New models
Google: Gemini Decision 1 is here, the first model built specifically for decisions. The model, which does not generate text token by token, answers closed questions (yes/no, multiple choice, scoring) and gives a probability and a confidence score for every option. In a single pass it can also evaluate as many as a thousand questions about the same text, image, audio, video or PDF, in 150–250 milliseconds. It is available through the new Gemini Decisions API in Google AI Studio, at $0.10 per million input tokens, with no charge for output; according to Google it is about five times cheaper than Gemini 3.5 Flash-Lite. Typical uses: which model or tool should get a request, whether a piece of content meets the rules, or how to score a submission. Inside Google it is already used for model selection in Gemini Notebook and for safety checks in Search. Source: Google, The Decoder.
Google DeepMind: AlphaGenome 2 also shows in which cell type a DNA difference causes trouble. Almost every cell in our body carries the same DNA, yet each switches on different genes, so the same genetic difference can do something quite different in heart muscle than in an immune cell. The new model was trained on more than a hundred million single-cell profiles from 68 tissues, and according to DeepMind it finds nearly twice as many variants that affect gene function as the best previous model. It can also give estimates for new cell types and states without retraining. It can help researchers find the cellular causes of diseases and drug targets; it is available for non-commercial research. Source: Google DeepMind, The Verge.
Alibaba: a preview of Qwen4-Max, the flagship of the new Qwen4 family, is available to try. In our previous day's roundup, Qwen4's arrival was still only a rumor; now the preview of the largest variant is available in Qwen Chat (including agent mode), in the Alibaba Token Plan subscription and in the Qoder coding tool. The company promises deeper knowledge, a better grasp of the world and stronger coding and agent skills, especially for complex programming and office work. There is still no official release date for the family's smaller members. Source: Alibaba Qwen (X).
Runway: the Praxis-1 robot-control model is now freely downloadable with open weights. The 27-billion-parameter "world action model" unveiled in September was trained on videos, and at once predicts what will happen in the image and what the robot should do. According to the company it leads Nvidia's RoboLab-120 ranking with a 46.2 percent success rate, ahead of every open-weight vision-language-action (VLA) model. It is under the Apache 2.0 license, so it can also be used freely for business purposes, and can be run and further trained on your own hardware; it is available on GitHub and Hugging Face. Source: Runway, Runway (X).
Tools & features
SpaceXSI: Grok Bot 1.0 is out, the first stable version of the AI colleague tuned for teamwork. In the 1.0 release, which had earlier been flagged as a rumor, the Team Bots feature lets a whole team work with the same bot, which learns who is working on what and can also be shared in Slack or Microsoft Teams. The bots follow multi-person conversations, can join video calls (they speak up through the chat), and also flag without being asked when something in a project needs attention, for example when a deadline is approaching. Source: SpaceXSI, SpaceXSI (X).
Cursor: teammates' AI agents can now work together. The code editor's Projects feature has been rebuilt for multi-user use: every member of the team and their agents see each other's work in a single shared workspace, and can coordinate the tasks. Source: Cursor (X).
ElevenLabs: voice agents are getting memory, routines and models that can be tailored to a company. The Memory feature announced at the New York ElevenLabs Summit lets a customer-service agent remember what a returning customer said earlier (for example that they are only free in the afternoon, or that they are moving in March), and Routines handle the work before and after the conversation, for example calling a loan applicant back or explaining a bill by SMS. The Interaction Engine chooses the right model for every step of the conversation: a cheaper one for simple questions, a frontier model for more complex ones (for example a refund). In the coming months companies will be able to fine-tune the speech and speech-recognition models on their own data; the company also unveiled Scribe v3 Medical, a transcription model that knows healthcare terms, and as a preview, ElevenAI, a workplace assistant that dictates, takes notes and helps in meetings. Source: ElevenLabs.
Google: Antigravity's agent can drive a browser, and developers can give Gemini API agents their own runtime. The Antigravity development environment's agent can now research in the built-in browser and check the written code live; the user decides what it may sign into, and downloads pause until they are approved. Alongside the Gemini API's Managed Agents, developers can also attach their own, self-hosted sandbox (an isolated runtime). Source: Google Antigravity (X), Google AI Studio (X).
