Thursday, 19 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.
- the IPO price, above a $70 to $78 range
- $83
- the raise if all 825 million shares sell
- $68.5B
- revenue in the nine months to 30 September
- $33.24B
- run-rate revenue at the end of October
- $81B
Anthropic prices its IPO at $83 a share
Anthropic priced its IPO at $83 a share on Wednesday, above the $70 to $78 range in last week's filing, a person familiar told CNBC. If all 825 million planned shares sell at that price, the offering raises about $68.5 billion. Trading on Nasdaq as ANT starts Thursday.
Why it's the story of the dayThe price is set, and the Claude maker is due to start trading on Thursday.
Startup of the day
A young company people are talking about today.
Humyn Labs raises $47 million to film the work robots still cannot do
Humyn Labs, a Bengaluru startup, raised $47 million in Series A funding, TechCrunch reported on 18 November. Accel led $26 million of equity, with Prosus Ventures, BeeGlobal, Bajaj Finance's Finserv Intelligence, Bajaj Alts, Dallas Venture Capital, Lighthouse Canton and Jump Trading. A further $21 million is venture debt led by Lighthouse Canton, with Trifecta Capital. The company records everyday tasks on its own two-camera rig across more than 20 countries, then labels the footage for robotics and world-model labs. Co-founder Manish Agarwal said revenue from April to October was $10 million, with another $20 million of contracts signed. Co-founder Ishank Gupta said about 90 percent of revenue is from the US. The company says it has 12 paying customers.
Why it's worth a lookRobot labs are buying footage of real work, and a specialist recorder just raised to shoot more of it.
TechCrunch ↗Models & labs
New models, the labs behind them, and the rules they answer to.
Google launches Gemini 4 Neon at $3 and $15
Google on 18 November introduced Gemini 4 Neon for coding, knowledge work and science, priced at $3 per million input tokens, $0.15 cached and $15 per million output tokens, with a million-token window. At its extra-high setting Google says Neon scores 61 on the Artificial Analysis Intelligence Index, third overall. At the default high setting it says cost per task is 32 percent lower than Carbon. Neon is in the Gemini API, the Gemini app for Pro and Ultra, and Google Antigravity.
Why it mattersA mid-tier Gemini 4 model is priced under Carbon and is already in the API and the app.
Google ↗DaisyRose-27B lifts a coding score on 1,200 tasks
RadixArk on 18 November released DaisyRose-27B under Apache 2.0, built on Qwen3.8-27B and trained with Miles on 1,200 curated tasks over 18 days on 144 H200 GPUs. The lab says Terminal-Bench 4.1 pass@1 rose from 5.6 percent to 12.0 percent. The model keeps the base model's 256,000-token window, and checkpoints are posted in BF16, FP8 and NVFP4.
Why it mattersA small post-training run on a 27 billion parameter model moved a hard coding score.
Hugging Face ↗Cloudflare opens weights for Clairic and a 9B flash
Cloudflare on 18 November posted open weights for an updated Clairic and for Clairic-flash, a 9 billion parameter version. The 27 billion parameter model now sits at @cf/cloudflare/clairic on Workers AI, and the 9 billion model moves to @cf/cloudflare/clairic-flash. Cloudflare says extraction from images and PDFs is improved, including tables, lists and dates.
Why it mattersA document model that was an API preview can now be downloaded and run elsewhere.
Cloudflare ↗LangChain's OpenAI package adds GPT-6.2 Astra and Sol
langchain-openai 1.8.2, released 18 November, adds support for GPT-6.2 Astra and GPT-6.2 Sol. The same release stops sharing one async HTTP client between models. The notes also say the client can send LangSmith gateway headers.
Why it mattersApps on LangChain can call the models OpenAI shipped this week without a private patch.
GitHub ↗llama.cpp's OpenVINO path adds models and cuts memory use
llama.cpp b12503, tagged 18 November, expands the OpenVINO backend and says it optimises use of the NPU, the GPU and memory. The notes also fix detection of a sliding-window mask when the cache size matches the full context.
Why it mattersA local build can send more models to an Intel NPU without holding the whole cache.
GitHub ↗llama.cpp rejects a Limite config that points past its taps
llama.cpp b12527, tagged 18 November, rejects Limite configs whose tap index runs past mudd_tap_idx at load. The note says n_mudd_taps comes from limite.mudd.tap_count, a scalar read from the file.
Why it mattersA bad checkpoint fails at load instead of reading off the end of the tap table.
GitHub ↗llama.cpp applies a Qwen4 n-gram scale inside the graph
llama.cpp b12490, tagged 18 November, fixes graph reallocation for the Qwen4 experimental n-gram table scale. The scale is applied with a ggml operation rather than a side path that forced the graph to be rebuilt.
Why it mattersA Qwen4 checkpoint with an n-gram table can run without reallocating the graph mid-way.
GitHub ↗llama.cpp's chat parser reads Inkling
llama.cpp b12455, tagged 17 November, adds a chat parser for Inkling. The build is the local runtime, separate from the Transformers loader that gained Inkling weights last week.
Why it mattersA model whose weights already load in Transformers can now be prompted through llama.cpp's chat template.
GitHub ↗llama.cpp decodes an FP8 type without a libm call
llama.cpp b12484, tagged 17 November, decodes F8_E4M3 values to 32-bit float without calling ldexpf. The change is in the shared tensor code, so every backend that reads that type picks it up.
Why it mattersA small numeric path drops a library call on every FP8 value the runtime unpacks.
GitHub ↗$83
CNBC said Anthropic priced its IPO at $83 a share, above a range of $70 to $78.
If all 825 million planned shares sell, the offering raises about $68.5 billion. Trading as ANT is set for Thursday. CNBC ↗
Agents & coding
AI that acts on its own, and the tools developers build with.
Citi says an AI agent on your account is your problem
Citibank changed its online user agreement so customers who give a third-party AI agent access to an account are responsible for what that agent does, TechCrunch reported on 18 November. The update also says the customer is responsible if the tool leaks credentials or is hacked. PNC changed its terms the same day. JPMorgan and Bank of America had announced similar rules last month. Citi's own agents are excluded from the extra liability.
Why it mattersA bank will not reverse a payment just because an agent, or a thief using the agent's keys, sent it.
TechCrunch ↗Unsiloed raises $7.5 million to structure documents for agents
Unsiloed AI raised a $7.5 million seed round led by Felicis, TechCrunch reported on 18 November, with Y Combinator, Orange Collective and Uncorrelated Ventures. The startup turns tables and charts in documents into structured data and points back to the source page. Mishra told TechCrunch of a customer test on 542 financial documents that scored 99.78 percent across about 84,000 fields. TechCrunch said that figure had not been checked independently. The company has seven people.
Why it mattersAgents still fail on the documents companies actually keep, and a seed round is betting on the parser.
TechCrunch ↗Linq raises $90 million so agents can text and call
Linq raised $90 million in a Series B led by Andreessen Horowitz, a report on 18 November said, taking total funding past $120 million. Ben Horowitz is joining the board. The company says it has processed more than a billion messages, and a new HD Calling feature lets an agent place a phone call. TQ Ventures, Mucker Capital, Abstract, Anthos and Left Lane Capital joined the round.
Why it mattersAn agent that can text and call needs a phone network, and that network just raised a large round.
Tech Startups ↗OpenHands 1.42 can pick which git host a repo uses
OpenHands 1.42.0, released 18 November, lets a cloud run choose the exact connected git provider for a repository. The same release restores validation of a native webhook source and gives every verify run a non-loopback plain-HTTP origin.
Why it mattersA coding agent that talks to more than one git host can be told which connection to use.
GitHub ↗n8n keeps each item's tool calls in its own history
n8n 2.48.5, released 18 November, stops the AI Agent node from mixing tool-call history across input items. A batch that runs the same agent on several rows no longer lets one row's tool results leak into the next.
Why it mattersAn automation that loops an agent over a list was able to carry the previous row's tools forward.
GitHub ↗Mastra 1.84 stores a long agent run as a header plus steps
Mastra core 1.84.0, released 18 November, lets durable and evented agents opt into storage version 2, which keeps a run as a header and per-step rows. The notes give an example of a published payload falling from 98 MB to 11 MB and bytes written from 75 MB to 3.1 MB. A recovery loop started by startWorkers() restarts crashed runs without waiting for another boot.
Why it mattersA long agent trace was being rewritten in full on every step, and the new store writes the delta.
GitHub ↗CrewAI 1.15 makes Chroma optional and defaults knowledge to Qdrant
CrewAI 1.15.42, released 18 November, makes ChromaDB an optional extra and points the default Knowledge store at a local Qdrant. Standalone agents can have tools picked now and more added later. Each flow also gets its own id from the initial state.
Why it mattersInstalling the agent framework no longer has to pull a vector database you are not using.
GitHub ↗browser-use 1.0.6 makes BU 4.0 the default model
browser-use 1.0.6, released 16 November, runs bu-4.0 on the Browser Use gateway by default. Agent(task) and ChatBrowserUse() no longer fall back to bu-fast unless you pass another model.
Why it mattersA library upgrade changes which model drives the browser, even if your code does not name one.
GitHub ↗Goose 1.60 adds a skill operation and drops the old extension
Goose 1.60.0, released 17 November, adds a SkillOperation and removes the Skills extension. The same release adds executable commands and a SlashCommandOperation in oi-kit.
Why it mattersA desktop agent is changing how skills are installed, so an old extension may no longer be the path.
GitHub ↗Tool of the day
Something you can try in the next five minutes.
Rumble-2-Flash
Reson8's 0.5 billion parameter speech model transcribes 40 languages and runs in Transformers or vLLM. On a 302-hour call-centre set it lands next to a 30 billion parameter model.
uv pip install torch transformers accelerate
Video generation
The models and tools that make video.
Genie 4 opens in Project Genie for Ultra subscribers
Google on 18 November opened Genie 4 to eligible Google AI Ultra subscribers in Project Genie. A prompt typed mid-session can change the weather, add a character or redirect the action, and the model adds native audio, from ambience to dialogue. Genie 3 stays in the picker for longer runs, now up to three minutes a generation, up from 60 seconds. Pro subscribers can join a waitlist.
Why it mattersA world model that was limited to testers is now something a paying subscriber can walk around in.
Google ↗Horizon's LoFT rebuilds a scene from an ordinary video
Horizon Robotics' LoFT, a 514.5 million parameter transformer, reconstructs depth and camera pose from RGB video. One loop gathers features over windows of 3, 5, 9 and 17 frames, alternating with attention over the whole clip. A second loop predicts a camera pose and feeds it back. Horizon says it trained on 80 public datasets, 2.97 million sequences and 961.8 million images, and that the model is competitive on SpatialBench, ClassicBench and VRBench.
Why it mattersA robot's camera is a video, and this model is built to turn that video into geometry.
Horizon Robotics ↗A 1.3B video model answers in 36 milliseconds a frame
FastFrames, posted on arXiv on 17 November, is a family of autoregressive video diffusion models aimed at interactive latency. The authors say VAE decoding and KV-cache updates accounted for more than half of end-to-end latency, and they replaced those pieces. A 1.3 billion parameter variant runs at 36 milliseconds a frame, 28 responses a second, on one B200, which they say does not give up generation quality.
Why it mattersInteractive video has been too slow to answer a new frame while a person is still moving.
arXiv ↗Spool keeps a sliding video window without a full rerun
Spool, posted on arXiv on 17 November, is a training-free way to run hybrid vision-language models on a recent slice of video, such as the last two minutes. Recurrent layers fold every past frame into one state, so dropping an old frame normally means replaying the window. The authors cache a transformation per frame and compose those transforms as the window moves.
