AI is moving into the physical and scientific world. Factory robots are learning by watching people, a model builds playable 3D worlds, and Anthropic is running its own drug programmes. Google's new security-focused model took first place on an independent intelligence ranking, but only selected cyber defenders can test it. Behind the data centre build-out, much of the money is borrowed.

IN THIS ISSUE · 5 SECTIONS

 1  Major AI News: Gemini 4 Carbon, Genie 4's playable worlds, Anthropic's drug programmes.

 2  Fun AI Topics: factory robots that learn by watching, an AI swarm checks a giant proof, luxury brands sell human effort.

 3  AI Tools: Liquid AI LFM3, GitHub's Copilot app, Nemotron 3.5 Transcribe.

 4  Expert Prompt: plan your user research before a product decision.

 5  Trending: New York's rules for chatbots and children, AI's debt problem, publishers against Meta.

LEAD STORY

Google unveiled Gemini 4 Carbon on 28 October, two weeks after Gemini 4 Argon. It ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 later. For now it is available to a subset of cyber defenders, and Google will add more trusted partners through its Fairwind Program, so the ranking is hard to check independently.

Why it matters: Businesses can't use the top-ranked model yet. Google says broader access is coming "soon" but hasn't given a date.

ALSO TODAY

Google DeepMind introduced Genie 4, a world model that turns text prompts into explorable, photorealistic scenes in real time. Users navigate with keyboard controls and change scenes through plain English, while the model adds ambient audio, music, and talking characters. Sessions can last up to two minutes, twice the previous version, and Gemini models plan multi-step events to keep scenes coherent. In one demo, a simulated humanoid robot played checkers while a Gemini model planned its moves.

Why it matters: Safe simulation could let robots and agents practise unfamiliar tasks before acting in the real world, cutting cost and risk.

ALSO TODAY

Anthropic revealed it is running its own drug programmes and has acquired Mithril Medicine, a one-person company developing treatments for two rare forms of muscular dystrophy. Mithril took a possible treatment for one form from an idea to promising results in mice in under 18 months. In partner work, Claude redesigned one section of a CAR T-cell therapy's genetic instructions, producing a design about 70% shorter that let immune cells make roughly ten times as many receptors, and wrote first drafts of two regulatory documents, one in four hours instead of about three months by hand. If the candidates hold up, Anthropic expects one to be ready for human trials as soon as late 2028. The work is still early and mostly preclinical.

Why it matters: An AI lab doing its own hands-on science signals that models are moving from writing text to running real research pipelines.

Standard Bots lets people teach its robots by demonstration instead of rigid programming. A person can teleoperate the robot, use gloves, or guide a handheld device. The company is deployed across hundreds of companies and roughly 350 use cases. Production-ready doesn't have to mean perfect. If a person only steps in once an hour instead of standing at a station all day, the return can still be large, though a mistake that ruins an expensive part changes the maths.

Why it is interesting: It shows robotics shifting from specialist-only programming to tools existing factory staff can operate.

Source: The Neuron

Startup Axiom says it directed a swarm of agents to produce a fully machine-checked version of the Classification of Finite Simple Groups in 24 days. The original proof spans roughly 15,000 pages across more than 500 papers, with no single expert believed to have read it all. The agents generated about 43 million lines of Lean code and found and repaired a false claim in one stage, leaving the theorem intact.

Why it is interesting: It suggests AI can help verify sprawling human proofs that no individual can fully hold in their head.

Source: i10x

As feeds fill with AI-generated content, luxury brands are showing visible human craft. Prada's holiday campaign, Drawing Prada: Animated by Hand, uses traditional hand-drawn animation with no AI, led by Disney veteran John Pomeroy, who joined Walt Disney Animation Studios in 1973. The tactic, called effort signalling, reveals the work behind a product to raise its perceived value.

Why it is interesting: It marks a deliberate split, with premium brands using scarcity of human labour as the selling point AI cannot fake.

Source: Laughing Place

[MODELS] Liquid AI LFM3 - a family of small multimodal models built to run on your own device within tight memory limits. liquid.ai (company site)

[CODING] GitHub Copilot app - a desktop app for running several AI coding agents at once, each in its own copy of your project. It sits alongside VS Code, the command line and GitHub.com, and is in technical preview for paid Copilot plans. helpnetsecurity.com (launch coverage)

[VOICE] Nemotron 3.5 Transcribe - an NVIDIA speech model that transcribes recordings of up to an hour in 40 languages, tells speakers apart, translates to and from English, and answers questions about the audio. huggingface.co (model page)

Use it to: plan your research before you commit to a product decision, so you test something real instead of guessing.

Act as a senior UX researcher and product strategist. I will give you context: the product, feature, audience, business goal, hypothesis, key unknowns, development stage, timeline, resources, existing evidence, and constraints. Recommend and justify an appropriate mix of interviews, usability tests, surveys, diary studies, prototype tests, and behavioural analysis. Then define objectives, participant criteria, sample size, recruiting approach, session formats, exact questions or tasks, metrics, and the evidence that would support or challenge the hypothesis. Keep questions neutral and separate exploratory research from validation research. Deliver a week-by-week plan, a discussion guide, a synthesis framework, key research risks, and clear Proceed, Iterate, Reconsider, or Test Further decision criteria.

Example use case: A product team planning user research for a new onboarding flow uses it to decide what to test, in what order, and what result would justify building versus rethinking the feature.

A new law bars chatbots available to minors from pretending to be human, claiming personal relationships, asking unprompted emotional questions, reusing private information, or providing sexually explicit content. The restrictions apply to all users unless a company verifies they are adults, and penalties can reach $25,000 per violation. Critics say the definition of chatbot is so broad it could block useful tools.

Why it is important: It is an early attempt to regulate emotionally manipulative AI, and the broad wording may force companies to add age checks widely.

Source: Governor Hochul's office

Much of the AI build-out is now paid for with debt, and investors are starting to push back. A CoreWeave unsecured bond fell below 80 cents on the dollar, dragging its yield towards 15%. Big players are financing sites through outside vehicles, such as the Blue Owl-backed vehicle behind Meta's Hyperion project, and about $18 billion of Oracle loans tied to a New Mexico site have become stressed after power-permit delays. Gartner expects more than $4 trillion of asset value to be exposed to writedown risk through 2032.

Why it is important: If borrowing costs rise, debt-heavy builders could struggle while cash-rich giants keep spending, reshaping who dominates AI compute.

Source: Enterprise AM

Meta has delayed new business terms from 30 October to 20 November after publishers warned that news content linked on Facebook and Instagram could be used to train AI. An Oslo court has ruled the legality of the terms must be reviewed before they can take effect, the UK competition watchdog says it will keep scrutinising them, and Danish and Swedish publisher groups have complained to their national competition authorities. Meta removed explicit AI-training language, but publishers say the revised wording still covers its generative AI.

Why it is important: The dispute tests whether platforms can claim content for AI by default, and could set the rules for how publishers are paid.

Source: Press Gazette

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas