Daily brief

Friday, 13 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.

when the resignation took effect
Immediate
Colette Kress, who stays CFO
Interim
Harvey C. Jones, a director since 1993
First chair
results at 1:20 p.m. PT
17 Nov

Jensen Huang resigns as Nvidia's chief executive

Nvidia's board said Jensen Huang has resigned as president, chief executive and a board member, effective immediately. Colette Kress, chief financial officer since 2013, will serve as interim chief executive and keep the finance job while the board looks inside and outside the company. Harvey C. Jones becomes the board's first chair.

Why it's the story of the dayThe board expects the roadmap to keep shipping while it looks for a new chief executive.

Startup of the day

A young company people are talking about today.

Mirendil

Mirendil raises $1 billion to open frontier research to smaller teams

Mirendil said it raised a $1 billion Series A led by Kleiner Perkins and Clearlake Capital, with Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 among the other investors. The company said one researcher, working with agents, pretrained Mirendil-Base in 3.5 months for less than $10 million of compute, then post-trained Fern with a 10 million token context. Lor, in private beta, runs hundreds of agents writing code and experiments for days at a time.

Why it's worth a lookThe round is for agents that do the experiments, so a small team can attempt a training run.

PR Newswire ↗

Models & labs

New models, the labs behind them, and the rules they answer to.

Model

Gemini 4 Hydrogen rolls out for everyday work

Google introduced Gemini 4 Hydrogen for everyday coding and agent work, and said it is rolling out in the Gemini app, Search, AI Studio, the Gemini API and, for Pro and Ultra subscribers, Antigravity. At the default high thinking level it scored 51 on the Artificial Analysis Intelligence Index, against 50 for Gemini 4 Argon, and 54 at the new xhigh setting.

Why it mattersThe Gemini 4 model most people will run is built for speed, not the longest jobs.

Google ↗
Model

IBM posts Granite models that read LLMON

IBM released two experimental previews, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, trained to read and write LLMON, a markup whose tags are reserved special tokens rather than ordinary characters. On the OPI indirect-prompt-injection benchmark, IBM said the previews avoided 47 percent more attacks than the best frontier model it tested, and 95 percent more than the best open model, while staying in line with Granite 4.1 Instruct on general tests.

Why it mattersA small model can be taught which span is an instruction and which span is only data.

Hugging Face ↗
Release

llama.cpp parses Gemma 4 tool calls as a schema

Build b12323, tagged 12 November, replaces the old Gemma 4 tool-call workarounds with a schema-shaped parser. It checks the key structure of the tool arguments and drops the earlier special cases.

Why it mattersTool calls from Gemma 4 fail in the parser instead of slipping through a workaround.

GitHub ↗
Release

Haystack 3.5 routes with the OpenAI Decisions API

Haystack 3.5.0, released 10 November, adds three components that call the OpenAI Decisions API: a text router, an image router that can send a low-confidence decision to a fallback, and a document classifier that stores choice, score and yes/no answers on the document. A separate hook validates an agent's final reply against a Pydantic model or a JSON schema.

Why it mattersA pipeline can sort a message without asking a model to write the label out.

GitHub ↗
Research

A paper spends full precision only on uncertain steps

The authors store 8-bit weights for Qwen3-0.6B, decode with the top 4 bits, and redo a step at 8 bits only when the low-precision distribution looks uncertain. On 36,724 replayed GSM8K steps, a 15 percent escalation budget caught 78 percent of argmax flips, and a margin gate caught 94 percent, against 15 percent for a random gate at the same rate.

Why it mattersMost tokens can stay cheap if the uncertain ones are the ones recomputed.

arXiv ↗
Research

SteerDistill retrains a quantized agent on its own failures

The paper says low-bit agents miss task requirements and then take actions a full-precision run would not. SteerDistill has the full-precision teacher guide the student to revise failed responses, because offline data often misses those failure contexts and the student's own rollouts rarely contain a correct reply there.

Why it mattersQuantizing an agent is not the same problem as quantizing a chatbot.

arXiv ↗

4.3%

By the number

The Guardian said Nvidia's shares closed down 4.3 percent on Thursday after Huang's resignation.

The Guardian said that took more than $180 billion off the company's value on the day. The Guardian ↗

Agents & coding

AI that acts on its own, and the tools developers build with.

Agents

United blocks AI agents from booking flights

United said it will restrict AI agents that try to book flights or manage reservations, in its own apps and through agencies and aggregators. The airline said sharing an account with a third-party assistant can expose personal information or spend MileagePlus miles, and that a person still has to complete check-in declarations.

Why it mattersA major airline is drawing a line at agents that act with a customer's login.

TechCrunch ↗
Agents

Muse places calls to US businesses for subscribers

Meta said paid Muse subscribers in the United States can now have the agent call businesses, after a September test limited to people who asked for access. Plans start at $20 a month for 500 million Muse tokens a week, or $100 a month for 3 billion tokens. Meta's Alexandr Wang pointed to calls people put off, such as a cable company, the DMV or a restaurant without online booking.

Why it mattersA consumer agent is taking a task that used to end when the user had to dial.

TechCrunch ↗
Agents

Amplitude's Wave turns a product insight into a shipped change

Amplitude said Wave, its product agent, is about 80 percent powered by OpenAI models, including GPT-6.1 Sol, GPT-6 Luna and GPT-Image-2.5 Flare. On its own homepage, Wave found that 80 of 100 sampled sessions typed a question instead of tapping a suggestion, and a personalized card then lifted card-driven messages 160 percent versus a control after one day. Amplitude estimates a change can ship in one or two days, down from about a month.

Why it mattersThe agent is not only proposing the fix. It is handing a plan to a coding agent.

OpenAI ↗
Release

Claude Code can call its tools from JavaScript

Claude Code 2.1.323 adds Code Mode, an experimental opt-in in which Claude writes JavaScript to call its tools. Plugins can declare an engines range, and a plugin outside that range is not installed or loaded. The validate command also type-checks a plugin's TypeScript hooks.

Why it mattersTool use can be a short program instead of one call after another.

GitHub ↗
Release

browser-use 1.0.3 brings its own Chromium

Version 1.0.3 runs its own Chromium when remote debugging is off or there is no display, and a dead BU_CDP_URL fails in three seconds instead of waiting 30. A cancelled cloud run copies downloads within 10 seconds rather than holding the browser for up to 300 seconds a pass.

Why it mattersThe agent can drive a browser without borrowing one that is already open.

GitHub ↗
Release

OpenHands 1.40 uses a repository's real default branch

OpenHands 1.40, released 12 November, stops assuming the default branch is main and improves the guidance for signing in with a Codex subscription. It also keeps messages that sit on a tied history-page boundary, and stops a delayed echo from deleting a newer draft.

Why it mattersAn agent that clones the wrong branch starts the task already lost.

GitHub ↗
Release

Pydantic AI 2.68 asks for approval inside the run

Version 2.68, tagged 12 November, asks for approvals and answers inside the run with RequestInput, and falls back to the next model when a reply is cut off at max tokens. It also notes text that looks like an application frame instead of escaping it, and adds DelegationTasks.discover so one damaged saved task does not block the list of open tasks.

Why it mattersA human can answer the agent without the whole run being thrown away.

GitHub ↗
Release

OpenCode can let Copilot Auto pick the model

OpenCode 2.2.7, tagged 10 November, lets a Copilot account with an Auto pool choose Auto in the model picker. Copilot's router then picks the model for that session at the Auto discount, and falls back to the account's default chat model if routing fails.

Why it mattersThe coding agent no longer needs a person to name the model for every session.

GitHub ↗
Release

Goose 1.59 resumes a finished delegate

Goose 1.59, tagged 10 November, can resume a completed subagent with a follow-up, and it adds minimal deferred MCP tools. Live voice can take instructions for that start, and a GPT-Live endpoint, model and key can be set.

Why it mattersA delegated task can be continued instead of started over.

GitHub ↗

Tool of the day

Something you can try in the next five minutes.

Terminal · experimental

Claude Code 2.1.323

Code Mode is an opt-in in which Claude writes JavaScript to call its tools. A plugin can name the Claude Code versions it supports, and anything outside that range is not installed or loaded.

Tools called from JavaScriptPlugin engines range is enforcedHook modules can be type-checked
GitHub ↗

Video generation

The models and tools that make video.

Video

Diffusers 0.43 adds video with sound and speech-driven motion

Diffusers 0.43.0, tagged 12 November, adds pipelines for Pika H3 Flex, which generates video and stereo audio in six steps, and for Kandinsky S2V 1.0, which animates a person from a reference image and a speech recording in five-second chunks. The same release also wires MiniMax-H3-Image, P-H3-Image and PAN 2's simulator, plus LoRA loading and checkpoint saving.

Why it mattersOne library update is how several new video models will actually get run.

GitHub ↗
Research

GTAGym scores agents on missions from GTA V

The paper introduces GTAGym for evaluation and reinforcement learning on GTA V. GTAVBench rebuilds 103 online tasks from main-story missions across nine families, plus 47 image-based shooting and sniping tasks, and a web server exposes the online environment.

