Daily brief

Thursday, 12 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.

Series C, as the government stated
>£500m
valuation after the round
>£5bn
the round Fractile closed
Series C
backers named in the announcement
5

Britain backs Fractile past £500 million

The UK government said Fractile, a chip firm in London and Bristol, has raised a Series C worth more than £500 million. Sovereign AI and the British Business Bank are in the round, with CapitalG, Lightspeed and Redpoint. The government said the round values the company at more than £5 billion.

Why it's the story of the dayA government fund is buying into the chips that would run British AI.

Startup of the day

A young company people are talking about today.

Mercor

Mercor raises $750 million at a $20 billion valuation

Mercor said its Series D is $750 million and values the company at $20 billion, including an employee tender. Named participants include Benchmark, Felicis, Fundomo, Apollo affiliates, G Squared, General Catalyst, Glade Brook, Nvidia, Salesforce, Sapphire Ventures, StepStone, XTX Ventures and Zoom. The company said it now has more than five million experts in its network and has opened offices in New York and London.

Why it's worth a lookThe round is for the evaluations and practice environments enterprises use to train their own agents.

Mercor ↗

Models & labs

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

Model

Gemini Decision 1 scores the choice in one pass

Google opened the Gemini Decisions API, powered by gemini-decision-1, in Google AI Studio. One request can score up to 1,000 yes/no, multiple-choice or rubric questions on text, images, audio, video or PDFs, and Google said the end-to-end latency is 150 to 250 milliseconds at $0.10 per million input tokens, with no output tokens billed.

Why it mattersA routing or policy check no longer has to wait for a generated answer.

Google ↗
Model

Baidu posts ERNIE 5.5 Flash Preview

Baidu released ERNIE 5.5 Flash Preview, the first post-trained model on its ERNIE 5.5 Flash Base, with 280 billion parameters, 14 billion active and a 1 million token context. The API is on Qianfan, and Baidu said the weights will be open-sourced under Apache 2.0. It called the preview an intermediate checkpoint while post-training continues.

Why it mattersA long-context coding and design model is on an API before the weights land.

Baidu ↗
Model

Qwen4-Max opens as a two-week preview

Alibaba's documentation describes qwen4-max-preview as a multimodal model with a 1 million token context and up to 393,216 output tokens, shared by thinking and the answer. It is available only through Token Plan, from 10:00 on 11 November to 24:00 on 25 November 2026, Beijing time, and Alibaba said the preview is then retired.

Why it mattersThe first Qwen4 model developers can call has a published end date.

Alibaba Cloud ↗
Markets

OpenAI's quarter reportedly passed $10 billion

TechCrunch reported, citing Bloomberg, that OpenAI's quarterly revenue crossed $10 billion for the first time, up from a reported $6.7 billion in the second quarter. The same report said the adjusted operating margin moved from -150 percent to -109 percent, and that the annual revenue run rate had reached $50 billion.

Why it mattersThe figure, if it holds, is the public marker for how fast ChatGPT's business is growing.

TechCrunch ↗
Model

BotPsy puts a 0.8B agent on the device

Tether AI Research introduced BotPsy, a 0.8 billion parameter model built on Qwen3.5-0.8B for tool use. On the post, BFCL v4 rises from 23.4 to 44.0 against its base, ACEBench from 30.1 to 58.7, and tau-bench from 23.2 to 51.5. Weights are on Hugging Face in safetensors and GGUF.

Why it mattersA personal agent that calls tools can run without a frontier API bill.

Hugging Face ↗
Access

More than 20,000 veterans claimed a free year of Plus

OpenAI said more than 20,000 transitioning servicemembers and veterans have claimed a year of ChatGPT Plus at no cost. OpenAI Academy and partners have hosted more than a dozen trainings for nearly 1,000 people in the military community, with further workshops listed through late November.

Why it mattersA free year of the paid product is already in use, with ID.me checking who qualifies.

OpenAI ↗
Research

A paper repairs a loop by editing a few weights

Tree-Ising treats a repair of a language model's reasoning loop as a sparse edit over its own BF16 weights. The authors contrast it with a GPT-6 Astra agent that corrected loops in Qwen3.5-0.8B without gradients, using 158 turns, 89.7 million tokens and 50 forward passes.

Why it mattersA loop in a small model can be a weight edit, not another long agent run.

Hugging Face ↗

>£5 billion

By the number

The UK government said that is Fractile's valuation after the Series C.

The round itself is worth more than £500 million. GOV.UK ↗

Agents & coding

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

Agents

Databricks' Genie takes a job and keeps going

TechCrunch reported that Databricks unveiled an updated Genie that can run multistep jobs over an extended period, using an ontology of the data a customer stores on the platform. Chief executive Ali Ghodsi showed it working through a weak part of an employee survey and a sales pitch that had gone badly. He said Unity Gateway can cap use and swap in a cheaper open-weight model.

Why it mattersThe agent is being pitched as something a business leader leaves running.

TechCrunch ↗
Funding

Plurai raises $13 million to rehearse an agent

TechCrunch reported that Plurai raised a $13 million seed led by Team8, with U&I, Velocity and Mercer. The company builds a digital twin of a customer's systems, maps the agent's flows, and prepares evaluations for edge cases. It works alongside Claude Code, Codex, Copilot and Cursor.

Why it mattersThe round is for practice runs before an agent touches the live tools.

TechCrunch ↗
Release

LangChain 1.5 drops Python 3.10

langchain 1.5.0 requires Python 3.11 and langchain-core 1.8.0. Installs on Python 3.10 still resolve 1.4.x. The release also adds model-router and auto-mode middleware, and marks ModelRouterMiddleware as beta.