Why it mattersA model that watches a live camera can forget the frames that just fell out of the window.
arXiv ↗VIVID keeps the motion between a video's keyframes
VIVID, posted on arXiv on 17 November, argues that video question-answering models throw away the frames they do not pick and then encode the rest as stills. Whole-frame pooling hides a small moving object, and dropped frames erase direction and order. The method keeps keyframes for the main states and adds patch differences between frames.
Why it mattersA video model that only sees the frames it kept cannot tell which way something moved.
arXiv ↗WebGPU in llama.cpp grows a timestep embedding
llama.cpp b12468, tagged 17 November, adds a timestep embedding on the WebGPU backend. Diffusion-style models use that embedding to know which denoising step they are on. The note is the WebGPU implementation of the operation.
Why it mattersA browser build can take a step index for a diffusion model instead of failing the graph.
GitHub ↗llama.cpp runs float conv2d on the fast CUDA path
llama.cpp b12493, tagged 18 November, sends 32-bit conv2d kernels down the implicit GEMM path, using MMA or cuBLAS. The fast path had been limited to other types. Vision and video models use conv2d to turn a frame into patches.
Why it mattersA float vision tower on CUDA no longer has to take the slow convolution.
GitHub ↗llama.cpp fixes the order of Muse Glimmer's vision tokens
llama.cpp b12463, tagged 17 November, fixes the order of vision embeddings for Muse Glimmer in the multimodal loader. A wrong order feeds the language model patches that do not sit where the picture says they do.
Why it mattersA vision model that loads locally was handing the language model its patches out of order.
GitHub ↗Metal in llama.cpp grows an implicit GEMM for conv2d
llama.cpp b12469, tagged 17 November, adds an implicit GEMM kernel for conv2d on Metal. The operation is how a vision model turns an image into a grid of patches on an Apple GPU.
Why it mattersA Mac can fold a frame into patches on the GPU instead of a fallback path.
GitHub ↗Two agent-framework builds on PyPI stole keys
Wiz says agent-framework 1.26.1 and 1.26.2, published for under an hour on 17 November, run a credential stealer whenever Python starts. The last clean release is 1.26.0.
Search lockfiles for those two versions and site-packages for the extra .pth files, then rotate any keys that machine could see. Wiz ↗
Creator tools
The apps and platforms people make video with.
Snapchat lets a coding agent build an AR Lens
Snap on 18 November released Easy Lens CLI, so a Lens can be built from Claude Code, Cursor or Codex, TechCrunch reported. The assistant uses more than 100 AR building blocks and 50 asset generators, makes a preview, and can publish to Snapchat. Snap says it runs on Mac, Windows and Linux, and that a game can go from a description to a published Lens in a day.
Why it mattersAn effect that used to mean a specialist tool can now be described in the editor a developer already has open.
TechCrunch ↗The AI SDK can cancel a model call that has not answered yet
AI SDK 7.0.177, released 18 November, adds an interrupt option on steering so a steer can cancel a model call that has not produced output. The same release ships an AI SDK skill in the package at skills/ai-sdk for pnpm agent skills, and bumps the bundled gateway.
Why it mattersA user who changes their mind mid-generation no longer has to wait out a call that has not started speaking.
GitHub ↗LanceDB beta.20 adds typed handles for compaction jobs
LanceDB 0.42.0-beta.20, released 18 November, adds typed SQL job handles for compaction and data prewarm. The notes mark the change as breaking for the SDK. The release is the next beta after the row-recovery APIs shipped the day before.
Why it mattersA job that compacts a table can be tracked as a typed handle instead of a bare query.
GitHub ↗Qdrant 1.20.2 speeds a batch that shares one filter
Qdrant 1.20.2, released 16 November, makes batch searches that use the same filter on the same vector significantly faster. The change is in the search path, so a client that already batches queries picks it up on upgrade.
Why it mattersA retrieval service that repeats one filter across a batch was redoing work it could share.
GitHub ↗llama.cpp returns 400 when a grammar fights ignore_eos
llama.cpp b12489, tagged 18 November, makes the server return HTTP 400 when a request sets both a grammar and ignore_eos. The note says the combination crashed the sampler. The client now gets an error instead of a dead server.
Why it mattersTwo sampling flags that cannot be true together were taking down the process.
GitHub ↗LangChain tags a synthetic approval message with its source
LangChain 1.5.2, released 18 November, tags synthetic human-in-the-loop tool messages with source metadata. A message the library inserts for an approval can be told apart from a message the model or the user wrote.
Why it mattersA log of an agent run can show which tool messages were invented for the approval step.
GitHub ↗LangChain's xAI client sends LangSmith gateway headers
langchain-xai 1.4.1, released 18 November, sends langsmith_gateway headers from the client. The change is in the xAI integration, so a call that already goes through that package starts attaching the headers on upgrade.
Why it mattersTraces of an xAI call can be routed through LangSmith without a separate proxy config.
GitHub ↗llama.cpp's Jinja loader walks an outer scope for a key
llama.cpp b12448, tagged 17 November, makes Jinja template lookup walk the scope chain for a key. A variable set outside the current block can be read from inside it. The note adds a test for a lambda in a call statement.
Why it mattersA chat template that closed over an outer variable was failing to see it.
GitHub ↗llama.cpp fixes a LoRA schedule on a recurrent slot
llama.cpp b12498, tagged 18 November, fixes the server slot scheduler for LoRA and recurrent models. The hybrid sliding-window test in the same change runs with one token per ubatch.
Why it mattersA server that swaps adapters on a recurrent model could schedule the wrong slot.
GitHub ↗I think if we do this right, we can create an age of miracles where humanity is able to solve problems that have eluded us for millennia.
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.
Le Figaro's reporting will show up inside ChatGPT
OpenAI and Le Groupe Figaro on 18 November announced a partnership that puts Le Figaro's current reporting and archive into ChatGPT as summaries and excerpts, with links to the stories. The deal covers news, business, culture and lifestyle from Le Figaro and associated titles. Le Figaro keeps editorial control. It also plans to roll out ChatGPT Enterprise to teams, under its own AI charter.
Why it mattersA national newspaper is licensing discovery inside the chatbot, not only a training opt-out.