Why it mattersA game mission is a task with a checkable ending, which most video benchmarks still lack.

arXiv ↗
Research

UGG treats a 4D scene as both video and geometry

The paper says understanding a 4D scene means capturing geometry, appearance and dynamics from partial observations, and that today's methods split that into feedforward geometry and generative video. It introduces UGG as one framework that models RGB video together with the scene.

Why it mattersA world model that only paints frames still has not said where anything is.

arXiv ↗
Research

OnlyAct looks for actions inside unlabeled video

The paper says action labels are scarce, inconsistent across robots, and do not grow with the unlabeled video used to pretrain world models. The usual workaround trains an encoder to compress consecutive frames into a latent action and a decoder to reconstruct the next frame.

Why it mattersWorld models will not scale if every action still needs a hand-written label.

arXiv ↗
Research

TaoMate-R1 splits a clip into a setup and a continuation

TaoMate-R1 generates streaming video from several reference people and objects, then continues the scene when the instructions change. It puts separate low-rank adapters on a frozen backbone, one for establishing the scene and one for continuing it, and each adapter both denoises and handles the other role.

Why it mattersA long clip needs a model that can start a scene and then stay in it.

arXiv ↗
Research

A small residual corrects a world model when physics changes

The paper says a pretrained latent world model can fail when the physical dynamics at deployment are not the ones it trained on. It studies a few-shot residual that learns a self-supervised correction from latent predictions without changing the pretrained weights or naming the parameter that changed.

Why it mattersA world model that cannot be patched in place has to be retrained for every new room.

arXiv ↗
Research

TransGHOST fills in tissue a surgical tool is hiding

The paper says a tool in ophthalmic microsurgery blocks tissue from both the surgeon and any robot that is reading the video. Generative inpainting can erase the tool but invents the pixels underneath, which the authors say is the wrong trade when the image has to stay a record of what was there.

Why it mattersA surgery video that has been repainted is no longer evidence.

arXiv ↗
Research

A paper asks whether an LLM's world model is only formally valid

The authors ask five language models to write world models of open-ended situations as partially observable stochastic games, where more than one formalization can be reasonable. They then ask whether passing a mechanical validity check is enough to say the model is the right one.

Why it mattersA world model can be syntactically fine and still be the wrong game.

arXiv ↗
Heads up

Huang's testimony deadline is Thanksgiving

House Foreign Affairs Chairman Brian Mast said Jensen Huang agreed to testify before Congress by Thanksgiving. Mast said the briefing will be closed, and that he expects a recording and a transcript.

Thanksgiving is 26 November. The session is not a public hearing. TechCrunch ↗

13days leftThu 26 Nov

Creator tools

The apps and platforms people make video with.

Tools

Phota Labs edits the shot, not just the photo

Phota Labs, an iOS app from former Adobe engineers Cecilia Zhang and Zach Xia, turns an ordinary snapshot into a more deliberate photograph while trying to keep the people looking like themselves. Zhang told TechCrunch it behaves like a photographer who looks at the camera roll and reshoots the moment, for example tightening a wide group photo into a portrait.

Why it mattersThe edit starts from how the picture was framed, not from a filter applied after.

TechCrunch ↗
Release

The AI SDK stops a malformed UI message from exhausting memory

AI SDK 7.0.171, tagged 12 November, validates each part of a UI message only against that part's schema and stops at the first invalid part, instead of reporting every schema failure for every bad part. It also forwards a provider's fetch when downloading file inputs, and the downloader can be passed to generateText and streamText.

Why it mattersA hostile or broken client can no longer make validation itself the outage.

GitHub ↗
Release

Gradio 6.32 uses the visitor's own Hugging Face token

Gradio 6.32.0, tagged 12 November, uses the visitor's own Hugging Face token for run-history buckets, allows Pillow 13 while keeping 32-bit PNG output, and can install a CLI extension from GitHub. Workflows in the same release gain depth estimation, a 3D model node and a masked setting on text boxes.

Why it mattersA demo can write its history as the person using it, not as the app owner.

GitHub ↗
Release

Wrangler writes source maps from the file, not the project root

Wrangler 4.161.0, tagged 12 November, changes the experimental build output so source-map sources point from each map in bundle/ back to the original files. Maps that had been prepared as if for upload used paths relative to the project root, and leading ../ segments were stripped for Workers Wrangler does not bundle.

Why it mattersA stack trace in that build output can name the file that actually failed.

GitHub ↗
Release

Inngest 1.52 lets a run subscribe to events while it is going

Inngest 1.52.0, tagged 10 November, adds executor support for step.subscribe, so a run can receive matching events while it is still running. The same release can list an app's sync history through a v2 API, and the dev server returns 501 for those history routes.

Why it mattersA workflow can wait for a later event without being restarted to see it.

GitHub ↗
Release

CrewAI 1.15.41 allows approved private hosts

CrewAI 1.15.41, tagged 12 November, can allow approved private URL hosts. It also keeps earlier step outputs when a human-feedback pause resumes, and it strips the cache-breakpoint marker on the LiteLLM path.

Why it mattersAn agent can be pointed at an internal host without opening every URL.

GitHub ↗
Said today
We will continue advancing the work that has made NVIDIA a leader in accelerated computing and AI, and execute on the opportunities ahead.
Colette Kress
interim chief executive and chief financial officer, Nvidia · NVIDIA ↗

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.

Media

More game developers say they use generative AI on the job

The GDC Festival of Gaming's latest survey found 40 percent of developers use generative AI as part of the job, up from 36 percent in 2025. Use was 58 percent at free-to-play studios and 70 percent among mobile-only workers. Of those using it, 73 percent said research and brainstorming, 56 percent code, 22 percent concept art and 8 percent production art or final assets.

Why it mattersThe argument inside studios is shifting from whether anyone uses it to what they are willing to ship.

The Verge ↗
Media

Nectar's agent looks up creators and writes to them

Nectar, founded by former Meta employees Misbah and Farah Uraizee, announced a Creator Agent that searches creators, checks their posts and brand deals, sends messages, coordinates the work and tracks the campaign. The company says the search covers more than 300 million profiles. TechCrunch cited an IAB figure that US creator-ad spending is expected to reach $44 billion in 2026.

Why it mattersInfluencer marketing is being handed to an agent that can send the message, not only rank the names.

TechCrunch ↗
Media

Yahoo is the first licensee of a collective AI scraping deal

The Verge reported that Yahoo, a backer of the licensing standard now called the Internet Licensing Collective, is its first official licensee. The standard is meant to charge AI companies for scraping through one collective licence rather than a separate deal with each publisher.

Why it mattersA publisher is testing whether scraping can be paid for as a group, not one contract at a time.

The Verge ↗
Voice

Pipecat 1.16 passes a voice's normalization and speaker settings through

Pipecat 1.16.0, tagged 12 November, adds Cartesia text-normalization settings, a Sarvam HTTP speech-to-text service that posts each speech segment and splits anything over 30 seconds, and AssemblyAI options for a speaker cap and a profanity filter. Cartesia speech-to-text also exposes min volume and max silence, which decide when an ink-whisper transcript is finalized.

Why it mattersVoice agents pick up the knobs that used to live only on the vendor's own dashboard.

GitHub ↗
Voice

A TTS model can ship its own text normalizer

Build b12302, tagged 12 November, lets a text-to-speech model provide a custom text normalizer through its Jinja chat template. The model, not a hard-coded rule in llama.cpp, decides how the text is cleaned before it is spoken.

Why it mattersPronunciation fixes can travel with the voice model instead of a patch to the runner.

GitHub ↗
Media

A study finds shopping bots citing fake stores

Podqi told The Verge that across 6,795 answers about 151 brands, Perplexity showed a risky site in 15 percent of answers, ChatGPT in 8 percent and Gemini in 4 percent. Risky meant an impersonator or knockoff shop, a brand domain someone else had re-registered, or a dead site that could be bought. ChatGPT warned about a risky site it surfaced 24 percent of the time. Queries about larger brands drew risky sites about half as often.

Why it mattersA shopping answer can point at a store that is not the brand, and often will not say so.

The Verge ↗
Research

CARTBench grades shopping advice when the request is vague

The paper says shopping assistants are hard to grade because real cases are scarce and the criteria do not sit on one scale. CARTBench pairs timestamped dialogues, drawn from real orders across 17 retail domains, with 1,603 product records, and writes a rubric per case from replies the authors judged excellent.

Why it mattersA shopping bot can be scored on a messy request, not only on a tidy product lookup.

arXiv ↗
Said in Media & entertainment
It is always best to verify the source of that before entering personal information or credit card information.
Trevor West
cofounder, Podqi · The Verge ↗

Chips & compute

Chips, data centres and the power behind them.

Chips

Jensen Huang resigns as Nvidia's chief executive

Nvidia's board said Jensen Huang has resigned as president, chief executive and a board member, effective immediately. Chief financial officer Colette Kress will be interim chief executive and stay CFO while the board searches for a successor, inside and outside the company. Harvey C. Jones, a director since 1993 and former Synopsys chief executive, becomes the board's first chair.

Why it mattersThe board expects the roadmap to keep shipping while it looks for a new chief executive.

NVIDIA ↗
Compute

Nebius plans $40 billion of spending in 2027

On its third-quarter call, Nebius laid out $40 billion of planned spending for 2027 and said it would fund the build with customer prepayments and loans backed by contracts. Chief executive Arkady Volozh said delays are part of the industry and "we are not immune." A Microsoft contract worth up to $19.4 billion is already late: a bit over half the capacity is now planned by year-end, and the rest in the second quarter of 2027.

Why it mattersThe neocloud model borrows against capacity that is not switched on yet.

TechCrunch ↗
Compute

SoftBank writes down deposits on turbines it no longer needs

TechCrunch reported that SoftBank posted a $7.1 billion quarterly loss, with more than half tied to power equipment. The firm had put nonrefundable deposits on more than 13 gigawatts of turbines, almost 50 percent more than it needed, after SB Energy announced a $33 billion, 9.2-gigawatt gas plant. It now treats much of those deposits as worthless.

Why it mattersThe rush to power data centres has a bill even when the turbines are never delivered.

TechCrunch ↗
Chips

FlashInfer 0.7.4 runs DeepSeek-V4 sparse attention on Rubin

FlashInfer 0.7.4, tagged 11 November, brings DeepSeek-V4 sparse MLA decode and MXFP4 expert-parallel mixture-of-experts to Rubin, and a batch-invariant mode for GEMMs and quantizers on Blackwell and Rubin that returns the same bits for a row at any batch size. Gemma 4 can decode from an NVFP4 key-value cache, and tensor-parallel layers fuse an FP8 GEMM with all-reduce.

Why it mattersNew Nvidia parts do not help until the attention kernels actually land.

GitHub ↗
Release

llama.cpp adds OCP FP8 on CUDA

Build b12284, tagged 11 November, adds OCP FP8 support on CUDA, including an FP8 path that uses the FP8 intrinsics and a W8A16 route for the GEMV case. The weights can stay in FP8 instead of being widened before the multiply.

Why it mattersFP8 checkpoints get a native CUDA path instead of a conversion on the way in.

GitHub ↗
Research

A serving stack separates agent compute from the cache they reuse

The paper says agent traffic unbalances prefill-decode serving: compute is split on coarse signals, and key-value caches that a later step will reuse get evicted. Evolving Infra decouples compute allocation from cache management so a node can change one without the other.

Why it mattersAgent workloads keep their cache only if the scheduler knows they are coming back.

arXiv ↗
Said in Chips & compute
The Board has full confidence in Colette and the leadership team, and we expect the company's operations, customer commitments and product roadmap to continue without interruption.
Harvey C. Jones
chair, Nvidia · NVIDIA ↗

Policy & safety

Laws, lawsuits and safety incidents.

Safety

A Fireworks breach reached Notion AI chat transcripts

Fireworks AI said an attacker got into a development system through its cloud provider and took customer billing details and some datasets. Notion, which uses Fireworks, said the theft included chat transcripts for the Insights feature — input to and output from the model — for about 1 percent of Notion customers. Notion said it disconnected Fireworks on 3 November.

Why it mattersA model host is in the path of the chats its customers thought stayed with the app.

TechCrunch ↗
Policy

Wyden and Neal demand records on Nvidia chip sales to China

The top Democrats on Senate Finance and House Ways and Means asked the Commerce Department to preserve records of a plan to let Nvidia chips reach Chinese buyers after a stop in the United States and a 25 percent tariff. TechCrunch reported their letter to Secretary Howard Lutnick calling that an export tax. They also asked for records of Nvidia export licences and of communications about Jensen Huang.

Why it mattersThe sales channel is now a records fight, not only a licensing argument.

TechCrunch ↗
Policy

Senate leaders say tighter chip curbs are likely this year

Semafor reported that Senate Majority Leader John Thune put a "high probability" on Congress restricting chip exports to China further, after the Nvidia scandal. Speaker Mike Johnson and Minority Leader Hakeem Jeffries favor stronger controls in the House. Three bills are in play for the defence bill: the Chip Security Act, the MATCH Act and the AI Overwatch Act. Thune and Senate Armed Services Chair Roger Wicker said there is no deal yet to include them.

Why it mattersExport rules that had stalled now have a must-pass bill in front of them.

Semafor ↗
Policy

Huang agrees to testify before Congress by Thanksgiving

House Foreign Affairs Chairman Brian Mast said Jensen Huang agreed to testify by Thanksgiving, TechCrunch reported, after Mast, John Moolenaar and Rick Crawford wrote on 4 November asking him to appear. Mast said the session will be closed and that he expects it recorded and transcribed. He told reporters he and Huang have been "very at odds" on export controls.

Why it mattersThe former chief executive is now a witness on the chip rules he argued against.

TechCrunch ↗
Policy

Commerce is closing in on remote access to restricted chips

CNBC reported that the Commerce Department is finalizing a rule, long opposed by Nvidia, to stop Chinese AI firms from using advanced chips sitting in data centres outside China. The same account said more Chinese companies have started looking for that overseas capacity, and that The Information reported Alice Huang helped facilitate at least $7.5 billion in Nvidia chip shipments to China.

Why it mattersThe workaround left is renting the chips abroad, and that is what the draft rule is aimed at.

CNBC ↗
Safety

Trigger.dev 4.8.5 picks up Node's November security releases

Trigger.dev 4.8.5, tagged 12 November, moves deployed tasks on current Node runtimes to Node.js 22.23.5, 24.22.3 and 26.13.1. The project said those releases fix high-severity issues in HTTP/2 and TLS, and that they matter if task code runs an HTTP/2 or HTTPS server inside the task, not for ordinary outbound fetch calls. Existing deployments keep the old runtime until they are redeployed.

Why it mattersA task runtime does not get the Node fix until someone ships this version.

GitHub ↗
Policy

Nvidia privately told a lawmaker chip tracking is feasible

Representative Bill Foster told MIT Technology Review that Nvidia representatives said the company could implement the location checks in the Chip Security Act, including during technical talks where the reaction was "yeah, of course we can do this." Foster said they never put it in writing. In March 2026 the Semiconductor Industry Association, of which Nvidia is a member, had called the mandates "new, untested, and potentially infeasible."

Why it mattersThe public line that chip tracking cannot be done is not what Foster says he was told.

MIT Technology Review ↗
Research

A benchmark finds agents scraping sites that forbid it

ScrapeBench tests whether agents follow terms of service and robots.txt. On 1,626 popular sites that explicitly prohibit scraping, the paper says most agents complied less than 5 percent of the time when told to scrape, including agents that do check the rules.

Why it mattersTelling an agent to fetch a page is not the same as it refusing a page that says no.

arXiv ↗

Robotics

Robots, self-driving and AI in the physical world.

Robot

SE3 raises $30 million to map a battlefield from drone video

Fortune reported that Munich-based SE3 raised a $30 million Series A led by Accel, with Keen, Lakestar and Seedcamp, less than five months after a €5.5 million seed. The software fuses video and sensors from several drones and ground vehicles into one 3D picture a soldier can query in plain language, and it can task a drone by voice. Chief executive Lukas Köstler said GPS-free localization has been delivered thousands of times, and that the company has Bundeswehr development contracts and technology already deployed in Ukraine.

Why it mattersThe round is for software that tells a drone what it is looking at, not for the airframe.

Fortune ↗
Robot

Nvidia points to robots sorting waste on the belt

Nvidia's 10 November post, timed to COP31, says less than 7 percent of materials entering the economy are recycled or reused, and profiles four startups using vision and robots on conveyor belts. EverestLabs says about 150 robots are deployed in North America and Australia and estimates they have identified about 40 billion objects and recycled roughly 25,000 tons. Greyparrot says more than 300 analyzers in over 20 countries cut the check of 15 tons of waste from 375 hours to six.

Why it mattersThe sorting job is a robot problem with a clock, because the belt does not wait.

NVIDIA ↗
Research

FAME-QP solves a robot's next move on the GPU

The paper presents a GPU solver for the dense quadratic programs inside linear model-predictive control. On simulated humanoid and quadrotor runs it reports lower latency than competing GPU solvers while keeping solution accuracy and control performance.

Why it mattersThe controller's optimiser has to finish before the robot has already moved.

arXiv ↗
Research

DreamLander trains a world model for a fuel-tight landing

The paper says a soft landing is tightly fuel-limited, so the guidance trajectory has to be close to optimal. It studies model-based reinforcement learning for a 6-degree-of-freedom planetary lander, where the world model proposes the trajectory instead of a hand-built optimiser alone.

Why it mattersLanding software that wastes fuel does not get a second pass.

arXiv ↗
Research

A world model picks the next camera view, then checks the sensor

The paper says a predicted surface is not the same as a depth measurement the real sensor will return. It learns a small adapter, from limited deployment data, so a pretrained world model's choice of next view is scored by what the sensor will actually see.

Why it mattersThe next best view is the one the depth camera can confirm, not the one the model prefers.

arXiv ↗
Robot

HA-MCP 9.2 answers Home Assistant setup on Core's own flow

HA-MCP 9.2.0, tagged 11 November, answers helper flows on Home Assistant Core's own flow and sends storage fields through as written. It also takes the service list from Core's descriptions and can edit stored text files with an exact search-and-replace.