Why it mattersA project still on Python 3.10 will not pick up this line.

GitHub ↗
Release

browser-use 1.0.2 brings its own Chromium

Version 1.0.2, tagged 11 November, drives its own Chromium when Chrome remote debugging is off, and tells the agent how to switch to a person's Chrome. CDP calls from the CLI and MCP wait 20 seconds, a page frozen by a JavaScript dialog fails fast, and open() in MCP browser code uses UTF-8 on Windows.

Why it mattersA server without a desktop Chrome can still run the browser agent.

GitHub ↗
Release

OpenHands 1.39 filters automations by owner

OpenHands 1.39.0, released 11 November, can filter the automations dashboard by owner, with a service-accounts option. Stop-hook feedback stays out of the chat, and resume timestamps are read as UTC.

Why it mattersA shared workspace can show one person's automations, including the service accounts.

GitHub ↗
Release

CrewAI 1.15.40 looks up a past state by row

Release 1.15.40, tagged 11 November, removes ChromaDB from the embedding-function layer, adds an exact historical state lookup by row id, and wires JSONL into RAG data-type discovery. Standalone agents scaffold as JSON by default. after_llm_call hooks run on async calls, and an agent stops when CrewAI AMP cannot resolve its apps.

Why it mattersA saved run can be opened by its row, and a missing app stops the agent.

GitHub ↗
Work

A law firm says 87 percent of enabled staff use ChatGPT

OpenAI's 12 November interview with Gilbert + Tobin chief executive Sam Nickless says 87 percent of enabled ChatGPT users at the Australian firm are active, more than twice the firm's usual rate for other tools. He said a confidential partner-review job for 120 partners, run in ChatGPT Work, cost about $62.

Why it mattersThe number is active use inside one firm, not a mandate.

OpenAI ↗
Research

A memory keeps the trajectory behind the lesson

WikiTrace stores each distilled lesson with a summary of why its trajectory succeeded or failed, and with the raw trajectory. On a new task the agent can walk those links before it acts. The authors say it scored 3.2 points higher on average, across five benchmarks, than the same lessons with no evidence attached.

Why it mattersAn agent can check a stored lesson against the run it came from.

Hugging Face ↗

Tool of the day

Something you can try in the next five minutes.

Open source · Python

browser-use 1.0.2

The patch lets the command-line agent drive its own Chromium when Chrome remote debugging is off, and it tells the agent how to switch back to a person's Chrome. CDP calls wait 20 seconds, a page stuck on a JavaScript dialog fails fast, and Windows file calls in browser code use UTF-8.

Own Chromium when debugging is off20 second wait on slow pagesUTF-8 file calls on Windows
GitHub ↗

Video generation

The models and tools that make video.

Video

PAN 2 simulates the next 20 seconds of a scene

The Institute of Foundation Models at MBZUAI released PAN 2 on GitHub, an open world model that takes an optional image plus a written action and simulates a video of up to 20 seconds. A language model writes the next state, and a 29 billion parameter diffusion model, decoded with the Wan 2.2 VAE, fills in how it looks. The repository is under Apache 2.0.

Why it mattersAn agent can watch a candidate action before anything moves in the real scene.

GitHub ↗
Video

NTT's video model checks a lift before the robot moves

NTT said it integrated a physics control adapter into a video generation model so a robot can score candidate actions for a collapsed load or a dropped object. In a shelf-loading test the success rate rose from 73.1 percent to 92.3 percent. Field trials with Mujin are planned, starting with airport baggage.

Why it mattersThe clip is a prediction of the lift, scored before the arm moves.

NTT ↗
Research

A paper distils streaming video without a second score network

The authors extend APEX to streaming video by using a shifted condition, inside the same generator, to estimate the velocity of the generated distribution. That removes a separate fake-score network. They keep S2-Avatar's frozen teacher, Self-Replay Forcing and perceptual regularization, and evaluate on 180-second clips.

Why it mattersFewer networks in the distillation means a long stream can be trained on less memory.

Hugging Face ↗
Research

A rubric scores which physical rule a clip broke

PhyRubric builds a caption-specific rubric before any video is seen, splitting the caption into semantic and physical items, then scores a clip from item-level verdicts with timestamps. The authors say a single intuitive score does not show which physical principle failed.

Why it mattersA text-to-video test can name the broken rule, not only the overall grade.

Hugging Face ↗
Video

pi-H3 makes a five-second clip in 3.8 seconds

Stanford researchers announced pi-H3 on 10 November, a few-step version of MiniMax H3 that predicts pixels on the last step and skips the video decoder. On one B200 it makes a five-second HD clip with stereo audio in 3.8 seconds, and a 15-second clip in 18.6 seconds. A Hugging Face demo is live; weights and inference code are planned for 16 to 22 November.

Why it mattersA short clip can finish before the video itself would have played.

Hansheng Chen ↗
Video

Avid puts video generation inside Media Composer

Avid announced Ready to Play on 10 November, an optional Media Composer extension built with Google's Gemini models and Veo. Editors can generate a clip from text or a still, change a shot with a prompt, or add a description to clip metadata. Avid charges nothing extra; model use is billed through an eligible Google Cloud account, and generated media is saved beside the original.

Why it mattersA placeholder or a cleanup can be made in the timeline the edit already uses.

Avid ↗
Video

Bandicam turns a script into a narrated video

Bandicam launched Super Narrator in Bandicam AI Studio on 10 November. A topic, or a script of up to 30,000 characters, can become a narrated video with a target length of up to 30 minutes, with a chosen voice, generated or uploaded visuals, captions and music. The result downloads as an MP4, or as media and a timeline for Premiere Pro or DaVinci Resolve.