AWS: a free local emulator for more than two hundred services, with the code written entirely by AI agents. AWS Emulator simulates Amazon's cloud services on your own machine, so developers can test their programs without a paid account. According to Amazon, a swarm of agents wrote the Rust code, with nearly thirty thousand changes in the first three weeks, and the team did not write or review a single line. For now it runs with Docker on macOS and Windows. Source: GitHub, Simon Willison.
Google Labs: Tabletop builds a shared travel plan, a vote or a bill-splitter from a single prompt. In the experimental service the group drops in its links, files and ideas, and Gemini turns them into an interactive itinerary, a dinner vote, a bill-splitter or a leaderboard that everyone can then edit at once, like a Google document. For now it is available only in the United States, through a waitlist. Source: Google, The Verge.
Smaller updates at a glance. The Gemini app has also appeared as a standalone app in Apple CarPlay, so a conversation started on the phone can be continued by voice while driving. ChatGPT's Finance feature has also arrived in the desktop app, and also builds interactive reports from your own financial data; for now only in the United States. OpenRouter's Extractions API pulls specified data (for example a receipt total) from text, images and PDFs in structured form. ByteDance's Seedance 2.5 video model has also become available in 4K in Runway and Pika. Source: Gemini (X), ChatGPT (X), OpenRouter (X), Runway (X).
Policy & safety
Chip exports: the White House and Congress would both tighten the rules after the Nvidia scandal. According to The Information, the White House has asked Congress to require the Commerce Department to write a rule that would stop Chinese companies from using frontier AI chips through remote access in data centers abroad. Citing procurement documents, Reuters wrote that Northwestern Polytechnical University, which is linked to the Chinese military and is on the US blacklist, still got access to Nvidia H200 chips. Several US lawmakers are calling for Jensen Huang to resign, and are threatening a subpoena if he does not appear voluntarily at a closed session. Source: The Information (X), Reuters, CNBC.
England: the higher-education regulator says AI has made it almost impossible to assess coursework fairly. According to an Office for Students (OfS) investigation, students have AI write essays and dissertations, often unnoticed, and get similar grades to those who do not use it. The regulator is urging a return to traditional exams and in-person presentations, and has held out sanctions for universities that do not put in place reliable, AI-free assessment. Source: The Guardian, The Decoder.
US data companies are selling training environments worth half a billion dollars to Chinese AI labs. A ChinaTalk investigation identified at least nine US firms, including Mercor, Surge AI, AfterQuery and Turing, that also sell so-called reinforcement-learning (RL) environments: simulated tasks on which models practice, for example, coding or office work. Chinese labs deliberately buy the same products previously used by Anthropic or OpenAI, and can thereby skip expensive experimentation and computing capacity. Source: The Decoder.
Perplexity: they let twelve AI models loose on their own agent sandbox, to see whether they could break out. The company tested the SPACE virtual-machine environment built for its Computer agent: the models had to obtain a password from the host machine and bypass network blocks. According to CEO Aravind Srinivas, one lesson was that models not specialized in cybersecurity proved better attackers than the cybersecurity models, and that in tests the model has to be told explicitly to break out. Source: Perplexity (X), Aravind Srinivas (X).
Radical Numerics: an AI model that also runs on a laptop, with no internet, spots natural and engineered pathogens. The Omnii Defense model can also be used in the field together with a USB gene sequencer, so researchers and public-health staff can quickly recognize new biological threats. The company's founders previously worked on the open Evo and Evo-2 DNA models; they say an arms race is starting in biological defense too, as in cybersecurity. Source: Wired.
Helsing: the German defense firm's Clarity platform uses AI agents to prepare military targets. In the system unveiled in Berlin, hundreds of specialized agents process sensor data, open-source intelligence and reports, identify enemy forces and facilities, and then prepare the data for weapons systems for human approval. According to Helsing, a person's name will be attached to every decision, and the system will run on the customer's own infrastructure as a European product; it is a rival to similar platforms from Palantir and Anduril. Source: The Decoder, Resilience Media.
Business & society
Nvidia: according to The Wall Street Journal, the board is preparing to oust Jensen Huang. Citing five employees, The New York Times wrote that the CEO also had an affair with a subordinate, Kimberly Powell, the vice president for healthcare; after that, the WSJ reported that the board is meeting as soon as Wednesday about Huang's future, and is preparing to remove him. Nvidia did not comment on the reports. The stock still rose nearly 3 percent during the day, but fell about 8 percent in after-hours trading. Huang has led the company he founded for 33 years, and it has no known succession plan; the quarterly earnings report is due on November 17. Source: The New York Times, The Wall Street Journal, TechCrunch, CNBC.