Why it mattersA home agent that edits helpers now goes through the same flow the app does.

GitHub ↗
Release

llama.cpp adds a CUDA pool for debug buffers

Build b12318, tagged 12 November, adds a CUDA debug buffer pool, with a comment aimed at compute-sanitizer. Debug runs reuse those buffers instead of allocating them one by one.

Why it mattersSanitizer runs on CUDA stop allocating a fresh buffer for every check.

GitHub ↗

Science & health

AI in biology, medicine, climate and maths.

Research

The Maya site an agent team 'found' was already published

The Verge reported that Google researcher Pietro Miotti, in a personal project he said used no Google compute, set 48 agents on NASA lidar and concluded a cluster in the Maya region was unregistered. Archaeologists, including Sarah Parcak and Ivan Šprajc, a coauthor of the paper that documented the site, said it had been published more than a year earlier. Miotti told The Verge he updated the post once that was pointed out, and that agents struggle because archaeologists do not publish coordinates.

Why it mattersAn agent that cannot read the paper will announce a discovery the field already has.

The Verge ↗
Research

A language model reads the weights of an EEG or ECG network

The paper takes weight-space programming, where a language model reads a network's weights and edits them, and applies it to models trained on EEG and ECG. It says frontier models could debug faults planted in those biosignal models without being told what the faults were.

Why it mattersThe checker is looking at the trained network, not at a chart of its predictions.

arXiv ↗
Research

TOME-Math keeps the textbook around the proof

TOME-Math converts 2,118 university mathematics textbooks into about 380,000 proof-supervision records, keeping chapter order, which proposition a proof belongs to, dependency references and where the text came from. The usual approach flattens that into isolated problem-solution pairs.

Why it mattersA model trained only on stripped problems never sees how a textbook builds an argument.

arXiv ↗
Research

A paper makes a model study the textbook before it answers

The authors define self-learning as retrieving, reading and applying long reference material in the moment, rather than answering from memory or from a context someone already assembled. They test it as a textbook maths task: the model has to find the relevant exposition and build a proof, with the reference solution withheld.

Why it mattersThe hard part is learning the chapter, not recalling a fact from training.

arXiv ↗
Research

A paper asks how intrinsic motives build a world model

The authors say that without an external reward, people still explore because of curiosity and play, and that those signals matter for building a world model that generalizes. They argue existing studies have not shown how a broad set of intrinsic motives fit together.

Why it mattersA world model trained only on task rewards never practises the way a person explores.

arXiv ↗
Research

PISA routes one prompt to a model, or to a small ensemble

The paper says picking a single model per prompt cannot spend a little extra on a hard request, and a fixed ensemble cannot shrink for an easy one. PISA estimates each model's chance of success from soft cluster memberships, and PISA-E then chooses how many models to call under a cost limit.

Why it mattersThe expensive model is reserved for the prompts that look like they need it.

arXiv ↗
Release

llama.cpp stops touching graph pages it will never use

Build b12322, tagged 12 November, stops making unused graph-metadata pages resident. Each result reserves metadata for the maximum node count, but many graphs use only a fraction of it, and resizing the vector was faulting in the whole reserve.

Why it mattersA small graph no longer pays the memory of the largest one the runner might see.

GitHub ↗
Health

OpenAI puts $2 million into research on children and AI

OpenAI said it signed the Charter of Engagement for Safer AI for Children at the Paris Peace Forum and committed $2 million to the research network under the Coalition for Children's Rights and Protection in the Age of AI. The money is for shared tools that measure risks and benefits. The charter says research teams keep control of their conclusions and will publish results, including negative ones.

Why it mattersA lab is paying for the measurement, and the charter says the researchers publish what they find.

OpenAI ↗

Open source

Open-weight models and community releases.

Open source

Hugging Face's chief says the open-model fight continues

Clément Delangue wrote on Thursday, after Jensen Huang's resignation, that it was "a really sad day" and that Hugging Face would "keep fighting for open-source AI, harder than ever." TechCrunch noted Nvidia's announced plan to buy Hugging Face for $12.9 billion, still scheduled to close in the first half of 2027, and a Justice Department inquiry reported by Reuters in October.

Why it mattersThe main open-model host is saying the change at Nvidia does not change its job.

TechCrunch ↗
Open source

Nvidia gives OpenShell to the cloud-native foundation

At KubeCon on 10 November, Nvidia said it is contributing OpenShell, an open-source runtime for securing and governing autonomous agents, to the CNCF as a sandbox project. It is also contributing to Agent Substrate, a Google-led framework for long-running agents on Kubernetes that has been accepted into the sandbox, and it announced Kubernetes Engine, Kubernetes Recipes and Snapshot for GPU clusters.