Why it mattersA long narration can be assembled in one pass, then opened in a normal editor.

Bandicam ↗
Heads up

Qwen4-Max preview closes on 25 November

Alibaba's documentation says qwen4-max-preview is available only through Token Plan, from 10:00 on 11 November to 24:00 on 25 November 2026, Beijing time. The preview model is retired when that window ends.

If you want to try the preview, call it before 25 November, Beijing time. Alibaba Cloud ↗

13days leftWed 25 Nov

Creator tools

The apps and platforms people make video with.

Tools

Claude Fable 5.2 rolls out in Microsoft Copilot

Microsoft said on 11 November that Claude Fable 5.2 is rolling out in Copilot Cowork and Copilot Studio. Work IQ grounds the model in an organisation's files, meetings, chats and business data, within existing permissions. Admins control access in the Microsoft 365 admin centre, and availability can vary by region and organisation.

Why it mattersThe model Anthropic shipped this week is now a choice inside Copilot, not only in Claude.

Microsoft ↗
Release

The AI SDK parses a long array as it streams

AI SDK 7.0.170, tagged 11 November, parses and validates completed elements as they arrive when streaming a long array with Output.array or streamObject, instead of reprocessing the whole array on every chunk. A later stop or dispose also cancels a pending send-now, so a replaced chat does not send the abandoned queue.

Why it mattersA long structured stream stops burning CPU on text it has already accepted.

GitHub ↗
Release

n8n 2.48.1 keeps an expression after a replacement

n8n 2.48.1, tagged 11 November, keeps expression transfers working after built-in replacements, and matches bare node references across expression engines. The Edit Image node registers a font once, rather than on every text operation.

Why it mattersA workflow that rewrites a value no longer drops the expression that referred to it.

GitHub ↗
Release

A2UI 0.3 can lint a catalog

The a2ui CLI 0.3.0, tagged 11 November, adds a2ui lint to check catalog JSON schemas against protocol rules. The release requires a2ui_core 0.7.0 or newer, because the CLI exposes those types in its public API.

Why it mattersA catalog of agent UI components can be checked before it is served.

GitHub ↗
Release

LiteLLM 1.109 retires the disk cache

LiteLLM 1.109.0, tagged 11 November, removes the disk cache backend, so cache_params.type must be local or redis. String fallback targets become lists of model names, a hard rpm cap needs optional_pre_call_checks set to enforce_model_rate_limits, and temperature, top_p and top_k must be dropped or drop_params set for Gemini 3.6 Flash, 3.5 Flash-Lite and newer. The release also adds mid-stream fallback, so a cut-off stream continues on the fallback model.

Why it mattersAn upgrade changes cache, fallbacks and which sampling fields Gemini will accept.

GitHub ↗
Tools

Hugging Face shows how to serve a decision model

A 10 November Hugging Face post describes decision models that return a probability over preset answers instead of generated text, and shows a Transformers pipeline named decision. The post also covers serving with llama.cpp, finding models on the Hub, and training with Hugging Face tooling. The worked example uses autotrust/GEV-26B-Decide.

Why it mattersThe Hub now has a path for models whose job is a labelled choice.

Hugging Face ↗
Billing

A Hugging Face bill can land on Google Cloud

Hugging Face said on 10 November that an organisation can subscribe on Google Cloud Marketplace and link the organisation, so Hub charges appear on the Google Cloud bill. The post lists private models and datasets, Jobs, Inference Endpoints, Inference Providers and GPU Spaces as the usage this covers.

Why it mattersA team that already pays Google Cloud can add the Hub without a separate card.

Hugging Face ↗
Release

A dev build of HA-MCP talks to Home Assistant

homeassistant-ai/ha-mcp tagged v9.3.0.dev3496 on 11 November, a development snapshot of the custom component and its MCP server, ahead of a 9.3.0 stable release. The install name in the notes is ha-mcp-dev==9.3.0.dev3496.

Why it mattersAn assistant can be pointed at a Home Assistant install from a pre-release build.

GitHub ↗
Said today
For too long, Britain has produced some of the world's best scientific ideas, only to see too much of the technology, industry and wealth they create built elsewhere.
Kanishka Narayan
AI Minister, United Kingdom · GOV.UK ↗

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

JioStar's assistant finds a clip across 15 languages

OpenAI's 11 November interview with JioStar chief architect Vijay Seshadri says the company built a conversational video and cricket-stat assistant into JioHotstar with the OpenAI API. JioStar's platforms are used by more than 500 million people and have supported 72.5 million concurrent live-sports viewers. He said more than 60 percent of users speak the request, and the team built the experience for more than 15 languages in about eight weeks.

Why it mattersDiscovery on a live-sports app is a spoken request, not another row of posters.

OpenAI ↗
Voice

DeepL's Aure follows a change in tone

DeepL unveiled Aure on 11 November, the model now powering DeepL Voice. It keeps updating the translated voice as a speaker's tone, emotion, emphasis and pace change, using both the translated words and the source audio. At launch it supports English, German, Spanish, French, Italian, Japanese, Portuguese and Korean, and it applies automatically in meetings, in-person conversations and the voice API.

Why it mattersA live translation can carry the delivery, not only the sentence.

DeepL ↗
Voice

ElevenLabs and Nvidia add 15 megawatts for voice agents

ElevenLabs said on 11 November that Nvidia Nemotron 3.5 Super VL is available in ElevenAgents, and that it plans to add 15 megawatts of Vera Rubin capacity on top of its Blackwell Ultra systems. The same post says the companies are using the stack for a sovereign voice agent on Poland's national e-health platform.

Why it mattersThe voice-agent host is buying power, not only another model name.

ElevenLabs ↗
Voice

Scribe v3 cuts word errors and labels speakers live

ElevenLabs launched Scribe v3 for recorded audio and Scribe v3 Live for realtime on 10 November. It said v3 makes 27 percent fewer word errors than Scribe v2 across 102 languages, and that v3 Live makes 43 percent fewer errors than Scribe v2 Realtime and adds speaker labels in real time. Both are in the API as scribe_v3 and scribe_v3_live.

Why it mattersA live transcript can say who is speaking, which meeting tools have been missing.

ElevenLabs ↗
Travel

Singapore Airlines says Kris satisfaction doubled

OpenAI said on 10 November that Singapore Airlines moved the Kris chatbot to the OpenAI API and saw a 2x rise in its customer-satisfaction score. AI-assisted tools for staff cut average handling time by about 10 percent, Flight Recommender monthly users grew about 3x over the past year, and more than 60 percent of employees use the internal Jarvis platform monthly.

Why it mattersThe airline's public number is a satisfaction score on the bot customers already use.

OpenAI ↗
Media

Paramount's chief calls AI a bigger shift than CGI

On Paramount Skydance's earnings call, chief executive David Ellison said it is hard to imagine AI having a smaller effect on film than CGI. He said Terminator cost $4 million, and that the same picture shot for shot would cost $60 to 70 million today. He said the company is building its own tools, and that AI is a tool for artists and will never replace them.

Why it mattersA major studio is tying AI to lower budgets for original films, not to replacing the cast.

The Verge ↗
Media

RYET puts RMB 2 million into AI scripts and films

Ruanyun Edai said on 11 November that its China entity and a co-investor will each put in RMB 1 million to make AI-assisted scripts and films on its YeeZo platform. Over 12 months it plans about 400 scripts and about 30 AI films. Profit is split 60 percent to the company and 40 percent to the co-investor, and neither side is guaranteed its money back.

Why it mattersThe test is whether the films sell, because the capital is paid back only from profit.

GlobeNewswire ↗
Media

Kaitoi's Motion Studio builds an ad from a sentence

Kaitoi Labs said on 11 November that Motion Studio, a tool the platform built for itself, turns a description into an animated logo, promo or social clip. Users can start from one of 22 templates or a conversation, and ChatGPT or Claude can drive it through MCP. Studio seats cost $199 a month; new accounts start with free credits and then pay as they go.

Why it mattersAn ad bumper can be revised in chat, including by an assistant the team already uses.

Kaitoi ↗
Said in Media & entertainment
Content discovery is inherently visual, but combining those visual options with spoken preferences and conversation creates a powerful new paradigm.
Vijay Seshadri
chief architect, JioStar · OpenAI ↗

Chips & compute

Chips, data centres and the power behind them.

Chips

Britain backs Fractile past £500 million

The UK government said on 11 November that Fractile, based in London and Bristol, raised a Series C worth more than £500 million, valuing the company at more than £5 billion. Sovereign AI and the British Business Bank joined CapitalG, Lightspeed and Redpoint. The government described the work as chips, software and machine-learning research aimed at faster and more energy-efficient hardware.

Why it mattersPublic money is in the round for a British inference chip.

GOV.UK ↗
Compute

IBM designs Nighthawk r3 for 10,000 operations

IBM said Nighthawk r3 will be built to run 10,000 quantum operations across 520 programmable qubits, with 2,034 physical qubits. Clients and partners have run more than 34 billion circuits on Nighthawk r2, with reported speedups of up to 19x. In 2027 IBM plans to link two r3 processors into a 1,040-qubit system.

Why it mattersThe next processor is specified by operations and a plan to couple two chips.

IBM ↗
Funding

Torus raises $500 million for modular inference sites

Torus said on 11 November that Foundry packages power, storage, cooling, networking and compute into a prefabricated system of 1 to 20 megawatts, aimed at 12 to 16 weeks from contract to compute. A $500 million growth round was led by Magnetar, with ICONIQ, Origin Ventures, Ridgeline Partners and Trinity Capital. Each Foundry is paired with Torus Reserve, the company's storage and grid-management product.

Why it mattersThe offer is inference capacity on power the grid already has, not a new campus.

Torus ↗
Release

llama.cpp picks a smaller CUDA tile for thin BF16 layers

Build b12287, tagged 11 November, uses the MMVF CUDA path for BF16 matrices with few rows at batch sizes 2 to 8, where a 32-row MMF tile would leave most SMs idle. The notes give the example of Qwen4exp hyper-connection projections, 320 rows with K of 10240. MMVF is used only while a budget that scales with SM count is met, and only when K is at least 1024.

Why it mattersA thin projection on a large GPU can use a kernel that actually fills the chip.

GitHub ↗
Release

llama.cpp turns Vulkan KV dequant back on for coopmat2

Build b12291, tagged 11 November, enables KV-cache dequantization with Vulkan coopmat2, guarded by a heuristic on query tokens and the GQA ratio. The notes say some head sizes had been treated as unaligned, and fixing that alignment helped even when the dequant path does not apply.

Why it mattersA Vulkan build can use the faster cache path on head sizes it used to skip.

GitHub ↗
Chips

HOPE prunes experts that work as a group

Amazon's 9 November post on HOPE says mixture-of-experts models keep every expert in memory, and gives Qwen3.5-122B as an example: 256 experts in each of 48 layers, with 95 percent of its roughly 250 GB footprint in those expert layers, while a token uses 8 experts. HOPE scores experts in pairs rather than one by one. On Qwen3.5-35B-A3B with half the experts removed, the post says HOPE scored 42.3 on SWE-Bench Pro, matching the unpruned model.

Why it mattersDeleting experts can keep a coding score if the prune respects which experts work together.

Hugging Face ↗
Chips

AMD's 256-core Epyc SP7 servers start shipping

AMD said on 10 November that 6th Gen Epyc SP7 processors, on the Zen 6 design, are generally available, with up to 256 cores. Servers using them are shipping from major vendors. The same announcement extends the chips into Helion racks for agent workloads, and says some configurations reach 73,728 threads in one rack.

Why it mattersThe CPU that schedules an agent is on sale, not only the GPU that runs the model.

AMD ↗
Chips

Infineon says the data-centre build will keep growing

Infineon said on 10 November that revenue rose 12 percent in the year to the end of September, and that it expects sales growth of 20 percent this year on demand for AI data centres. Chief executive Jochen Hanebeck said US opposition will not stop the build, and that operators will also look for sites outside the United States. Infineon sells the power chips, not the AI processors.

Why it mattersThe power-chip supplier is planning for a faster build, even where new sites are unpopular.

Euronews ↗
Said in Chips & compute
Fractile is pushing the frontier of AI forward by rearchitecting the hardware stack.
Leandros Kalisperas
chief investment officer, British Business Bank · GOV.UK ↗

Policy & safety

Laws, lawsuits and safety incidents.

Safety

Helsing's Clarity keeps a person on every strike

Helsing unveiled Clarity on 11 November, an agentic intelligence platform it says is built in Europe and runs on infrastructure the customer owns. It processes sensor, satellite, drone, signals, open-source and report data, and nominates a target only when independent sources agree and an operator confirms. Every decision is recorded with the person who made it. Helsing did not say whether Clarity is already in service.

Why it mattersThe product is a military workflow whose last step is still a person's authorisation.

Helsing ↗
Policy

OpenAI cites a gap between students using AI and being shown how

In a 10 November note, OpenAI said ThinkYoung found 79 percent of surveyed students used AI for homework or assignments, while 47 percent said they had been shown how. The note describes education partnerships in Estonia, Slovakia, Greece and Italy, and names teacher preparation and local curricula as the work those partnerships share.

Why it mattersThe published gap is instruction, measured in one student survey.

OpenAI ↗
Research

A paper tests whether an agentic robot will refuse

LEASH introduces RoboRisk, with 92 simulated tasks and 13 tasks on a physical robot, in five groups: refusing a task, repairing a plan, hazards during execution, injection, and sequential memory. The authors say earlier robot-safety benchmarks cover at most two of those groups.

Why it mattersA robot agent can be scored on refusal and on what happens mid-motion.

Hugging Face ↗
Work

Freelance briefs are asking for AI across more of the job

OpenAI and Upwork said on 10 November that by late 2025 clients were requesting AI in more kinds of freelance work, including chatbot development, video production and marketing automation. Of 63 work categories, 43 lost share on Upwork where they were relatively common in ChatGPT. AI-related postings carried larger fixed-price budgets and more hours, and their mix of tasks became 45 percent more diverse from early 2023 to late 2025.

Why it mattersThe shift shows up in what clients post, before it shows up in employment totals.

OpenAI ↗
Policy

Australia will require safety rules for frontier models

The Prime Minister said on 10 November that mandatory safety and security rules for frontier AI companies will be written into Australian AI Standards law. A new regulator, working with the Australian AI Safety Institute, could impose significant penalties. Companies would have to manage serious risks, allow independent evaluation, report incidents quickly and publish a safety framework.

Why it mattersA lab that wants to operate in Australia would have a safety duty it can be fined for missing.

Prime Minister of Australia ↗
Policy

Australia weighs a criminal offence for model harm

ABC News reported on 11 November that the government will consider a criminal offence for an AI company whose model causes harm and that did not take reasonable steps to prevent it. A second option is criminal penalties for serious breaches of safety standards. The federal police told a parliamentary inquiry they had received no formal report of a crime over the Medicare statistics incident and had opened no investigation.

Why it mattersThe open question is whether a lab can be prosecuted, not only fined, for a model's conduct.

ABC News ↗
Policy

Pritzker asks Congress for a federal AI framework

Illinois governor JB Pritzker's agenda for the incoming Congress, reported on 9 November, asks for federal privacy law, federal rules for AI risk assessment, testing and monitoring, and energy and water standards for data centres. He said federal action should build on state laws, not override them. Illinois law he signed defines a catastrophic risk as one that could kill or seriously injure more than 50 people, or cause more than $1 million in property damage, in a single incident.

Why it mattersA large state is asking Washington to set a floor without wiping out the state rules already passed.

WTTW ↗
Policy

Grok's text watermark is off unless a team turns it on

Tesla North reported on 11 November that SpaceXAI's API docs, updated 10 November, now say text watermarking is off by default. Earlier docs had said supported Grok models would mark answers from 7 November, with an opt-out. Admins turn it on per model, and the change applies within five minutes. Voice output is still watermarked for the EU endpoint and for users in the EU, the EEA or California, and those audio marks cannot be switched off.

Why it mattersA developer who assumed every Grok reply was marked after 7 November needs to check the header.

Tesla North ↗
Policy

EU sandboxes must be free for small AI firms

Commission Implementing Regulation (EU) 2026/2487, adopted on 10 November, sets the rules for AI Act regulatory sandboxes. Each member state is to have at least one national sandbox by 2 August 2027. Guidance, advice and supervision are free for small firms, start-ups and public bodies, with exceptional costs billed only if disclosed first. Other companies may be charged the full cost. An exit report is not a declaration of conformity.

Why it mattersA small company can test a system with a regulator without paying the supervision fee a larger firm may owe.

EUR-Lex ↗

Robotics

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

Robot

Runway opens Praxis-1 weights for robots

Runway said on 11 November that Praxis-1, a 27 billion parameter mixture-of-experts world-action model, is available under Apache 2.0, including commercial use. On Nvidia's RoboLab-120 it reached 46.2 percent success across 1,200 episodes, against 42.9 percent for FLUX 3 Action, 36.8 percent for Cosmos3 Nano and 28.0 percent for pi0.5. Fine-tuned on Google's RT-1, which Runway said was absent from training, it reached 80.0 percent on SimplerEnv, against 52.4 percent for the original RT-1 policy.

Why it mattersA world-action model for arms and humanoids can be downloaded and adapted.

Runway ↗
Robot

Alpamayo 2 Edge plans a path inside the car

Nvidia's 10 November technical post describes Alpamayo 2 Edge, a 2 billion parameter vision-language-action model for DRIVE AGX Thor. Closed-loop reinforcement learning in AlpaGym improved driving performance by 48.9 percent over the best checkpoint without that training. After TensorRT optimisation, pixel-to-trajectory latency is 160 to 190 milliseconds depending on how much it reasons.

Why it mattersA driving model that explains itself is sized for the computer in the car.

Nvidia ↗
Research

A paper teaches a hand policy to foresee the grasp

DexForesight trains an action-conditioned spatial world model on scene depth, hand geometry and demonstrations, then freezes it and aligns a policy to those predictions. At deployment only the policy remains, on RGB, language and robot state. On 11 DexJoCo tasks the base policy's success rose from 49.4 percent to 60.5 percent, and on three real-robot tasks it reached 58.3 percent against 50.0 percent for the strongest baseline.

Why it mattersThe foresight is baked into the policy, so the robot does not have to render a future frame.

Hugging Face ↗
Research

A robot memory can mistake a pause for progress

The authors say memory-augmented vision-language-action policies that sample frames by time treat a pause or a wrist fidget as work. On RoboMME counting tasks, success fell from 47 to 16 of 50 on one task and from 44 to 10 on the other. Indexing memory by events, without retraining, recovered half or more of the loss.

Why it mattersA memory stamped by the clock can stop a task that has only been waiting.

Hugging Face ↗
Research

A paper adapts a long action chunk by how long it runs

ExRL shifts adaptation of a long-chunk robot policy from choosing the actions to choosing how long to execute each predicted chunk. The authors say that keeps the search small while the chunk itself stays expressive. The method is for online adaptation when real-robot samples are expensive.

Why it mattersA deployed policy can be tuned by shortening or lengthening a chunk, not by rewriting it.

Hugging Face ↗
Robot

AutoE2E opens a small model for hands-free driving

The Autoware Foundation and TIER IV released AutoE2E on 11 November, an open driving model with 79.9 million parameters under Apache 2.0. On a validation set its fine-tuned reactive layer, which plans 6.4 seconds of steering and acceleration ten times a second, had an average error of 0.744 metres over three seconds, second of eight open configurations. A world model and a reasoning layer are still being built, and further road trials are planned for 2027.

Why it mattersA hands-free driving model small enough for a car chip is available to retrain.

Autoware ↗
Robot

Atlas rolls on scooter wheels in a control test

The Robot Report said on 9 November that Boston Dynamics published a video of Atlas with scooter rims on its ankle actuators. The company called a wheeled Atlas a research testbed, not a product. One adapter is the only new part. The behaviours were learned in simulation, and the robot switches between rolling on flat ground and stepping over obstacles.

Why it mattersThe test is whether the whole-body controller still holds when the feet can spin.

The Robot Report ↗
Robot

Theseus says two store robots restock the shelves

Y Combinator published a launch post from Theseus Robotics, a Zurich start-up, saying two wheeled humanoids are already in stores. The company says it buys the robots and writes the software: restocking at up to 400 units an hour overnight, and order picking at up to 40 units an hour. It says pilots are lined up for up to 10 more robots in 2027.

Why it mattersThe claim is a night shift in a store that was not rebuilt for robots.

Y Combinator ↗
Robot

Hasbro and Dynamic Creatures plan character robots

Dynamic Creatures and Hasbro's Sixth Wall Studio said on 10 November that they will build robots that greet guests, share an adventure and remember them on a return visit at theme parks and resorts. Sixth Wall's CharacterOS is meant to keep a character's voice and behaviour inside set limits. The first uses are aimed at audiences aged 13 and older, and memory features would depend on the guest's permission.

Why it mattersA licensed character would have a body, with the studio still setting what it is allowed to say.

GlobeNewswire ↗

Science & health

AI in biology, medicine, climate and maths.

Health

AlphaGenome 2 predicts a variant inside one cell type

DeepMind announced AlphaGenome 2 on 11 November, a DNA model that predicts how a variant affects a specific cell type or state. It was trained on more than 100 million individual cell profiles across 68 tissues, and DeepMind said it identifies nearly twice as many variants that change gene expression as the previous model. It is available for non-commercial research through the AlphaGenome API, with a private preview for commercial research, and it is not validated for clinical use.

Why it mattersA variant's effect can be asked for one cell type, rather than a tissue average.

Google DeepMind ↗
Health

OpenAI takes caregiver workshops to five US cities

OpenAI said that on 10 November, with the Caregiver Action Network, the Alzheimer's Association and the American Diabetes Association, it is hosting free ChatGPT workshops in Oakland, Tampa, Pittsburgh, Atlanta and Detroit. An OpenAI analysis of 9 to 29 October estimated that more than 40 million weekly ChatGPT messages about healthcare in the US ask about care for someone else.