Anthropic: the IPO has been oversubscribed several times over, and it emerged that quarterly revenue approached $17 billion. According to Bloomberg, investors are often asking for many more shares than are available, and according to the Financial Times the banks are therefore considering raising the price range, which would take the company's value above $1.5 trillion. The retail offering has opened in Australia. According to The Information, Anthropic's third-quarter revenue rose 47 percent, to nearly $17 billion, and its operating loss was only $567 million, versus OpenAI's $10 billion in revenue and $10.9 billion loss. Source: Bloomberg, Financial Times, CNBC, The Information (X).
TSMC: the Taiwanese chipmaker is investing $1 billion in OpenAI and Anthropic. The board approved buying up to $500 million of OpenAI preferred shares, and another $500 million of Anthropic shares at the IPO price. Analysts say this kind of circular financing, in which suppliers invest in their customers, makes it harder to judge real AI demand. Source: Ars Technica.
Anthropic: its own AI model ported Claude to AMD chips in three days. According to the investor presentation, bringing up a new chip platform used to occupy dozens of engineers for months, but the Mythos model now did the work with almost no human intervention. According to co-founder Tom Brown, the company is already training and running Claude on all four major chip families, which could weaken the previous advantage of Nvidia's CUDA software environment. Source: The Decoder.
United Airlines: the airline is blocking personal AI agents from booking tickets on passengers' behalf. The agents can still show the flights, but they cannot complete the purchase. The company cites privacy risks and federal rules, for example that the passenger must personally certify there are no explosives in their luggage. CEO Scott Kirby says they do not want another middleman. The decision affects assistants such as Meta's Muse and OpenAI's Dots; after eBay, another large company is shutting out third-party shopping agents. Source: The Verge, The Wall Street Journal.
ByteDance: according to Bloomberg, TikTok's parent will launch its own Muse-like personal AI assistant in a matter of weeks. According to the paper, co-founder Zhang Yiming told a closed internal circle that by 2029–2030 they want to turn the company into a frontier lab that can compete with OpenAI and Anthropic. Source: Bloomberg.
Data centers: 181 gigawatts of on-site gas plants are being planned next to AI data centers in the US and Europe. According to a study, developers would build local fossil-fuel plants because they are slow to get a connection to the national grid. Climate advocates say these plants are unlikely to be temporary, and could lock in high emissions for years. Source: The Guardian.
Big Walk: a program bug fixed with AI has set off a dispute among independent game developers. The House House studio later acknowledged that, at the end of development of their popular co-op game, they used AI to a minimal extent to fix a bug. Some Independent Games Festival jurors say even that is too much, and the dispute is about how and when players should be told about the use of AI. Source: Ars Technica.
Investments at a glance. Britain's Fractile, which designs chips tailored to running AI models, raised $720 million (according to the UK government at a valuation of more than £5 billion, with the British public AI fund taking part). Energy-storage firm Torus raised $500 million for container-sized AI data centers, Rox, which automates sales research and outreach with agents, raised $100 million at a $1.3 billion valuation, and Edelbyte, which is building a data-analytics platform optimized for AI agents, came out of stealth with $70 million. Source: TechCrunch, Bloomberg, Fortune, Forbes.
Rumors
A new Google video model, BouncyBohr, has turned up in tests. According to a well-known leaker, the model, which handles 4K resolution and multiple reference images, can extend a video three times by ten seconds; based on a first test, its skill sits among Google's existing fast video models. The same person also places the as-yet-untried Gemini 4 Helix (Flash-Lite) and Gemini 3.9 Live among Google's next announcements. Google has not confirmed any of them. Reliability: low–medium (a single source whose earlier signals have partly been borne out, but the models have not been officially named). Source: chetaslua (X), chetaslua: Google models (X).
What this means for you: the team-ready Grok Bot, Cursor's shared agent workspace and ElevenLabs' customer-service agents that remember and also handle follow-up work show that the AI agent is increasingly becoming a colleague, so it is worth thinking now, on our AI agents page, about which recurring customer and internal processes you would entrust to AI agents, and with an AI governance framework, you can set safe and acceptable limits in time, as the United and eBay bans and the English university crackdown already show that partners and authorities are increasingly regulating where and how AI may work.