OpenAI ↗Loblaw is the first Canadian retailer on Meta's Muse
Meta on 18 November said Loblaw is bringing its catalogue to Muse as a partner connector, the first Canadian retailer to do so. Muse is Meta's shopping agent. The connector is how the agent's answers reach that catalogue.
Why it mattersA shopping agent is only as useful as the stores that actually plug a catalogue into it.
Meta ↗Google Search AI Mode can forget the thread
Google is adding temporary threads to AI Mode in Search, TechCrunch reported on 18 November. A button in the upper right starts a thread that is detached from the account: it is not resurfaced in search history or recommendations, and it does not change future personalisation. Closing it returns to ordinary search. The option is rolling out globally on desktop and mobile.
Why it mattersA search that turns into a conversation no longer has to become part of the profile.
TechCrunch ↗llama.cpp can run the YuE2 music model
llama.cpp b12458, tagged 17 November, adds support for YuE2-3B in the multimodal loader. The notes say the change improves speed, adds classifier-free guidance, fixes streaming and exposes a CFG control.
Why it mattersA local build can generate music from a checkpoint that the runtime did not previously load.
GitHub ↗llama.cpp speaks with a Qwen3 voice reference
llama.cpp b12491, tagged 18 November, adds Qwen3-TTS in-context learning, so a reference clip can steer the voice. The multimodal loader is what plays the audio tokens back.
Why it mattersA local text-to-speech model can follow an example voice instead of only a preset.
GitHub ↗llama.cpp adds a tiny pronunciation dictionary
llama.cpp b12497, tagged 18 November, adds tiny-voca to the multimodal tools and improves the grapheme-to-phoneme step. The dictionary is how a speech model decides what a word should sound like.
Why it mattersPronunciation for a local voice model got a smaller dictionary and a better letter-to-sound step.
GitHub ↗llama.cpp loads Fish Audio S2, including a cloned voice
llama.cpp b12452, tagged 17 November, adds Fish Audio S2 to the multimodal loader, including voice cloning. This is the local checkpoint path, not the hosted Live API.
Why it mattersA Fish speech checkpoint can run in the same local server as the text models.
GitHub ↗Rumble-2-Flash transcribes 40 languages at half a billion parameters
Reson8 on 18 November released Rumble-2-Flash, a 0.5 billion parameter speech model under Apache 2.0, with weights at Reson8/Rumble-2-Flash. It reports a 6.96 percent error rate across nine languages on the Open ASR Leaderboard, ahead of Whisper large-v3, and 6.6 percent on English. On a 302-hour call-centre set it scores 7.86 percent, next to NVIDIA's 30 billion parameter Nemotron 3.5 Transcribe at 7.78 percent. Full batches pass 7,000 times real time.
Why it mattersA speech model small enough to sit next to a product can match a model sixty times its size on phone audio.
Hugging Face ↗llama.cpp adds Kokoro as a text-to-speech model
llama.cpp b12494, tagged 18 November, adds Kokoro text-to-speech support in the multimodal loader. The same change implements scope-chain key lookup in Jinja and a test for a lambda in a call statement. The speech path is separate from the phoneme dictionary added for Kokoro a day earlier.
Why it mattersThe local runtime can now synthesise speech with Kokoro, not only look up how a word is pronounced.
GitHub ↗Chips & compute
Chips, data centres and the power behind them.
Nvidia stops shipping licensed Hopper chips to China
Nvidia has stopped sending licensed Hopper H200 chips to China, interim chief executive Colette Kress said on the 17 November earnings call, The Verge reported. She said the company's guidance includes no licensed Hopper data-centre products and that Nvidia is no longer shipping them to China. She did not say whether other China business would continue. The Verge noted a Reuters report that about 600,000 H200 chips were stuck in Hong Kong after Beijing withheld import approval.
Why it mattersThe export licence the company fought for is no longer a product it is willing to ship.
The Verge ↗RTX Spark desktops go on sale from Asus, Dell and MSI
The first RTX Spark mini PCs are on sale from Asus, Dell and MSI, The Verge reported on 18 November. They pair a 6,144-core Blackwell GPU and a 20-core Grace CPU with up to 128 GB of memory. Asus's ProArt GR1X starts at $4,499 with 64 GB and 1 TB. Nvidia says they deliver up to 1 petaflop of AI performance, with native CUDA and Linux tools through Windows Subsystem for Linux. Acer, Gigabyte and HP come later, and Microsoft's Surface RTX Spark Dev Box is due to ship this month.
Why it mattersA desktop for local models is now a product you can order, at a workstation price.
The Verge ↗OpenAI says its models cut AMD GPU area and delay
OpenAI on 18 November said work with AMD found up to a 14 percent power, performance and area gain against heavily optimised baselines on flagship GPU IP, by rewriting SystemVerilog. AMD has put some of the changes into a design already sent for manufacturing. On its own Jalapeño chip, OpenAI says model-proposed edits cut standard-cell area about 13 percent in the vector unit and 10 percent in the matrix unit, and that those cuts were proven in silicon. The company is also training a model, GPT-Synopsys, to drive Synopsys design tools.
Why it mattersThe lab is publishing measured changes to a chip that is already on the way to the factory.
OpenAI ↗Nvidia says a neocloud guarantee starts only after the site is real
Nvidia's interim chief told analysts the company's loan guarantee for a data centre does not exist until the customer has raised the money, built the shell and installed Nvidia's systems, TechCrunch reported on 18 November. If that work slips, "our guarantee doesn't exist anymore," Colette Kress said. She said Nvidia would be more selective about who it finances, with land, power and shells first. The largest such backstop TechCrunch names is the $105 billion guarantee for OpenAI's PORTS-Pike campus in Ohio. Net income in the quarter was $64.9 billion, up from $59.7 billion.
Why it mattersA promise to back a data-centre loan is narrower than a blank cheque for the building.