Why it mattersThe sessions are for people already using the product to look up someone else's care.

OpenAI ↗
Research

A world model proposes the next mutation in a protein

ProTEUS is a mutation-resolved world model pretrained on deep mutational scanning assays. Variants are states in a joint representation of sequence, evolution and structure, and mutations are the moves between them. The authors built it for closed-loop protein design, where only a small fraction of sequences can be tested in the lab.

Why it mattersThe next edit to a protein can be proposed from assays the model has already seen.

Hugging Face ↗
Health

ERIS2 predicts how a mutation shifts a protein

The University of Virginia said on 9 November that a team led by Nikolay Dokholyan released ERIS2, a model that predicts how a gene mutation changes the stability of a protein. It was trained on a database of more than 100,000 known mutations. The university said it showed strong predictive performance, and the tool is free at eris2.dokhlab.org with no login. It is described in the journal Structure.

Why it mattersA lab can screen which variants are worth a wet-lab test before it synthesises them.

UVA Health ↗
Health

GlioTrace follows tumour cells through living brain

Uppsala University said on 11 November that GlioTrace uses machine learning to follow glioblastoma cells in living mouse-brain slices. Cells from six patients were photographed at least every two hours for up to five days. Some tumours spread along blood vessels and others through the tissue, and the same cell could switch. The software is free to other researchers and is not a guide for treatment.

Why it mattersA drug's effect on invasion can be watched cell by cell, not only as a before-and-after image.

Uppsala University ↗
Health

An AI-designed enzyme takes a step in an HIV drug

The University of Manchester said on 11 November that, with Gates Foundation support, it engineered an enzyme for a step in making lenacapavir. Directed evolution of more than 12,000 variants reached more than 98 percent conversion, 90 percent isolated yield and 99.5 percent of the desired form. A generative model, RFD4-Proteina, yielded 67 designs for a harder step; four were active, and the best converted 79 percent of the starting material.

Why it mattersA manufacturing step for an HIV drug was proposed by a model and then checked in the lab.

University of Manchester ↗
Health

Claude designed binders for 14 of 15 protein targets

A 9 November Genetic Engineering News report says Anthropic's Claude models designed miniprotein binders for 15 targets and sent them to Adaptyv Bio and Twist Bioscience. Binders worked for 14 targets. Of 1,320 designs, 354 bound, a 27 percent hit rate, including several against TNFα. The tests measured binding only, not whether the designs work as drugs, and none was solved as a structure.

Why it mattersThe designs left the model and were made and tested, and one hard target still failed.

Genetic Engineering News ↗
Health

No AI-designed drug has an FDA approval yet

Chosun Biz reported on 10 November that McKinsey counted $8.4 billion of venture and private-equity investment in AI drug development in 2025, more than double about $4 billion two years earlier. It said no AI-designed drug has FDA approval. The furthest candidates it named are MindRank's MDR-001 and Insilico Medicine's rentosatib, which entered phase 3 in February and July respectively.

Why it mattersThe money has scaled, and the first approval is still ahead of the late-stage trials.

Chosun Biz ↗

Open source

Open-weight models and community releases.

Release

Unsloth posts Qwen3.8-27B weights for local chat

Unsloth 0.1.917-beta, tagged 11 November, adds Qwen3.8-27B Dynamic v3.1 GGUFs and says Claude Code and Codex can run on MLX models through unsloth start. The same notes say LTX-2.5 FP8 runs on 24 GB GPUs, Cosmos3-Nano can continue a clip from its first or last frame, and FLUX.2 and FLUX.3 use 36 to 43 percent less VRAM when decoding images.

Why it mattersA local app update ships a new Qwen quant and a lighter image decode together.

GitHub ↗
Release

llama.cpp reflects padding on WebGPU

Build b12292, tagged 11 November, adds PAD_REFLECT_1D on the WebGPU backend and drops a redundant contiguity check. The change is the one used by the Qwen3-TTS speaker encoder and the Chatterbox vocoder.

Why it mattersA browser build can run the reflection those speech models ask for.

GitHub ↗
Release

llama.cpp's SYCL clamp accepts a strided view

Build b12283, tagged 11 November, routes the SYCL clamp through the shared unary helper so it handles non-contiguous or strided tensor views, matching other unary ops. The old flat kernel and an unused block-size preset were removed.

Why it mattersA clamp on a sliced tensor no longer takes a different, stride-blind path.

GitHub ↗
Release

Ultralytics speeds compiled SAM image encoding

Ultralytics 8.4.211, tagged 11 November, says compiled SAM 3 and SAM 3.1 image encoders run 7 to 9 percent faster under torch.compile, with eager mode unchanged. The release also adds a static INT8 export for ExecuTorch that keeps detection, segmentation, pose and oriented-box heads in floating point. An OCR path can use a cached FP16 TensorRT engine, which it says is about 2.5 times faster than ONNX Runtime on dense pages after the engine is built.

Why it mattersA compiled segmentation pass and an on-device export both get a measured speedup.

GitHub ↗
Release

any-llm 1.50 can rerank through OpenRouter

any-llm 1.50.0, tagged 10 November, adds reranking through OpenRouter and lists every chat model that Hugging Face Inference Providers serve. A provider can say, per model, whether completions accept a response format. Listings can also report the context window the server is actually offering.

Why it mattersOne client can rerank and can see the context limit the provider is serving.

GitHub ↗
Release

Open WebUI 0.13.3 schedules a coding agent

Open WebUI 0.13.3, tagged 8 November, lets an automation pick a terminal, a working folder and a coding agent available there. A scheduled run uses that saved setup and checks that the terminal and agent are still available before it starts. The agent's result is sent to a chat.

