place and finish a real phone call as part of a task. People who already use Muse or Instinct can connect them. Pine said the calls run on Pine Voice 1.0, which it ranks first on τ³-Voice in English at 87.3 percent Pass@1.
Why it mattersThe agent is no longer stopping at the browser. It is dialling a business.
PR Newswire ↗Tool of the day
Something you can try in the next five minutes.
browser-use 1.0.5
Every run can be judged by your own model. One more low-effort request checks the run against the task and fills result.judgement.
pip install 'browser-use==1.0.5'
Video generation
The models and tools that make video.
Google Flow adds five tools for a shot
Google Labs said on 12 November that Flow is getting five tools. The post names Character Poster and Character Arc for keeping a character consistent, Pixel Machine for a retro look, Lucid Glass for light effects, and Pocket Studio, which generates extra camera angles from one phone take while keeping the actor's performance.
Why it mattersA phone clip can be turned into several angles without reshooting the person.
Google ↗Seedance 2.5 on Runway can render a 4K clip
Runway's API changelog for 11 November says Seedance 2.5 supports 4K on text-to-video, image-to-video and video-to-video in reference mode, with ratios such as 3840:2160 or 2160:3840. Output is billed at 110 credits a second, plus 57 credits a second of input and reference video, with an 80-credit minimum. Draft previews stay at 480p and enhanced drafts at 1080p.
Why it mattersThe 4K option has a published credit price, so a longer clip is a known bill.
Runway ↗Runway bills Seedream layers per image, not per job
The same changelog says that on 11 November Seedream 5 Pro Layers moved from 18 credits per decomposition to a charge per returned image: 3 credits at 1K or 1.5K, and 5 credits at 2K or auto, counting the base image and each layer. A job reserves credits for as many as 17 images, then settles to the number returned.
Why it mattersA layer split that comes back with fewer pieces should cost less than the old flat fee.
Runway ↗HyperFlow H3 Flash draws video and stereo sound in a few steps
TPS Report said on 10 November that Video Rebirth released HyperFlow H3 Flash, a 14B-class model that generates video and synchronized stereo audio in as few as 4 sampling steps. Weights are at VideoRebirthAI/hyperflow-h3-flash. The report says clips run at 24 frames a second with 32 kHz stereo, using 4, 6 or 8 steps, and that the weights are not licensed in the EU, the UK, South Korea or the US without prior authorization from MiniMax.
Why it mattersA short clip with sound is being offered as a few-step model, with a license limit written on the card.
TPS Report ↗pi-H3 renders a five-second clip in 3.8 seconds on one B200
Hansheng Chen's post, published 10 November, says pi-H3 generates video in 2 or 4 steps and predicts pixels at the last step, skipping the video VAE decoder. On one B200 it produces a 5-second HD video in 3.8 seconds and a 15-second video in 18.6 seconds, which the post says is more than twice as fast as Sol-H3. A Hugging Face demo is up. The post says weights and code will follow.
Why it mattersThe speed claim is for a demo that is already public, with the weights still listed as coming.
Hansheng Chen ↗Phosphene 5.0.20 uses less memory for a local video
Phosphene 5.0.20, tagged 13 November, says that on a 16 GB Mac an LTX-2.5 video now peaks at about 5.7 GB instead of 9 GB, and a render with LoRAs at about 6.7 GB instead of 19 GB, with the same frames. The Mac app can also keep models on an external drive.
Why it mattersA 16 GB Mac can finish a local clip that previously ran out of memory.
GitHub ↗llama.cpp corrects the shape and rotation of a video frame
llama.cpp build b12381, tagged 14 November, fixes video pixel aspect ratio and video rotation in mtmd. The note is the whole change: the frame was being read with the wrong shape or turned the wrong way.
Why it mattersA local vision model was being shown a stretched or sideways frame.
GitHub ↗A dataset ties floor plans, video and object boxes to one set of rooms
Spatial Foundry is a collection of 5,280 furnished residential 3D scenes with editable assets. The modalities include floor plans, perspective and panoramic images, metric depth, point clouds, meshes, 3D boxes and camera trajectories paired with video. Shared coordinates link views across time, and the scenes can be edited to add camera paths or move objects.
Why it mattersA video model can be trained on a room whose geometry and objects are already named, not only on pixels.
arXiv ↗United is blocking third-party assistants
United said sharing an account with a personal AI assistant can expose information and can lead to an unintended purchase, an account change, or MileagePlus miles being used. It also said check-in needs a person to confirm bags do not contain banned items such as explosives.
Do not give a United login to a third-party agent. Book in United's own app. United Airlines ↗
Creator tools
The apps and platforms people make video with.
Qdrant 1.20.1 speeds cold graph search
Qdrant 1.20.1, tagged 13 November, builds HNSW graphs of large segments with PiPNN and makes multi-condition filters faster on cold on-disk payload indexes. The notes say it halves the number of reads in HNSW search on cold on-disk storage, and quantizes Turbo8 and Turbo16 vectors about twice as fast.
Why it mattersA filtered search that misses the cache does less disk reading than the previous build.
GitHub ↗LanceDB adds exist_ok when creating a table or a secret
LanceDB 0.42.0-beta.13, tagged 14 November, lets Python callers pass exist_ok when creating secrets, views, functions and when cloning tables. The same tag adds negotiated batched blob range reads across the SDKs.
Why it mattersA setup script can create a table that might already be there without failing the second run.
GitHub ↗LiteLLM 1.108.9 publishes the command that checks its image
LiteLLM 1.108.9, tagged 14 November, says its Docker images are signed with cosign, using the key from commit e82a3c6. The note prints a verify command for ghcr.io/berriai/litellm:v1.108.9 that pins that commit's public key.
Why it mattersA deploy can check the image signature against a commit, not only against a tag name.
GitHub ↗LangGraph 1.2.22 streams a child graph on its own thread
LangGraph 1.2.22, tagged 14 November, streams a child graph on its own thread when the caller is already streaming subgraphs. The same tag skips message conversion in add_messages for plain messages, and filters pending writes to the tasks being read.
Why it mattersA nested agent can stream without sharing the parent's thread.
GitHub ↗browser-use 1.0.5 asks your model whether the run worked
browser-use 1.0.5, tagged 14 November, adds one more low-effort request that checks the run against the task and fills result.judgement with success, a failure reason, a description, and flags for a prompt injection or a bot wall. judge_context passes your criteria. judge=False turns the check off. The install line is pip install 'browser-use==1.0.5'.
Why it mattersThe same model that drove the browser can say whether the task actually finished.
GitHub ↗Goose 1.59.1 says the desktop notes are still on the way
Goose 1.59.1, tagged 14 November, says macOS desktop builds and the full release notes are coming soon. The tag is up. The note does not list feature changes yet.
Why it mattersThe version number is published before the desktop build notes are.
GitHub ↗llama.cpp fixes a 32-bit count of trailing zeros
llama.cpp build b12387, tagged 14 November, fixes ctz64 on 32-bit targets. The release note is that one change, pull 32273.
Why it mattersA 32-bit build was using a trailing-zero count that did not match the 64-bit one.
GitHub ↗WellSaid Producer takes a voiceover brief from a chat
WellSaid said on 10 November that Producer works through Claude, Copilot, ChatGPT or any agent harness using its MCP server. The company said it generates audio natively at 48 kHz, and that it has offered 96 kHz output since 2025. Style guides use a pronunciation model trained on more than 500,000 Oxford Languages transcriptions. WellSaid said customers include nearly half of the Fortune 100, and that the platform has produced more than 100 million minutes of voiceover since 2019.
Why it mattersA voice track can be requested from the chat where the script is already being written.
PR Newswire ↗We think a model should learn this from real conversations, the same way language models learn grammar.
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.
Pine open-sources a studio where agents cut a series
Pine AI said on 13 November that Pine Drama Studio, under Apache 2.0, turns a screenplay and one instruction into a story bible, characters, shots and episodes. People approve the designs, storyboards, takes and cuts. An Australian short-drama team moved onto it in October, after a September beta. Pine said each person had been producing 2 to 3 minutes a day and now produces 30 to 50. Generation runs on Pine Computers in the cloud, which is the paid part. Pine Computer drew about 24 million views on its 9 October launch day.
Why it mattersThe unit of work is a series that stays consistent, not a single generated clip.
PR Newswire ↗House House says well under 0.1 percent of Big Walk's code came from a chatbot
Game Informer on 10 November printed House House's answer on Big Walk. The studio said snippets from a chatbot were copied as a starting point and then modified, and that well under 0.1 percent of the code had anything to do with that use. Generative AI was not used in any other part of the game. One example is a Unity shadow-culling bug that made the game run 50 percent slower for about 20 seconds at sunrise and sunset. House House said it had no internal AI policy at the time and called that a failure.
Why it mattersThe disclosure is a fraction of the code, and the studio is treating the missing policy as the mistake.
Game Informer ↗Reddit is testing review pages built from its product threads
Reddit said on 10 November that it is testing Reddit Reviews, which organizes product conversations into a page with a summary, a split of positive, mixed and negative comments, and topic filters. The first set is about 10,000 products in electronics, health and beauty, home and garden, and fashion, rolling out over the coming weeks to US users on iOS, Android and the web. Reddit said half of US shoppers say they come to Reddit to check an AI recommendation. Businesses cannot pay to be featured.
Why it mattersThe page is a summary of existing threads, and a brand cannot buy its way onto it.
Reddit ↗iCloud+ will describe Home camera clips on the largest plans
9to5Mac reported on 11 November that in iOS 27, iCloud+ adds Apple Intelligence for compatible Home cameras on the 2 TB, 6 TB and 12 TB plans. Apple's text, quoted in the article, says the Home app can generate descriptions of recorded clips and search them in natural language. The article lists camera caps of 1, 2 and 5, at $9.99, $29.99 and $59.99 a month. It says the features need Private Cloud Compute, a home hub and an iPhone that supports Apple Intelligence.