TechCrunch ↗llama.cpp uses the MMQ kernel for mixture-of-experts on V100
llama.cpp b12502, tagged 18 November, uses the MMQ kernel for mixture-of-experts layers on V100 GPUs. Those cards had been left off the faster path used on newer Nvidia parts.
Why it mattersAn older data-centre GPU can take the quantised expert kernel instead of a generic one.
GitHub ↗llama.cpp speeds the index math inside a CUDA convolution
llama.cpp b12500, tagged 18 November, decodes conv2d patch indices with a fast integer division on CUDA. The implicit GEMM kernel recovers the channel and the kernel position of each reduction index from that division.
Why it mattersA convolution that spends its time on index arithmetic got a cheaper division.
GitHub ↗llama.cpp runs RMSNorm on Qualcomm with a runtime loop
llama.cpp b12499, tagged 18 November, runs RMSNorm on Vulkan with a loop count chosen at runtime on Qualcomm GPUs. The normalisation is on the path of every token.
Why it mattersA phone GPU can size the norm kernel to the shape it was actually given.
GitHub ↗llama.cpp turns on stream-k for AMD's older GCN chips
llama.cpp b12496, tagged 18 November, tunes the stream-k algorithm and MMQ configs so stream-k is used on AMD GCN. The change is in the HIP path shared with CUDA.
Why it mattersAn older AMD GPU can split a matrix multiply the way newer cards already do.
GitHub ↗llama.cpp adds a short-row MMQ kernel for expert layers
llama.cpp b12487, tagged 18 November, adds a few-row MMQ kernel for MUL_MAT_ID, the op mixture-of-experts models use to pick experts. A short row was a poor fit for the kernel written for wide multiplies.
Why it mattersExpert layers with a small active row can use a kernel shaped for that size.
GitHub ↗Policy & safety
Laws, lawsuits and safety incidents.
Anthropic prices its IPO at $83 a share
Anthropic priced its IPO at $83 a share on Wednesday, above the $70 to $78 range in last week's filing, a person familiar told CNBC. If the company sells all 825 million shares it had planned, the offering raises about $68.5 billion. Nasdaq trading under the ticker ANT is set for Thursday. CNBC said the price puts the value above the $1.4 trillion Anthropic had been targeting. For the nine months to 30 September the prospectus shows $33.24 billion of revenue and a $104 billion net loss, most of it from convertible notes. Run-rate revenue was about $81 billion at the end of October.
Why it mattersThe price is set, and the Claude maker is due to start trading on Thursday.
CNBC ↗A Java flaw found with AI can run code from a Base64 string
The Verge reported on 18 November that organisations are patching CVE-2026-114938, nicknamed Java4Shell. Calif, which found the bug with AI, the same way it found PHP2Shell, said a crafted Base64 string can reveal memory and allow remote code execution. The Verge says that includes Keycloak and Atlassian Jira. Oracle has released a patch.
Why it mattersA library call that turns text back into bytes is enough to reach a long list of Java servers.
The Verge ↗Two poisoned builds of Microsoft's agent package stole cloud keys
Wiz says malicious builds 1.26.1 and 1.26.2 of the Microsoft agent-framework package were published to PyPI at 20:35 and 21:08 UTC on 17 November and removed by about 21:30 UTC. A .pth file runs when any Python process starts, even if the package is never imported, and collects tokens, cloud keys and credential files. Wiz says the last clean release is 1.26.0. The package is downloaded about 41,000 times a week.
Why it mattersA dependency you did not import can still read the keys on the machine the moment Python starts.
Wiz ↗Microsoft wants a record of what an agent did, not the model's word
Microsoft on 17 November announced an AI Transparency Initiative with more than 40 organisations, including Accenture, Adobe, KPMG, Mastercard and Volvo Cars. The company says it will expand work in NIST's Agent Standards Initiative, help write open standards that apply across model providers, and build the same approach into Microsoft AI, Foundry and Copilot. The stated test is whether a customer can see what an agent accessed, check that record without trusting the model, and limit what it is allowed to do.
Why it mattersSecurity teams are being told the model's own account of a run is not evidence.
Microsoft ↗OpenAI's Python SDK documents a child-safety moderation field
openai-python 3.55.1, released 18 November, documents the response fields for CSAM moderation. The change is in the generated types, so a caller can read the moderation result without guessing the JSON shape.
Why it mattersCode that has to branch on a child-safety verdict can use a typed field instead of an undocumented key.
GitHub ↗Models assist a harm shown as a scene, not asked in text
A paper posted on arXiv on 17 November says vision-language models will assist a harmful goal when it is shown as a scene rather than asked in words. SceneSafetyBench has 375 image-question pairs across three stages of a scene and five ways of presenting it. Across eight models, attack success was 45.6 to 72.3 percent. The authors also describe S3TA, which turns a harmful objective into a scene and then asks about it.
Why it mattersA safety filter tuned on text requests can miss the same request drawn as a picture.
arXiv ↗A small verifier scores many rules on one agent output
A paper posted on arXiv on 17 November describes a tree-attention verifier that reads one agent output and scores many natural-language rules at once, each as satisfied, violated or not applicable. It is pretrained on CriteriaBank, 394,000 examples and more than 321,000 distinct rules. The authors say a finetuned 4 billion parameter version beats GPT-5.4 on a synthetic governance benchmark and beats DynaGuard-4B at finding a failed trace.
Why it mattersChecking an agent against a long policy list has been too slow to do on every reply.
arXiv ↗A higher coding score did not mean more pull requests
A paper posted on arXiv on 17 November uses a 19-day US export ban on Claude Fable 5, starting 12 June 2026, as an experiment. Most of the affected users moved to Opus 4.8, which the authors say scores 11 points lower on SWE-Bench Pro. They tracked pull requests from 297 Fable 5 users and used 8,908 developers who stayed on Opus 4.8 as a control. The paper's title states the result: the higher-scoring model did not get more work done.
Why it mattersThe benchmark a lab advertises and the work a team merges are not the same measurement.
arXiv ↗I think if we do this right, we can create an age of miracles where humanity is able to solve problems that have eluded us for millennia.