Why it mattersA recurring coding job can be tied to one machine and one folder, and it stops if either is gone.

GitHub ↗
Release

llama.cpp adds an FP8 cache for keys and values

Build b12274, tagged 11 November, adds FP8 key-value cache support, including tests for scaled FP8 caches, and allows BF16 scales as well. The notes say Gemma 4 Assistant can run with FP8 by storing and reading the scale pointers.

Why it mattersA long chat can keep its cache in 8-bit numbers, including on a Gemma 4 assistant.

GitHub ↗
Release

Hermes 0.23.1 reads the live plugin catalog again

Hermes Agent 0.23.1, dated 9 November, raises the plugin-catalog cap from 2 MiB to 16 MiB. The project said the published catalog had grown past the old cap, so every install silently used the catalog shipped with the release and could not see plugins added since. Installs that stay on 0.23.0 or older keep using that fallback.

Why it mattersA plugin published after the last update can be installed again, instead of an older pin.

GitHub ↗

Cost saving tip

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

Gemini Decision 1 · tokens

Score the choice, and skip the output tokens

Google priced the Gemini Decisions API at $0.10 per million input tokens, with no output tokens billed. On the shared benchmark suite, Google said that is nearly 5 times cheaper than Gemini 3.5 Flash-Lite through the Interactions API.

Google ↗
Do this

Use it for yes/no, multiple choice and rubric scores, and keep a generative model for the open-ended reply.

Quick hits

Smaller stories worth a glance.

Deals & funding

Who raised money, and who bought whom.

2 deals

Funding rounds 2

Investors put money into a company in exchange for a share of it.

Benchmark backs Mercor

Raised
$750M
Series D

Mercor now valued at $20B

Mercor raised a Series D, including an employee tender, to expand evaluations and environments where agents practise professional work.

Other investors

FFelicisFFundomoAaApollo affiliatesGSG SquaredGeneral CatalystGBGlade Brook CapitalNvidiaSSalesforceSVSapphire VenturesSGStepStone GroupXVXTX VenturesZZoom

· Mercor ↗

JMI Equity backs Blue J

Raised
$176M
Series E

Blue J now valued at >$700M

Blue J raised a Series E to scale Workflows, its agentic tools for tax work, and to expand in the United Kingdom. The company said annual recurring revenue grew 165 percent in the 15 months after its Series D.

Other investors

OHOak HC/FTSVSapphire VenturesIGIntrepid Growth Partners

· Blue J ↗

On the calendar

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

39 things
MTWTFSSNov91011121314151617181920212223242526272829Dec3012345678910111213141516171819202122232425262728293031123
  1. 2–13NovProtoStar climate challengeCarnegie Mellon UniversityToday
  2. 9–20NovCOP31 climate conference in AntalyaAntalya, TurkeyToday
  3. 9–12NovCoRL meets in AustinAustinToday
  4. 12NovUpdated Claude usage policy takes effectAnthropicToday
  5. 15NovComments close on China's draft tech-export listChina MOFCOM3days away
  6. 15NovWorkBuddy and CodeBuddy free period endsTencent3days away
  7. 18NovWebinar on running a dock with StretchOnline · Online6days away
  8. 18NovNvidia invited to brief CongressUS House chairs6days away
  9. 18NovDots become the default in workspacesOnline · ChatGPT6days away
  10. 19NovKinference demo livestreamOnline · Online7days away
  11. 19NovClaude Data turns on for EnterpriseAnthropic7days away
  12. 20NovAI Film Award submissions close1 Billion Followers Summit8days away
  13. 20NovConductor cloud service shuts down8days away
  14. 20NovUpdated OpenAI usage policies take effect8days away
  15. 21NovInaugural edition in HollywoodAfrican AI Film Festival9days away
  16. 22Nov'The Challenge Sverige' originally scheduled (now delayed to 2027)Mastiff / TV4Rumoured10days away
  17. 25NovQwen4-Max preview endsAlibaba13days away
  18. 29NovDaybreak Blue sessions can be used for trainingOpenAI2weeks away
  19. 30–3Nov–DecContent London 2026C21Media3weeks away
  20. 30NovRose d’Or Awards3weeks away
  21. 30NovGemini 4 Carbon introductory price endsGoogle3weeks away
  22. 1DecHearing on Apple's bid to block testimonySuno and Apple3weeks away
  23. 2–3DecWorkshop on autonomous surgical robotsFDA3weeks away
  24. 4DecAI-made film opens in limited releaseGods Don’t Give Gifts3weeks away
  25. 4DecGods Don’t Give Gifts premiereShowrunner (Fable)Expected3weeks away
  26. 4DecSingapore AI Safety Fellowship ends3weeks away
  27. 31DecHy4 half-price offer endsTencent7weeks away
  28. 31DecJules shuts downGoogle7weeks away
  29. 1JanFrontier labs must register under the RAISE ActNew York7weeks away
  30. Later in NovGuided Learning pilot in US schoolsGoogleExpected
  31. Mid-NovXREAL Aura shipments beginExpected
  32. DecemberNothing's first AI-native productExpected
  33. 2027Two Nighthawk r3 processors linkedIBMExpected
  34. H1 2027Digit 5 early customer accessAgility RoboticsExpected
  35. Nov–Dec 2026Fort Benning robotaxi pilotFort Benning
  36. Later this yearAge checks reach ThreadsMetaExpected
  37. 2028R200 systems ready to deploySouth KoreaExpected
  38. through Feb 2028Language Explorer in Real TalkBerlin
  39. 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.