Why it mattersThe description feature is tied to the storage plans, not to every iCloud account.
9to5Mac ↗A Nasdaq firm budgets RMB 2 million for AI scripts and films
Ruanyun Edai Technology said on 11 November that Jiangxi Ruanyun and a co-investor each committed RMB 1 million, with no guaranteed return. A 12-month plan aims for about 400 new and adapted scripts and about 30 AI films. The release calls those planning targets, not financial guidance. The budget splits RMB 1.2 million to production, RMB 450,000 to screenplay rights, RMB 200,000 to model and compute costs, and RMB 150,000 to script tools on YeeZo.
Why it mattersThe filing puts a number on scripts and films, and says the number is a plan.
GlobeNewswire ↗Otter adds 27 languages to its transcripts
Otter said on 10 November, at Web Summit Lisbon, that 13 languages are available now and 14 are coming, 27 in all, on top of English, French, Spanish and Japanese. New ones available now include Mandarin, German, Italian, Cantonese, Korean and Portuguese. The company said English tests averaged 96 percent on the Open ASR Leaderboard, and that it does not send audio to third parties. Automatic switching between languages is described as coming soon. The new languages cost nothing extra.
Why it mattersA meeting recorder can take a language it previously left out, without a separate plan.
Otter.ai ↗Gaxos launches ArtGen Muse for cheaper video
Gaxos said on 10 November that ArtGen Muse, a video model on its ArtGen platform, is available. The release says it is meant to produce video at a lower generation cost as the company scales the product. It does not publish a benchmark table in the announcement.
Why it mattersA smaller video shop is putting its own model on the platform it already sells.
GlobeNewswire ↗Beat Breeze lets a store change the music through chat
BMAsia Music launched Beat Breeze in Bangkok on 12 November, an in-store music and media platform. Store teams can control it through ChatGPT or Claude. An AI Music Director reviews changes, skips and volume each night. When enabled, it can adapt music using weather, temperature and store activity. Players are free apps for iPhone, iPad, Android, Windows and a browser.
Why it mattersThe person on the shop floor can ask for a change without opening a separate music console.
MARKETECH APAC ↗Chips & compute
Chips, data centres and the power behind them.
Six Chinese models still carry a Blackwell starting pattern
Will DePue of Unemployment Labs wrote on 12 November that six open models from DeepSeek, Moonshot, MiniMax and Ant Group match Nvidia Blackwell layouts in their starting weights. DeepSeek-V4-Flash and MiniMax-M3 match the 148-processor B200 or B300 pattern. DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ling-3.1-flash match the 152-processor GB200 or GB300 pattern. A B300 reproduced preserved rows of V4-Flash, and a rented GB200 reproduced Kimi K3's. The report says the fingerprint shows the hardware that created the starting weights, not where the chips were, who owned them, or whether an export rule was broken. A revised estimate puts the five runs at about 12,000 Blackwell GPUs, and says that figure rests on assumptions the weights cannot confirm.
Why it mattersThe weights are being read as a record of the chip that initialized them, with the legal conclusion left open.
Unemployment Labs ↗AMD says an MI455X beat Nvidia's published Rubin inference number
AMD said on 12 November that an Instinct MI455X on a Helios rack delivered 21,333 tokens per second per GPU on DeepSeek-R1 in an MLPerf Inference offline configuration, up to 30 percent higher throughput than Nvidia's latest published Vera Rubin offline result. The footnote compares AMD's October lab test with MLPerf Inference v6.1 entry 6.1-0106, retrieved 11 November. AMD also said Helios has up to 20 percent higher HBM bandwidth than Vera Rubin, from its own November calculations. A new MI450X has 288 GB of HBM4 and the same core compute and connectivity as the MI455X.
Why it mattersThe comparison is AMD's lab number against a published MLPerf entry, not a head-to-head run on the same day.
AMD ↗AMD's MI430X keeps full FP64 and 432 GB of HBM4
In the same 12 November post, AMD describes the Instinct MI430X as a part for scientific computing, with native FP64, 432 GB of HBM4 and up to 23.3 TB/s of memory bandwidth. AMD said it looked at 47 workload configurations across 11 applications where emulated FP64 is not enough. An editor's note says the MI430X details were added after the first version of the post.
Why it mattersThe scientific part is being sold on full-precision math and memory, not on the inference tokens of the MI455X.
AMD ↗Coded weights shrink an edge model without changing the GPU kernels
LazyWeights keeps GPTQ-style integer codes but Huffman-codes them and decodes them on demand between DRAM and L2, so the GPU kernels are unchanged. At 1 percent and 3.5 percent perplexity budgets, the paper says coded Phi-3-mini, Mistral-7B and SmolLM-1.7B are 5.7 to 31.4 percent smaller than the best llama.cpp format. A 0.21 mm² decoder unit, at 1 GHz in 7 nm, cuts Orin latency and energy per token by up to 36 percent and 38 percent.
Why it mattersThe saving is in the bytes moved from memory, not in a new matrix kernel.
arXiv ↗Malaysia is checking whether imported AI chips were re-exported
The Straits Times reported on 13 November that Malaysia's investment ministry is asking Customs whether chips brought into the country were later re-exported legally, and that a verification visit does not mean a company has been cleared. Since July 2025, US AI chips leaving Malaysia need permits. The paper said YTL Power has lost an eighth of its market value, RM5.5 billion, since markets opened on 10 November, as investors weighed exposure to Megaspeed.
Why it mattersA customs check is now attached to data-centre projects that were financed on the assumption the chips would stay.
The Straits Times ↗A Bloomberg report says tighter US rules could hit Southeast Asian data centres
The Star, carrying a Bloomberg report on 14 November, said the US Commerce Department is looking at new export rules and licensing for Malaysia, Thailand, Indonesia and Singapore, citing the Wall Street Journal. A Goldman Sachs report cited there forecasts Southeast Asia's data-centre capacity will grow fourfold in five years, with Johor the largest ASEAN hub by 2030. Thai prime minister Anutin Charnvirakul said on Friday that US and Chinese firms had reaffirmed investment plans.
Why it mattersThe next restriction being discussed is not only on China. It is on the countries the chips ship through.
The Star ↗llama.cpp applies the Qwen2-VL feed-forward scales it had been dropping
llama.cpp build b12385, tagged 14 November, says the Qwen2-VL graph passed null for the feed-forward scales, so per-tensor NVFP4 scales loaded by the generic pass were never applied. The model produced garbage, with perplexity around 1e6 on CUDA and NaN on CPU. The fix passes the scales into the feed-forward build, as Qwen2 already did.
Why it mattersQuantized Qwen2-VL was running without the scales that make the numbers meaningful.
GitHub ↗Hexagon flash attention in llama.cpp fixes a no-mask tail
llama.cpp build b12386, tagged 14 November, fixes the Hexagon flash-attention path when a tail of the sequence has no mask. The release note is that one change.
Why it mattersA Qualcomm build was mishandling the end of a sequence that did not use a mask.
GitHub ↗AMD Instinct MI400 Series GPUs bring leadership performance and new capabilities to AI and scientific computing. With advances in high-precision compute, new milestones in AI inference performance and an expanded range of configurations, we’re giving customers more options to power their next generation of AI and HPC workloads.
Policy & safety
Laws, lawsuits and safety incidents.
Nonprofits ask a California court to pause frontier training
Al Jazeera reported on 12 November that a complaint filed Thursday in San Francisco Superior Court, by the Irreplaceable Education Fund, Public Citizen, Heat Initiative and the Alliance for Secure AI, names Anthropic, Google, Meta, OpenAI and SpaceX. The suit asks the companies to pause the technology until safeguards are in place. Phil Aroneanu, founder of the Irreplaceable, said a freeze needs government inspections and the authority to halt prohibited work, not only a ban on public releases.
Why it mattersThe ask is a pause on training, filed in a state court rather than as a bill.
Al Jazeera ↗United tells third-party agents to stay out of its app
United said on 11 November that it is restricting personal third-party AI assistants on its own platforms and through agencies and aggregators, including booking, ticketing and distribution rules. It said sharing an account with such a tool can expose information and can lead to unintended purchases, account changes or use of MileagePlus miles. It gave flight check-in as a case where a person must acknowledge that bags do not contain banned items such as explosives.
Why it mattersAn agent that holds a United login is now something the airline says it will block.
United Airlines ↗A petition asks Nvidia's board to bring Jensen Huang back
TechCrunch reported on 14 November that a petition to reinstate Jensen Huang had more than 115,000 signatures by Saturday afternoon. The petition site listed 5,965 current employees, 1,108 former employees and 63,063 shareholders. Named signers include Elon Musk, Replit chief executive Amjad Masad, Epic Games chief executive Tim Sweeney and Y Combinator chief executive Garry Tan. The petition says Huang was removed without meaningful due process and asks the board to bring him back while an independent investigation establishes the facts. TechCrunch said the stock fell 2 percent on Friday.
Why it mattersThe resignation is already public. This is a demand that the board reverse it.
TechCrunch ↗Asian governments check what Huang's exit does to chip plans
TechCrunch reported on 14 November that South Korea's science ministry said Nvidia had reaffirmed a plan to supply 260,000 advanced GPUs and to continue joint projects, and that a government AI council met on Friday. Taiwan's economy minister Kung Ming-shin said the departure was unlikely to change the growth of the global AI industry, and Taipei said Nvidia's planned headquarters there was still on track. Japan's industry ministry said it would watch Nvidia closely. China's foreign ministry said on Friday it was not aware of reports tying restricted Nvidia chips to a Shanghai data center associated with Alice Huang. Singapore said its investigation into Megaspeed and six other people is still open.