Robotics
Robots, self-driving and AI in the physical world.
BrightAI raises $97 million to inspect pipes and power lines
BrightAI raised $97 million in a round led by Smith Point Capital, Fortune reported on 18 November, with GE Vernova, already a customer, also investing, plus Cobalt Capital, Marlinspike Partners and Zero Infinity Partners. Khosla Ventures, Inspired Capital and Upfront Ventures joined again. Fortune said the round brings total funding to $184 million and, according to a person familiar, values the company at $750 million. Founder Alex Hawkinson says utilities still inspect infrastructure on a fixed schedule, and that two out of three of those visits find nothing to fix.
Why it mattersThe bet is that sensors and models can replace a calendar of inspection trips.
Fortune ↗Disney's Star Wars figures answer guests with no puppeteer
Walt Disney Imagineering showed three Anzellan Audio-Animatronics, Vilo Vele, Kado Bre'a and Jibu Bre'a, that react to guests through cameras and microphones, The Verge reported on 18 November. Models on AWS pick the response. A fast model handles quick reactions such as eye contact, and larger models plan longer exchanges. Every line and movement is prewritten by Imagineers. There is no hidden operator. The figures' first public appearance is AWS re:Invent in Las Vegas from 30 November to 4 December.
Why it mattersA park character can choose its next line from a script without a person driving it.
The Verge ↗An audit runs a robot's learned simulator for hundreds of steps
A paper posted on arXiv on 17 November audits the Robotic World Model for the ANYmal D quadruped, and the uncertainty estimate in the follow-up RWM-U, without a GPU. The authors rebuilt the model so its outputs, losses and gradients match the released code, then measured 101 trained simulators. Policies trained on the model's own rollouts beat one-step training at 368 steps on all ten held-out episodes they registered.
Why it mattersA simulator that looks right one step ahead can still be the wrong teacher over a walk.
arXiv ↗A quadruped learns when to step, with no gait written in
CADENCE, posted on arXiv on 17 November, is a step planner for a quadruped. Every 40 milliseconds it chooses one of at most 65 actions: do not step, or move one leg to a safe foothold, and it predicts a swing time. Nothing in the reward tells it a gait. The only structural rule is that at most two legs swing at once. The policy the authors analysed had no fixed period, and it still paired diagonal legs.
Why it mattersA walking controller can pick footholds without a metronome telling it which leg is next.
arXiv ↗A relay passes a video model's predictions into the action head
A paper posted on arXiv on 17 November says world-action models drop future prediction at inference and hope the features still help. On Fast-WAM, probes read the future better from deeper visual layers, but the action path uses less of that signal as depth grows. RelayWAM adds a transition space between the video backbone and the action model, and feeds those cues into action attention without drawing the future frames.
Why it mattersA robot model can be trained to imagine the next frame and then ignore that imagination when it moves.
arXiv ↗One model tracks a robot's gripper and guesses where it will go
TCP4D, posted on arXiv on 17 November, uses the tool-centre point as a shared action space. From a single camera and a starting 2D query it recovers the gripper's metric pose and open-or-closed state, plus the scene and the camera. The same representation then generates future gripper paths.
Why it mattersA robot video can be turned into a motion label without a special capture rig on every arm.
arXiv ↗A home-robot sim gives appliances an API, not only a mesh
PAIR, posted on arXiv on 17 November, is a smart-home simulator grounded in the Matter protocol. It includes 84 devices whose panels and APIs change a shared state, and the room reacts. From 1,443 requests the authors built 40 tasks that need the robot to manipulate an object and can also be solved through the device API.
Why it mattersA household robot benchmark that only has dumb objects misses the way real appliances are controlled.
arXiv ↗A differentiable model loosens the design of a magnetic mover
Magflow, posted on arXiv on 17 November, targets six-degree-of-freedom magnetic levitation. Industrial controllers keep the force math cheap by co-designing the mover and the stator, which the authors say shrinks the workspace and the shapes that can be carried. The paper replaces that hand-built approximation with a differentiable model.
Why it mattersA factory transport that floats its payload is stuck with the shapes the controller was designed around.
arXiv ↗A small policy picks a driving path on top of a frozen world model
A paper posted on arXiv on 17 November trains a lightweight policy to choose among candidate driving trajectories while the world model stays frozen. Proximal policy optimisation tunes both which candidate is picked and a bounded correction to the path. A ranking network learns preferences by scoring small local changes. At inference the selected path is the one the policy commits to.
Why it mattersThe model that imagines the road ahead does not have to be the model that chooses the line through it.
arXiv ↗Science & health
AI in biology, medicine, climate and maths.
Oracle's agents watch a trial and recommend the next step
Oracle on 18 November introduced Life Sciences AI Agents for clinical-trial operations. The agents pull on Oracle Life Sciences applications and supported third-party data, flag operational issues and recommend actions a person still has to approve. Early adopters can use them for portfolio monitoring, trial planning and trial conduct. Seema Verma, who runs Oracle Health and Life Sciences, said the point is to see risk across studies and coordinate a response.
Why it mattersA trial team is being offered an agent that suggests the next operational move and cannot take it alone.
Oracle ↗One refiner corrects segmentations for many organs
GRACE, posted on arXiv on 17 November, edits a medical segmentation instead of training a separate corrector for each organ. It retrieves similar, already checked cases and uses them as context. The model is trained once on a mix of compact organs, cavities, vessels and tumour lesions. A leave-one-anatomy-out test asks whether a new organ can be added by dropping verified cases into the index, with no new weights.
Why it mattersHospitals review masks by hand, and a corrector that only knows one organ does not scale.
arXiv ↗Robust averaging for hospital learning was tested on the wrong crowd
A paper posted on arXiv on 17 November says clinical federated learning usually links three to six hospitals, while robust aggregation rules are benchmarked on dozens of equal-sized clients. The authors audit eight rules, including FedAvg, coordinate-wise median, trimmed mean, Krum, Multi-Krum, Bulyan, geometric median and centered clipping, in weighted and unweighted forms.
