8–14 November 2026
The week's stories that will still matter next week, with the noise taken out.
Anthropic opens IPO roadshow
Anthropic began its roadshow to sell up to 825 million Class A shares, targeting a Nasdaq listing under ANT.
Agents & codingAgents learn to pay
OpenAI launched ChatGPT Wallet so a ‘dot’ agent can check out after a user approves the purchase.
Video generationDiffusers 0.43 adds video with sound and speech-driven motion
The release adds pipelines for Pika H3 Flex, which generates video and stereo audio, and for Kandinsky S2V 1.0, which animates a person from a reference image and a speech recording.
Creator toolsAdobe opens its apps to other chatbots
At MAX, Adobe shipped a local MCP server so ChatGPT, Claude or Gemini can call Creative Cloud apps.
Story of the week
The single most consequential story of the week.
- Colette Kress role
- Interim CEO
- Harvey C. Jones on board
- Since 1993
Jensen Huang resigns as Nvidia's chief executive
Nvidia said Jensen Huang resigned as president, CEO and a director, effective immediately. CFO Colette Kress becomes interim CEO while remaining CFO, and long-time director Harvey C. Jones becomes the board’s first chair.
Why it's the story of the weekThe leader of the dominant AI chip supplier has stepped down, with the board signaling the roadmap must keep shipping during the transition.
Models & labs
New models, the labs behind them, and the rules they answer to.
Anthropic opens the roadshow for its IPO
Anthropic began its IPO roadshow to sell up to 825,000,000 Class A shares at an expected $70–$78 each. It has applied to list on the Nasdaq Global Select Market under the symbol ANT, and the registration statement is not yet effective.
Why it mattersA frontier lab is asking public investors to price its growth.
Anthropic ↗Gemini 4 Carbon opens to paying users
Google made Gemini 4 Carbon available to paying developers and businesses, and it is rolling out to AI Ultra subscribers. Through 30 November it matches Argon’s introductory price at $2 per million input tokens and $10 per million output tokens.
Why it mattersA model previously limited to cyber defenders is now a public API at a low introductory price.
Google ↗Meituan posts LongCat-2.5 weights
LongCat-2.5 is a multimodal mixture-of-experts model with 1.6T parameters and ~48B active per token. The MIT-licensed weights include a W4A4 route that quantizes expert weights and activations to 4 bits.
Why it mattersA trillion-parameter open-weight model is available for commercial use.
Hugging Face ↗Agents & coding
AI that acts on its own, and the tools developers build with.
ChatGPT Wallet lets a ‘dot’ agent check out after you approve
OpenAI launched ChatGPT Wallet for a limited set of US Pro subscribers, 18 and older, with a US-issued Visa, Mastercard or AmEx. An agent can help choose and then check out; the user reviews the cart and approves with Face ID or a fingerprint, and payment uses a tokenized card.
Why it mattersAgents can handle purchases with explicit human approval.
TechCrunch ↗Companies can host Claude's agent service themselves
Anthropic now lets customers run Claude Managed Agents in their own cloud or data centre. Agents, session histories and credentials stay on those systems, while model calls go to a chosen provider.
Why it mattersEnterprises can keep agent state inside their network.
The Next Web ↗Automattic's Spacefast hosts the app an agent built
Spacefast adds auth, DB, files, email and analytics to an app made with ChatGPT, Claude, Cursor or a git repo. It soft-launched in September and opened on Monday.
Why it mattersThere’s now a ready host for agent-generated apps.
TechCrunch ↗Video generation
The models and tools that make video.
Diffusers 0.43 adds video with sound and speech-driven motion
The release adds pipelines for Pika H3 Flex, which generates video and stereo audio, and for Kandinsky S2V 1.0, which animates a person from a reference image and a speech recording. It also wires MiniMax-H3-Image and PAN 2’s simulator, plus LoRA loading and checkpoint saving.
Why it mattersMultiple new video models now land in the library many apps already use.
GitHub ↗SwitchX 3.0 relights a shot and keeps the actor
Beeble’s SwitchX 3.0 is a video-to-video model rebuilt on MiniMax H3. It accepts up to four references, moves the background and locks the actor inside the alpha, outputting 4K 10-bit video.
Why it mattersEditors can change location while preserving the subject’s pixels.
Beeble ↗World Labs opens the Atlas world model to everyone
Atlas now offers public access with tools like Viewfinder, Chisel and Relight, and an API that returns SPZ Gaussian splats. A refined reconstruction costs 5,000 credits (about $5) and typically finishes in 15–25 minutes.
Why it mattersA previously gated world model now has public pricing and access.
Radiance Fields ↗Creator tools
The apps and platforms people make video with.
Adobe opens its apps to other chatbots
At MAX, Adobe shipped a local MCP server so ChatGPT, Claude or Gemini can call Creative Cloud apps. Photoshop gains relighting and style sliders, and Premiere gains object removal and a new colour grade.
Why it mattersChat interfaces can now drive Adobe editors directly.
TechCrunch ↗Claude Sheets and Whiteboards open in beta
Anthropic launched Sheets and Whiteboards in beta on paid plans and made Claude Data generally available. Sheets can build tabs, formulas and drop-downs and export to .xlsx or CSV; on Enterprise, Data turns on by default on 19 November.
Why it mattersSpreadsheets and data work can stay inside the chat.
Anthropic ↗Google Maps drops into an app inside AI Studio
Developers can now call Maps Platform from AI Studio without setting up billing first. Prompts can load live maps, and initial usage needs no card.
Why it mattersLocation-aware agents can be prototyped in the same window as the model.
Google ↗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.
Disney and Adobe will train models on Disney's worlds
Adobe became Disney’s official AI innovation partner and will build Firefly Foundry models tuned on Disney characters, places and styles. Imagineers are testing sketch-to-image for theme-park art.
Why it mattersA major studio is building private generators on its IP.
Adobe ↗Elvis and Eminem rights holders sue Suno
Presley’s rights holders sued under a Tennessee voice law, and Eminem’s publisher sued over songs. One suit says damages could exceed $50 million; the other seeks about $41 million.
Why it mattersA music generator faces tests under voice-likeness and song rights.
Los Angeles Times ↗Chips & compute
Chips, data centres and the power behind them.
Jensen Huang resigns as Nvidia's chief executive
Nvidia said Jensen Huang resigned as president, CEO and a director, effective immediately. CFO Colette Kress becomes interim CEO and remains CFO, and Harvey C. Jones becomes the board’s first chair.
Why it mattersLeadership change at the dominant AI chip supplier lands in the middle of an industry build-out.
NVIDIA ↗Google adds a locked-down TPU and a denser one
Google introduced TPU 8tx, which protects data and weights while in use, and TPU 8i Max, two 8i chips in one package. It is the first year Google has released four TPU systems.
Why it mattersCompute is splitting into secure enclaves and higher-density parts for AI workloads.
Google ↗Anthropic is buying $3.4 billion of gas turbines
Anthropic will buy Caterpillar generators OnePWR had ordered, enough for 1.7 GW, according to the Financial Times. Anthropic did not buy the supplier itself.
Why it mattersAI labs are securing power, not just GPUs.
TechCrunch ↗Policy & safety
Laws, lawsuits and safety incidents.
New York opens registration for frontier labs
New York launched a portal for developers with more than $500 million in revenue to register by 1 January 2027. They must publish a safety framework and report incidents within 72 hours; first penalties can reach $1 million.
Why it mattersLarge labs now have a filing date and penalties in one US state.
TechCrunch ↗Australia will require safety rules for frontier models
Australia said mandatory safety and security rules for frontier AI will be written into standards law, with a new regulator working with the Australian AI Safety Institute. Companies would need to manage serious risks, allow independent evaluation, report incidents quickly and publish a safety framework.
Why it mattersOperating in Australia will come with enforceable safety duties.
Prime Minister of Australia ↗A Fireworks breach reached Notion AI chat transcripts
Fireworks AI said an attacker accessed a development system and took customer billing details and some datasets. Notion said about 1% of customers had AI Insights chat transcripts exposed and it disconnected Fireworks on 3 November.
Why it mattersModel hosts can sit in the path of sensitive customer chats.