Why it mattersGovernments are asking whether supply promises survive the change of chief executive.
TechCrunch ↗Daybreak agents test a web app and draft the patch
SecurityWeek reported that OpenAI announced Daybreak Red Team and Defense Factory on 12 November at its Cyber at the Frontier event. Red Team uses agents that, in OpenAI's words, use source code and app context to investigate attack paths and confirm vulnerabilities, inside isolated environments with Guardian review. Defense Factory can take findings from Wiz, Snyk and Endor Labs, scan a GitHub repository and prepare a patch. SecurityWeek said engineers review the patch and that it is not deployed automatically. Red Team is early access. Defense Factory is a preview plugin for ChatGPT workspaces that already have Daybreak.
Why it mattersThe finding and the proposed fix stay in review, which is the line between a scan and a change to the repository.
SecurityWeek ↗Tens of thousands of Grok apps opened without a sign-in
Claroc said on 10 November that it tried 43,551 Grok app addresses, found 37,571 still online, and that 35,578 of those, 94.7 percent, opened with no sign-in. Twenty-four of the open apps were built by SpaceXSI and X staff, including a sales territory map and an advertiser board tracking progress toward a $1.6 million sales goal. Claroc said it also saw internal data from 30 other organizations. It said SpaceXSI took down the 24 staff apps after the report.
Why it mattersAn app an agent built on a public address was readable by anyone who had the link.
Claroc ↗A bargaining rule for agents that may be under the wrong privacy law
The paper models one data exchange between two machine-learning agents, with an enforcer, when it is unclear which statute applies. Rules enter as hard constraints. In two of three stress cases where residence and citizenship differ, a citizenship-based rule violates on 21 to 37 percent of shared records, and the authors' legally constrained bargaining does so on under 1 percent. They say that low rate follows partly because, under their default economics, the method shares almost only aggregates.
Why it mattersThe interesting number is not the violation rate alone. It is how little the method is willing to share.
arXiv ↗More than $3 billion in Chinese projects depend on restricted Nvidia chips
TechCrunch said on 14 November that Huang's exit raised uncertainty over more than $3 billion in Chinese AI computing projects that rely on Nvidia's restricted processors. The figure comes from a Reuters review of procurement notices, investment plans and corporate disclosures. Many of those chips are difficult for Chinese buyers to obtain under US export controls.
Why it mattersThe sum is planned projects and contracts, not a new quarter of Nvidia revenue from China.
TechCrunch ↗Robotics
Robots, self-driving and AI in the physical world.
Ring announces an indoor security drone and a patrol robot
TechCrunch reported on 10 November that Ring, owned by Amazon, is launching Ring Robotics. The indoor drone records 2K and docks when idle. The Rover streams continuous 4K and is aimed at businesses. Both use a system Ring calls AI Pilot. The Rover goes on sale on 16 December at $1,000. The drone goes on sale on 27 January at $270.
Why it mattersThe prices and the two sale dates are public, ahead of either device shipping.
TechCrunch ↗DoorDash's Dot starts deliveries in Fremont
DoorDash said on 13 November that Dot is delivering in Fremont, where it is built at Sonic Manufacturing Technologies. Phase 1A runs at most 10 robots, with a local operator alongside. An amended permit for phase 1B allows the number to rise in steps, with the robots driving themselves inside downtown, Centerville and Irvington. Dot is described as one-tenth the size of a car and fully electric.
Why it mattersThe first phase still has a person next to the robot. The permit is what lets that change.
DoorDash ↗DoorDash is teaching Dot to deliver in rain and snow
DoorDash said on 10 November that a Salt Lake City pilot is for deliveries in bad weather. Dot is described as one-tenth the size of a car, fully electric, and manufactured in the United States. Ashu Rege, vice president of autonomy, said the pilot is how the company builds the safety case for more cities and more days of the year.
Why it mattersThe new variable is weather, not another clear-day route.
DoorDash ↗TIER IV posts smaller end-to-end driving models
TIER IV said on 11 November that it is releasing open-source end-to-end driving models. Alpamayo 2 Edge takes camera images and has 2 billion parameters, against 34 billion in Alpamayo 2 Super. AutoE2E's released reactive model has 79.9 million parameters. METEOR has 16 million and can run on Nvidia or Renesas chips. OnePlanner plans about six seconds ahead. LatentDrive plans about four.
Why it mattersThe open weights are the small ones meant to sit in a car, not only the largest checkpoint.
TIER IV ↗XPENG talks up a 2027 driver assist, and a robotaxi, for the UK
Robotics and Automation News reported on 12 November that XPENG is aiming at 2027 for its Next Generation Pilot system, which it says is powered by VLA 2.0, and that it is exploring a robotaxi service and the IRON humanoid in the UK. Brian Gu, vice chairman and president, said the same capabilities behind the cars are what the company is using for robotaxis and humanoid robots. The company also said it wants about 100 retail and service sites in the UK within 2027.