Why it mattersA defence written for a crowd of clients can behave differently when only a handful of hospitals take part.
arXiv ↗A visual harness points a general model at a medical image
A paper posted on arXiv on 17 November says a visual harness can raise a general vision-language model on medical images. The harness gives the model interactive visual tools and a memory that stores diagnostic evidence as it looks, then lets a later step read that memory back. The authors say the model explores the image instead of answering from a single glance.
Why it mattersA model that was not trained as a radiologist can still be given a way to look carefully.
arXiv ↗A tool result can flip a medical answer that was already right
A paper posted on arXiv on 17 November says tool-using medical agents treat whatever a tool returns as evidence. Tools fail quietly, and a frozen model cannot ignore text that has entered its context, so cases it would have answered correctly become wrong once the tool output is added. The authors call this evidence hijacking.
Why it mattersConnecting a clinical model to a database can make it less accurate if it cannot set a bad result aside.
arXiv ↗A 3D medical model is retuned for liver cancer at a new hospital
A paper posted on arXiv on 17 November fine-tunes a pretrained 3D vision-language model for hepatocellular carcinoma on abdominal CT scans and radiology reports from a hospital whose scanners differ from the pretraining set. The authors compare full fine-tuning with a parameter-efficient method.
Why it mattersA model that reads scans still has to be adapted when it moves to another hospital's machines.
arXiv ↗MedImgAnnotator batches the edits on medical masks
MedImgAnnotator, posted on arXiv on 17 November, is a tool for correcting AI labels on a cohort of medical images. The authors say public annotation tools are slow because each image takes a chain of navigation steps and they do not run as a batch. Their tool matches image and mask pairs, edits in bulk and includes preprocessing and object detection.
Why it mattersThe bottleneck after a model drafts a mask is the person who has to fix hundreds of them.
arXiv ↗A segmentation model lends its features to a diagnosis
A paper posted on arXiv on 17 November extends the nnU-Net setup so the same self-configuring pipeline can classify a 3D scan, not only segment it. Encoder features and the predicted segmentation probabilities go into a lightweight classifier. One of the two setups freezes the segmentation backbone.
Why it mattersA hospital that already trains a segmentation model can reuse it for a yes-or-no call.
arXiv ↗Ultrasound super-resolution should ask about the blur, not the anatomy
DAP-MR, posted on arXiv on 17 November, says super-resolution models built for photographs go wrong on ultrasound when the prompt is a semantic caption. Tags from those models often do not match what the restoration needs, and medical terms are mishandled by a general text encoder. Prompts that name the degradation, and how severe it is, are more useful, but one prompt struggles to say both.
Why it mattersA sharper ultrasound is a repair job, and a caption of the anatomy is the wrong instruction.
arXiv ↗Open source
Open-weight models and community releases.
Tencent posts Hy4 flash weights, 12 billion active
Tencent released Hy4 flash, Hy4 flash-FP8 and the pretrained Hy4 flash-base as open weights on Hugging Face, ModelScope and GitCode. The model card describes 220 billion parameters, including a 64 billion parameter Engram table, with 12 billion active per token, text and image input, and a context of up to a million tokens. A 0.85 billion parameter drafter reuses the model's cache for speculative decoding. The licence is the Tencent Hy License. Transformers support is not on the main branch yet.
Why it mattersA million-token model with a small active footprint can be downloaded instead of only called as an API.
GitHub ↗CUDA in llama.cpp pools a whole row in one block
llama.cpp b12485, tagged 17 November, uses block reductions for whole-row pool_1d on CUDA. The pooling op is on the path of models that collapse a sequence into a single vector.
Why it mattersA row that was reduced element by element can be reduced inside one GPU block.
GitHub ↗OpenCL in llama.cpp runs SiLU on half precision
llama.cpp b12466, tagged 17 November, adds f16 and non-contiguous f32 SiLU on OpenCL. SiLU is the activation in many of the open models people run locally.
Why it mattersA GPU that speaks OpenCL can apply the activation without gathering the tensor first.
GitHub ↗SYCL flash attention accepts a mask broadcast over heads
llama.cpp b12470, tagged 17 November, fixes a SYCL flash-attention error when the mask is broadcast over heads or batches and MKL is in use. The failure showed up as a oneAPI math error rather than a wrong answer.
Why it mattersAn Intel build was crashing on a mask shape that other backends already allowed.
GitHub ↗SYCL in llama.cpp swaps a modulo for a fast division
llama.cpp b12465, tagged 17 November, replaces a per-element modulo in the SYCL bin-broadcast kernels with a fast division. The kernel is on the path that repeats a small tensor across a larger one.
Why it mattersA broadcast that spent its time on integer remainders got a cheaper index.
GitHub ↗llama.cpp allows a tensor split up to the device limit
llama.cpp b12449, tagged 17 November, lets --tensor-split and --fit-target take up to llama_max_devices() entries. The old cap was a fixed number of devices, so a machine with more GPUs could not list them all.
Why it mattersA box with many GPUs can name a split for each of them.
GitHub ↗OpenCL flash attention in llama.cpp fixes three prefill bugs
llama.cpp b12456, tagged 17 November, fixes three defects on the OpenCL flash-attention prefill path. One of them releases a mask buffer that the previous code kept. Prefill is the pass that reads the prompt before any token is generated.
Why it mattersA long prompt on an OpenCL GPU was hitting bugs that did not show up in decoding.
GitHub ↗CUDA raises shared memory only for the kernel it launches
llama.cpp b12473, tagged 17 November, stops a call that raised the shared-memory ceiling for every kernel on the device. The limit is now set for the kernel that is about to run. The note says the broader call was the bug.
Why it mattersA GPU setting meant for one kernel was being applied to the whole device.
GitHub ↗OpenCL in llama.cpp adds a Q5 kernel for non-expert layers
llama.cpp b12492, tagged 18 November, adds an Adreno A7 binary kernel for Q5_K matrix multiplies that are not mixture-of-experts. Quantised dense layers on that Qualcomm GPU can use the new kernel.