TechCrunch ↗Robotics
Robots, self-driving and AI in the physical world.
Runway opens Praxis-1 weights for robots
Praxis-1, a 27B MoE world-action model, is available under Apache 2.0. It scored 46.2% on RoboLab-120 and reached 80.0% on SimplerEnv when fine-tuned on RT-1.
Why it mattersAn open world-action model for arms and humanoids can be adapted for real tasks.
Runway ↗Tesla's robotaxis open in San Antonio
Tesla said on X that robotaxi service would be available in San Antonio from 5 a.m. local time on Monday. It did not say whether cars there have safety monitors.
Why it mattersAnother city gets driverless rides amid ongoing permitting questions.
TechCrunch ↗Science & health
AI in biology, medicine, climate and maths.
AlphaProof Code checks Signal's own source
DeepMind said AlphaProof Code, using Gemini Carbon, proved key properties of Signal’s production code in weeks, generating over six million lines of Lean.
Why it mattersMachine-checked proofs are reaching real-world cryptographic code.
Google DeepMind ↗Google publishes a yardstick for health answers
Google shared a framework for judging whether health AI is safe and useful for both chat use and clinical settings. It is an evaluation method, not a diagnostic product.
Why it mattersHealth answers need repeatable tests, not just confident prose.
Google ↗Claude designed three gadgets that worked on the first build
Anthropic said Fable 5.2 laid out a speakerphone, a voice-assistant hub and a handheld virtual pet with partner Diode; all three prototypes were built as designed without circuit fixes.
Why it mattersElectronics design joins the list of tasks models can draft for fabrication.
Anthropic ↗Open source
Open-weight models and community releases.
llama.cpp 1.0.0 freezes a public C API
The 1.0 release removes deprecated APIs and promotes the NextN speculative API into llama.h, adds new text architectures and an OpenAI-style /v1/audio/speech endpoint, and updates ggml to 1.0. From here, breaking C-header changes will require a major version.
Why it mattersBindings can rely on a stable API line for local inference.
GitHub ↗Buoyant ships Kinference for vLLM on Kubernetes
Kinference is an Apache-2.0 control plane in front of vLLM that rate-limits tokens, tracks cost by tenant and routes with the KV cache in mind. On a shared-prefix test it cut median time-to-first-token by 95% versus plain Kubernetes balancing.
Why it mattersOpen-weight serving gains the missing multi-tenant control layer.
Buoyant ↗Hugging Face shows how to serve a decision model
A guide introduces decision models that return probabilities over preset options and adds a Transformers decision pipeline. It covers serving with llama.cpp and training with HF tooling.
Why it mattersThere’s now a standard path to deploy models that score choices, not text.
Hugging Face ↗Paper of the week
One research result, explained in plain words.
ArchSpace: Recording, Reviewing, and Referencing LLM Architecture Research
Jie Wang, Tong Zhu, JiaHao Kuang, Zhe Xu, Yunkai Chai, Dehong Zheng, Yu Cheng, Tao Ji, ZhouHan Lin, Qi Zhang, Bowen Zhou
- What they tried
- They created a public platform where architecture ideas are logged with experiments and reasoning, reviewed by the community, and projected to larger scales using scaling laws.
- What they found
- Using shared baselines (Qwen3.5 and OLMo3), fitted scaling laws and a reusable recipe, the platform estimates loss at larger runs from small experiments and turns validated ideas into new baselines.
- Why it matters
- Architecture choices often lack transparent records; this gives researchers a common place to test, review and scale promising ideas before expensive training.
Money this week
The week's deals at a glance. The monthly review has the full picture.
Next week
What's coming from 15 November to 21 November, and how sure we are about each.
- 15NovComments close on China's draft tech-export listTomorrow
- 17NovThird-quarter results and earnings call3days away
- 18NovDots become the default in workspaces4days away
- 18NovNvidia invited to brief Congress4days away
- 25NovQwen4-Max preview ends11days away
- 30–3Nov–DecContent London 20262weeks away
Stories are summarized in plain language; follow the link before quoting a figure. Story images are shown from the publishers' own sites.