Why it mattersThe UK date is an aim for the driver assist, with the robotaxi and the humanoid still described as prospects.
Robotics and Automation News ↗Tristar raises $12.4 million to inspect a production line
Tristar AI said on 12 November that it raised a $12.4 million Series A led by Isogon Ventures, with Blackhorn Ventures, Ford Street Ventures, ISAI Cap Venture, Capgemini's venture fund managed by ISAI, Las Olas, NextView, Seraph, SNR and TenOneTen. The company said it is on more than 500 production lines in North America and Europe, and that live deployments report 95 to 99 percent defect detection. It said the system typically pays for itself in about four months.
Why it mattersThe round is for cameras on a factory line, with the accuracy figure coming from the company's own deployments.
PR Newswire ↗MuJoCo 4.0.1 fixes a bad Hessian in the Newton solver
MuJoCo 4.0.1, dated 11 November, fixes a sparse Cholesky bug introduced in 4.0.0. The note says rows with different descendants were merged into one supernode, so the Newton solver got a wrong Hessian factor when the constraint Jacobian was sparse, which caused slow convergence, wrong accelerations and divergence. The release also enables the Studio web viewer on Windows, and adds mjs_getInStringVec to read a string vector by index.
Why it mattersA physics step that looked unstable may have been the factorisation, not the robot model.
GitHub ↗A dataset pairs a person's gaze with the robot's own cameras
HEIR records speech, measured gaze and egocentric video together with the robot's observations and teleoperated navigation and manipulation. Sessions keep successive episodes, and a request can take several execution steps. The authors run offline tests of the extra human context and real-robot tests of whether the task is completed. A demo and the dataset are online.
Why it mattersThe robot is being scored on whether it finished the request, not only on whether it predicted the next joint angle.
arXiv ↗A flow model proposes where each LEGO brick has to sit
BrickFlow takes a set of parts and a photo of the finished object and predicts a pose for each part. The authors train a set-conditioned SE(3) flow on generated assemblies, then finetune it on images with 1 percent as much data. An analytic connectivity field pulls the prediction toward an assembly that actually connects. Models and data are released for research.
Why it mattersThe hard part is not the photo. It is that a brick one millimetre off does not click in.
arXiv ↗By piloting Dot in rain and snow in Salt Lake City, we're building the safety case and operational knowledge needed to expand this technology responsibly — not just to more cities, but to more days of the year when people need it most.
Science & health
AI in biology, medicine, climate and maths.
An AI reading prostate MRI matched radiologists, and varied by hospital
Radboudumc said on 12 November that SCARLET-1, published in Radiology, analysed 22,445 prostate MRI exams from 21,252 patients in 46 cities and 22 countries. Overall, the system matched experienced radiologists in clinical and screening settings. Across hospitals in 25 cities it did exceptionally well in 32 percent, moderately in 40 percent, and less accurately in 28 percent. The centre tied that spread to image quality and to who the patients were.
Why it mattersA single accuracy number hides the hospitals where the same system falls behind.
Radboudumc ↗A surgical blood-pressure target is chosen so the low readings stay rare
The paper fits a small drug model from a short record of mean arterial pressure and norepinephrine, then searches for the lowest target at which a bolus and an infusion keep the predicted chance of a reading below 65 mmHg inside a set limit. If no such target exists, the decision goes back to the clinician. The authors say they check calibration of the predicted minimum pressure on retrospective surgical data.
Why it mattersThe output is a target and a dose, and an explicit hand-back when the bound cannot be met.
arXiv ↗Silscore predicts siRNA potency with 26 fitted parameters
Silscore combines a preference for overall duplex stability, a position-by-position stacking profile, and a few nucleotide-identity terms chosen from what the thermodynamics leave unexplained. The main fit uses 26 parameters, which the authors say is fewer than half as many as any trainable comparator they evaluated. It is a guide-only model: it scores the siRNA sequence, not a large learned network.
Why it mattersThe claim is that a short thermodynamic model still competes with heavier ones on potency.
arXiv ↗Four organoid screens are released for drug-response models
The authors say most drug-response models are trained on immortalized cell lines, which can differ from a patient's tumour. They ran four screening studies on patient-derived organoids and are releasing the data so those models can be trained and tested closer to the clinical target. The page presents the dataset as the contribution.
Why it mattersThe gap being filled is the training set, not another architecture on the same cell lines.
arXiv ↗MutAtlas compares mutation scores without experimental labels
MutAtlas enumerates single-site substitutions on experimental protein structures and lines up scores from energy models, protein language models and inverse-folding models. The authors say the sources disagree in ways that look like different biases, not noise. A distillation method that never sees experimental mutation labels is, they report, the best of the zero-shot baselines and of naive fusions they tried on ProteinGym. The dataset and the evaluation pipeline are released.