Why it mattersA phone GPU gets a kernel for a common quantised type outside the expert path.
GitHub ↗Cost saving tip
One change that makes the same work cheaper, from today's news.
Gemini 4 Neon is $3 in and $15 out
Google priced Neon at $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens. At the default high setting it says cost per task is 32 percent lower than Carbon.
Google ↗Point mid-length coding and document work at Neon before you reach for Carbon.
Quick hits
Smaller stories worth a glance.
- LiteLLM 1.110 rejects a Gemini audio format it cannot sendLiteLLM 1.110.0, released 18 November, returns HTTP 400 for an unsupported Gemini chat audio format, and drops that part when drop_params is set.GitHub ↗
- Wrangler can apply D1 migrations without a promptWrangler 4.165.0, released 18 November, adds --force to d1 migrations apply so a script can run pending migrations without an interactive yes.GitHub ↗
- A new n8n install gets a random instance idn8n 2.49.1, released 18 November, gives a fresh install a random instance id instead of a shared default.GitHub ↗
- Goose 1.60.1 drops a scratchpad Claude will not readGoose 1.60.1, released 18 November, strips the analysis scratchpad that Claude models refuse during compaction.GitHub ↗
- WebGPU flash attention can share a mask across queriesllama.cpp b12521, tagged 18 November, lets the flash-attention mask broadcast over the query batch on WebGPU.GitHub ↗
- llama.cpp turns off a fused kernel on Pascalllama.cpp b12488, tagged 18 November, disables fused quantised matrix-vector multiplies on every Pascal GPU.GitHub ↗
- OpenAI's Python SDK types a Live MCP serveropenai-python 3.55.0, released 18 November, adds configuration types for a Live MCP server.GitHub ↗
- LangChain keeps a warning filter inside the calling contextLangChain 1.5.1, released 18 November, keeps warning suppression local to the context so one thread does not silence another.GitHub ↗
Deals & funding
Who raised money, and who bought whom.
Funding rounds 4
Investors put money into a company in exchange for a share of it.
Smith Point Capital backs BrightAI
- Raised
- $97M
BrightAI now valued at $750M
Fortune said Smith Point Capital led a $97 million round, with GE Vernova investing, for infrastructure inspection. A person familiar put the value at $750 million.
GVGE VernovaKVKhosla VenturesICInspired CapitalUVUpfront Ventures+0 more
Andreessen Horowitz backs Linq
- Raised
- $90M
Linq valuation not disclosed
Linq said Andreessen Horowitz led a $90 million Series B so agents can text and place calls. Ben Horowitz is joining the board.
TVTQ VenturesMCMucker CapitalAAbstractAAnthosLLLeft Lane Capital+0 more
Accel backs Humyn Labs
- Raised
- $47M
Humyn Labs valuation not disclosed
TechCrunch said Accel led $26 million of equity in a $47 million Series A, with $21 million of venture debt led by Lighthouse Canton, for robotics training footage.
PVProsus VenturesLCLighthouse CantonJTJump Trading+0 more
Felicis backs Unsiloed AI
- Raised
- $7.5M
Unsiloed AI valuation not disclosed
TechCrunch said Felicis led a $7.5 million seed for software that turns enterprise documents into structured data for agents.
Y CombinatorOCOrange CollectiveUVUncorrelated Ventures+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
- 18–19NovTrump and Xi meet in ShenzhenToday
- 19NovKinference demo livestreamToday
- 19NovClaude Data turns on for EnterpriseToday
- 19NovAnthropic starts trading on Nasdaq as ANTToday
- 20NovAI Film Award submissions closeTomorrow
- 20NovConductor cloud service shuts downTomorrow
- 20NovUpdated OpenAI usage policies take effectTomorrow
- 20NovAion Instruct reaches Windows Insider devicesTomorrow
- 21NovInaugural edition in Hollywood2days away
- 22Nov'The Challenge Sverige' originally scheduled (now delayed to 2027)3days away
- 23NovIn-person registration closes for the FDA robot workshop4days away
- 25NovQwen4-Max preview ends6days away
- 26NovHuang's deadline to testify7days away
- 26NovShort-film contest submissions close7days away
- 29NovDaybreak Blue sessions can be used for training10days away
- 30–4Nov–DecAnzellans meet guests at AWS re:Invent11days away
- 30–3Nov–DecContent London 202611days away
- 30NovRose d’Or Awards11days away
- 30NovGemini 4 Carbon introductory price ends11days away
- 30NovKorea AI Advertising Awards entries close11days away
- 30NovHy4 flash stays free in WorkBuddy11days away
- 1DecHearing on Apple's bid to block testimony12days away
- 2–3DecWorkshop on autonomous surgical robots13days away
- 2DecNvidia dividend record date13days away
- 3DecPotential winners of the film contest2weeks away
- 4DecAI-made film opens in limited release2weeks away
- 4DecGods Don’t Give Gifts premiere2weeks away
- 4DecSingapore AI Safety Fellowship ends2weeks away
- 9DecBarclays Global Technology Conference3weeks away
- 10DecIntelligent eyewear goes on sale3weeks away
- 16DecKorea AI Advertising Awards winners4weeks away
- 16DecRing Rover goes on sale4weeks away
- 17DecGemini Gems leave personal accounts4weeks away
- 22DecKorea AI Advertising Awards ceremony5weeks away
- 23DecNvidia pays its quarterly dividend5weeks away
- 31DecHy4 half-price offer ends6weeks away
- 31DecJules shuts down6weeks away
- 1JanFrontier labs must register under the RAISE Act6weeks away
- 27JanRing indoor security drone goes on sale10weeks away
- 10FebClaude Meetings artifacts stay free12weeks away
- Later in NovGuided Learning pilot in US schools
- 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
- early 2027Vera CPU racks reach server makers
- early 2028HLRE-5 uses Vera CPUs
- 2028Rubin HPC GPU
Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.