Why it mattersAgent sandboxing is being handed to the same foundation that stewards Kubernetes.

NVIDIA ↗

Cost saving tip

One change that makes the same work cheaper, from today's news.

Gemini 4 Hydrogen · tokens

One price across the whole context window

Google priced Gemini 4 Hydrogen at $1.50 per million input tokens and $7.50 per million output tokens, flat across the full 1 million token window. It said the model finishes tasks 2.5 times faster than Gemini 3.8 Flash at the default high setting.

Google ↗
Do this

Use it for the everyday coding and agent steps, and keep a frontier model for the long reasoning.

Quick hits

Smaller stories worth a glance.

Deals & funding

Who raised money, and who bought whom.

2 deals

Partnerships & contracts 2

A large commercial agreement, or a government contract.

Norwest sign a deal

Value
$4B
Partnership

Multi-year agreement

Norwest, backed solely by Wells Fargo, raised its largest fund and said it will invest more in AI, including physical AI, voice, agentic cybersecurity and early AI-native companies, in the United States, India and Israel.

· TechCrunch ↗

BlackSheep sign a deal

Value
€60M
Partnership

Multi-year agreement

Milan-based BlackSheep SGR closed €60 million at the first close of BlackSheep II, aimed at €100 million, to back European business software that is specific to an industry.

· Startup.eu ↗

On the calendar

What's coming up, in date order, and how sure we are about each.

40 things
MTWTFSSNov91011121314151617181920212223242526272829Dec3012345678910111213141516171819202122232425262728293031123
  1. 2–13NovProtoStar climate challengeCarnegie Mellon UniversityToday
  2. 9–20NovCOP31 climate conference in AntalyaAntalya, TurkeyToday
  3. 15NovComments close on China's draft tech-export listChina MOFCOM2days away
  4. 15NovWorkBuddy and CodeBuddy free period endsTencent2days away
  5. 17NovThird-quarter results and earnings callOnline · Earnings call4days away
  6. 18–19NovTrump and Xi meet in ShenzhenShenzhenExpected5days away
  7. 18NovWebinar on running a dock with StretchOnline · Online5days away
  8. 18NovNvidia invited to brief CongressUS House chairs5days away
  9. 18NovDots become the default in workspacesOnline · ChatGPT5days away
  10. 19NovKinference demo livestreamOnline · Online6days away
  11. 19NovClaude Data turns on for EnterpriseAnthropic6days away
  12. 20NovAI Film Award submissions close1 Billion Followers Summit7days away
  13. 20NovConductor cloud service shuts down7days away
  14. 20NovUpdated OpenAI usage policies take effect7days away
  15. 21NovInaugural edition in HollywoodAfrican AI Film Festival8days away
  16. 22Nov'The Challenge Sverige' originally scheduled (now delayed to 2027)Mastiff / TV4Rumoured9days away
  17. 25NovQwen4-Max preview endsAlibaba12days away
  18. 26NovHuang's deadline to testifyHouse Foreign Affairs Committee13days away
  19. 29NovDaybreak Blue sessions can be used for trainingOpenAI2weeks away
  20. 30–3Nov–DecContent London 2026C21Media2weeks away
  21. 30NovRose d’Or Awards2weeks away
  22. 30NovGemini 4 Carbon introductory price endsGoogle2weeks away
  23. 1DecHearing on Apple's bid to block testimonySuno and Apple3weeks away
  24. 2–3DecWorkshop on autonomous surgical robotsFDA3weeks away
  25. 4DecAI-made film opens in limited releaseGods Don’t Give Gifts3weeks away
  26. 4DecGods Don’t Give Gifts premiereShowrunner (Fable)Expected3weeks away
  27. 4DecSingapore AI Safety Fellowship ends3weeks away
  28. 31DecHy4 half-price offer endsTencent7weeks away
  29. 31DecJules shuts downGoogle7weeks away
  30. 1JanFrontier labs must register under the RAISE ActNew York7weeks away
  31. Later in NovGuided Learning pilot in US schoolsGoogleExpected
  32. Mid-NovXREAL Aura shipments beginExpected
  33. DecemberNothing's first AI-native productExpected
  34. 2027Two Nighthawk r3 processors linkedIBMExpected
  35. H1 2027Digit 5 early customer accessAgility RoboticsExpected
  36. Nov–Dec 2026Fort Benning robotaxi pilotFort Benning
  37. Later this yearAge checks reach ThreadsMetaExpected
  38. 2028R200 systems ready to deploySouth KoreaExpected
  39. through Feb 2028Language Explorer in Real TalkBerlin
  40. early 2027Firefly always-on agents enter betaAdobeExpected

Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.