Why it mattersThe useful output is where the models disagree, because that is the part an experiment should check.
arXiv ↗A texture descriptor learns its radius and still outputs hard binary codes
FLEX-LBP learns the radius, sampling angles and thresholds of a local binary pattern by backpropagation, while every forward pass still computes exact binary codes. The deployed descriptor has 90 learnable parameters. One configuration, used unchanged on six colour-texture benchmarks, improves nearest-neighbour accuracy of the fixed descriptor by 3.55 points on average over five seeds. On the Kather colorectal histology set, the fully learned descriptor reaches 93.8 percent, against 92.8 percent for the best classical descriptor they cite. Code is released as a PyTorch layer.
Why it mattersThe histology gain is small, and it comes from a descriptor a person can still inspect.
arXiv ↗Ground-reaction estimates get worse when the same person is kept out of training
The paper trains three sequence models to estimate ground-reaction forces from motion-capture markers, on a locomotion set of 6 subjects and a patient-handling set of 10, in two laboratories. It compares a trial-level split, which can put the same person in train and test, with leave-one-subject-out and with transfer across labs. The point of the study is how much those choices change the reported error.
Why it mattersA force estimate that was validated on another trial from the same person is not the same test as a new person.
arXiv ↗A robot's own mechanics can bias a smoothness score used in rehab
The paper says robot hardware can bias kinematic measures of recovery and make devices hard to compare. The authors built a closed-loop human-robot simulation calibrated to published healthy and stroke kinematics, held the motor plan fixed, and measured bias in seven metrics against a low-impedance reference. The result is a correction for that bias, estimated in simulation.
Why it mattersA patient can look smoother because the robot is smooth, not because the movement recovered.
arXiv ↗Hist2Prot reads a routine slide as a spatial protein map
Nature Cancer published Hist2Prot on 13 November, a model that predicts single-cell protein maps from an ordinary hematoxylin and eosin slide. It was trained on about 51 million paired cells and tested on further sets of about 72 million cells. Across seven cohorts and 1,774 patients, the virtual proteins improved biomarker, prognosis and immunotherapy-response predictions over the stain alone.
Why it mattersThe scan a hospital already makes is being asked to stand in for a protein assay it did not run.
Nature Cancer ↗The standard of care in prostate cancer management has continuously evolved alongside new guidelines, imaging technologies and workforce changes. Prostate AI systems will likely require the same approach, with structured evaluation, monitoring and periodic refinement to ensure they remain aligned with contemporary clinical practice.
Open source
Open-weight models and community releases.
TiHon-TTS speaks Vietnamese from 100,118 parameters
A Hugging Face post on 13 November says TiHon-TTS has 100,118 parameters, that the network is 402 KB in float32, and that it runs 40 times faster than real time on one x86 core and 83 times on one Apple M4 core. It was trained on text, not on recordings: 80,000 steps across 78,700 Vietnamese sentence chunks, with four voices. The license is CC BY-NC 4.0.
Why it mattersA speech model this small can sit on a microcontroller-class budget, and it was never played a waveform in training.
Hugging Face ↗uv 0.13.8 installs CPython 3.16.0a2
uv 0.13.8, released 12 November, adds CPython 3.16.0a2 and updates the managed 3.11 through 3.15 builds to pip 26.3, zlib 1.4 and Expat 2.10.0. A preview path can install wheel variants from a local path, a direct URL or a Git repository, and can pick them using the machine's x86-64 level or AArch64 version.
Why it mattersA preview Python and a CPU-specific wheel can be installed with the same tool that already manages the stable versions.
GitHub ↗Weaviate 1.40.5 checks every verb before it grants a role
Weaviate 1.40.5, tagged 10 November, requires every verb of a manage permission before that role can be assigned, and before it can be seen or edited. The notes list no new features. The other changes are fixes, including dropping a nil error from a read-repair message.
Why it mattersA partial permission no longer counts as the right to administer a role.
GitHub ↗Ruff 0.17.5 redirects rules that were renamed
Ruff 0.17.5, released 12 November, redirects renamed rules in its preview features, so a config that still names the old rule is sent to the new one. The release is a preview-rule update, not a new linter.
Why it mattersA rename in the rule list does not have to be a silent miss in an existing config.
GitHub ↗Pixi 0.84.2 fixes a memory climb and builds for AMD GPUs
Pixi 0.84.2, released 10 November, fixes a memory-use regression introduced in 0.82.1 and adds support for building AMD GPU packages. It is a patch on the package manager, not a new solver.
Why it mattersProjects that started using much more memory after 0.82.1 have a patch, and an AMD GPU build is now part of the same release.
GitHub ↗Claude's agent SDK 0.3.322 matches Claude Code 2.1.322
Claude Agent SDK 0.3.322, tagged 10 November, says it was updated to parity with Claude Code 2.1.322. The note does not list a separate feature beyond that match. Install is npm install @anthropic-ai/[email protected].
Why it mattersA host pinned to the SDK can sit on the same build as that Claude Code release.
GitHub ↗llama.cpp starts a new compute split at a weight nothing reads
llama.cpp build b12383, tagged 14 November, starts a new split when it hits an incompatible weight inside a split that has no inputs. The change is in the backend splitter.
Why it mattersA graph that mixed incompatible weights was being kept in one split.
GitHub ↗LiteLLM 1.106.10 stops a Google MCP login loop
LiteLLM 1.106.10, tagged 14 November, says a fix stops Google MCP servers from asking a user to sign in again when the access token expires. The note says the patch is two pull requests, one of them that fix, and that no action is required to upgrade. The image is signed with the same cosign key as later tags.
Why it mattersA connector that expired its token was sending people back through login on every call.
GitHub ↗Cost saving tip
One change that makes the same work cheaper, from today's news.
Half off 4K images during launch week
MiniMax said that for launch week, image requests at 4K and above are 50 percent off the standard price. A 4096 by 4096 image is the size the post calls 4K.
MiniMax ↗If you were going to try H3-Image at 4K, the launch-week price is the one to use.
Quick hits
Smaller stories worth a glance.
- A custom video decoder can plug into llama.cppBuild b12380 lets mtmd use a custom video decoder, and adapts the path for wllama.GitHub ↗
- Jinja templates in llama.cpp allocate less while they renderBuild b12379 cuts value allocation and lookup overhead when a chat template is rendered.GitHub ↗
- Whisper input skips chunks that are only paddingBuild b12378 stops mtmd from feeding Whisper audio chunks that contain nothing but padding.GitHub ↗
- DFlash speculative decoding moves into its own fileBuild b12377 moves the DSpark Markov head and backbone graphs into dspark.cpp. The note calls the change a move.GitHub ↗
- Hexagon in llama.cpp gains IQ2_XXS and NVFP4Build b12376 adds IQ2_XXS and NVFP4 layouts and HVX matmul kernels on Hexagon.GitHub ↗
- The chat template drops a Mistral workaroundBuild b12375 removes a Mistral-specific workaround from llama.cpp's chat handling.GitHub ↗
- Metal matrix multiplies stop overflowing a destination offsetBuild b12374 fixes a destination-offset overflow in Metal's mul_mv kernel.GitHub ↗
- Prefect's dev build fails a benchmark that only hungPrefect 3.9.4.dev1 fails the CodSpeed job when the benchmark fails, instead of timing out during the import.GitHub ↗
Deals & funding
Who raised money, and who bought whom.
Funding rounds 2
Investors put money into a company in exchange for a share of it.
Standard Capital backs Luminal
- Raised
- $25M
Luminal valuation not disclosed
Standard Capital said it led a $25 million Series A in Luminal, which builds a compiler that searches for an efficient way to run a model on existing chips. Founders named are Joe Fioti, Matthew Gunton and Jake Stevens.
Isogon Ventures backs Tristar AI
- Raised
- $12.4M
Tristar AI valuation not disclosed
Tristar AI said it raised a $12.4 million Series A led by Isogon Ventures for visual inspection on production lines. The company said live deployments report 95 to 99 percent defect detection.
BVBlackhorn VenturesFSFord Street VenturesICISAI Cap VentureCVCapgemini VC FundLOLas Olas VCNVNextView VenturesSPSeraph PartnersSVSNR VenturesTTenOneTen+0 more
On the calendar
What's coming up, in date order, and how sure we are about each.
- 9–20NovCOP31 climate conference in AntalyaToday
- 15NovComments close on China's draft tech-export listToday
- 15NovWorkBuddy and CodeBuddy free period endsToday
- 17NovThird-quarter results and earnings call2days away
- 18–19NovTrump and Xi meet in Shenzhen3days away
- 18NovWebinar on running a dock with Stretch3days away
- 18NovNvidia invited to brief Congress3days away
- 18Nov‘Dot’ agents become the default in workspaces3days away
- 19NovKinference demo livestream4days away
- 19NovClaude Data turns on for Enterprise4days away
- 20NovAI Film Award submissions close5days away
- 20NovConductor cloud service shuts down5days away
- 20NovUpdated OpenAI usage policies take effect5days away
- 21NovInaugural edition in Hollywood6days away
- 22Nov'The Challenge Sverige' originally scheduled (now delayed to 2027)7days away
- 23NovIn-person registration closes for the FDA robot workshop8days away
- 25NovQwen4-Max preview ends10days away
- 26NovHuang's deadline to testify11days away
- 29NovDaybreak Blue sessions can be used for training2weeks 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
- 4DecAI-made film opens in limited release3weeks away
- 4DecGods Don’t Give Gifts premiere3weeks away
- 4DecSingapore AI Safety Fellowship ends3weeks away
- 10DecIntelligent eyewear goes on sale4weeks away
- 16DecKorea AI Advertising Awards winners4weeks away
- 16DecRing Rover goes on sale4weeks away
- 22DecKorea AI Advertising Awards ceremony5weeks away
- 31DecHy4 half-price offer ends7weeks away
- 31DecJules shuts down7weeks 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.






