A clearer view of AI.
THE INDEPENDENT AI.

AI NEWS FOR BUILDERS · 19 STORIES

AI news, 28 October 2026

Google DeepMind's Genie 4: a video you can walk into; Anthropic moves into drug discovery, and buys a one-person biotech run with Claude; Gemini API CLI: Google's media models in one terminal command.

Top stories

Google DeepMind's Genie 4: a video you can walk into

Genie 4 turns a text or image prompt into an interactive world you can move through and change with plain-language instructions in real time, now with native audio and longer sessions of a few minutes (TechCrunch reports up to two minutes, against one for Genie 3). It's built with Gemini Omni and Gemini 3.8 Flash, and Google also showed a robotics experiment with Gemini 4 Argon. It's a research preview for trusted testers for now, coming to Google AI Ultra subscribers in Project Genie and to early-access partners in the Gemini API and AI Studio, and Google lists its own limitations. The 'looks like real camera footage' reactions are creator enthusiasm, not a test.

Why it mattersEarly uses are training scenarios, product walk-throughs and pitch demos you can explore instead of just watch. Most people can't try it yet, so note it and wait for API access before planning anything.

Sourcesx.com post ↗ · deepmind.google ↗ · x.com post ↗

Anthropic moves into drug discovery, and buys a one-person biotech run with Claude

Anthropic published early results from six pharma and biotech partners (BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group). Examples: Divergence produced about 35,000 antibody designs in roughly a month, largely autonomously, and Enveda says Claude caught all 11 issues its reviewers found in 1,061 pages of manufacturing documents plus 7 they missed, cutting review time by 89%. These are the partners' own figures and mostly early lab (preclinical) work, as Anthropic itself says. It also acquired Mithril Medicine, where founder Cody Tranbarger used Claude to run contracts, logistics and experiments with 11 outside labs, going from idea to promising mouse results in under 18 months; human trials could come 'as soon as late 2028'.

Why it mattersMithril is the clearest example yet of one person plus an AI coordinating many outside suppliers. The same pattern fits an agency or consultancy managing freelancers, vendors and paperwork: start by letting Claude track contracts and deadlines for one project.

Sourcesx.com post ↗ · anthropic.com ↗ · x.com post ↗

Gemini API CLI: Google's media models in one terminal command

Google released an experimental command-line tool, gemini-api, a single free, open-source (Apache 2.0) program that makes images, voiceovers, music and video, analyses PDFs and YouTube videos, transcribes audio (including to SRT subtitles) and runs Deep Research, all from the terminal. It has modes built for coding agents (dry runs, usage reports, an agent mode), so Claude Code or Codex can call Gemini's media models as a tool. Fact vs hype: it's real, but beta, and its own docs say it is 'not a supported Google product'. Costs are Gemini API prices: Gemini 3.8 Flash and text-to-speech have a free tier, but image generation is paid only (about $0.03-$0.11 per image depending on size), video is about $0.10 per second and a Lyria song $0.08. The '$1-3 per Deep Research run' figure going around isn't on Google's pricing page, so treat it as an estimate; Deep Research is billed by tokens.

Why it mattersOne tool lets your coding agent produce a voiceover, a thumbnail and subtitles for a piece of content in one go. Try it on something free first (transcribe a meeting recording to SRT) and keep an eye on image costs before automating anything.

Sourcesx.com post ↗ · x.com post ↗ · github.com ↗Fact Hype or Fact? check

Big lab updates

ChatGPT can now read your iCloud Mail and Calendar

OpenAI added separate iCloud Mail and iCloud Calendar plugins to ChatGPT on all plans, on the web, desktop and iOS. You connect each one under Settings > Plugins; Mail needs an @icloud.com, @me.com or @mac.com address, and you sign in with an Apple app-specific password.

Why it mattersIf you or your clients live in Apple Mail, 'find that invoice email' or 'what's on Thursday?' now works inside ChatGPT. Use a dedicated app-specific password so you can revoke access at any time.

Sourcesx.com post ↗ · x.com post ↗

Anthropic: Claude at 40% off for US government

Under a deal with the US General Services Administration (OneGov), US federal, state, local, tribal and territorial governments get 40% off token-based Claude usage with no minimum purchase, covering Claude Enterprise, Claude for Government and the API.

Why it mattersOnly relevant if you sell to the US public sector, but it's another sign big buyers are getting real price cuts. Ask your own vendors about volume pricing.

Sourcesx.com post ↗ · anthropic.com ↗

Now Anthropic claims an AI refuted a major computer-science hypothesis too

A day after OpenAI's claim (see the previous day's news), an Anthropic paper says an internal Claude research model found an algorithm refuting the Exponential Time Hypothesis on its own, using about 14 million output tokens with no human input, and Anthropic has published the Lean proof code. Lean checking is strong evidence, but this is Anthropic's claim and independent experts still need to review it. Spotted by @kimmonismus.

Why it mattersNothing to act on this week. The story is two labs in one week showing long, machine-checked AI work, and the lesson for business is the same: pair AI output with an automatic check.

Sourcesx.com post ↗ · x.com post ↗ · github.com ↗

Quick lab notes

• Gemini can now check whether an image, video or audio clip was made with Google's AI: ask 'Is this AI generated?' It uses Google's SynthID watermark, so it can't detect content from other companies' tools: x.com

• Artificial Analysis rankings: Cartesia's Managed Agents top its Speech Agent Arena with 91.7% task success (against 77.5% for Cartesia Line), and Microsoft's MAI-Image-2.7-Preview is ranked #2 for image editing: x.com and x.com

• Rumours, none confirmed: a new OpenAI model codenamed 'Fe-pichu' and Qwen 4 'Kerria' are said to be on the LMArena testing site (@Mr_Salio): x.com and x.com

Why it mattersUse Gemini's check on suspicious images clients send you, but remember a 'no' only means it wasn't made with Google's tools.

Sourcesx.com post ↗ · x.com post ↗ · x.com post ↗ · x.com post ↗ · x.com post ↗

Agents & use cases

WebMCP isn't replacing MCP: use both

Chrome's developer team clarified the difference. MCP is the back end: a server where your data and actions live, reachable by agents on any platform at any time. WebMCP is a proposed browser standard (a bit of JavaScript or HTML attributes) that exposes actions on your live website to the browser's built-in agent, only while the page is open and the user is there. Their analogy: MCP is the call centre, WebMCP the expert in the shop. Put the core logic in MCP and the in-page actions in WebMCP. The comparison guide itself dates from March (updated May); the post is a reminder.

Why it mattersIf you want agents to book, quote or fill in forms on your site, WebMCP is how you tell them exactly how; MCP is how they reach you when nobody's on the site. It's early, so experiment on one form rather than rebuilding anything.

Sourcesx.com post ↗ · developer.chrome.com ↗

Sentient OS: a private, on-device 'second brain' for Macs

Sentient OS (YC F26) runs a local model on Apple-silicon Macs (Gemma 4 E4B, or your own via LM Studio or Ollama). Overnight it reads the files, messages, mail, calendar and transcripts you choose, builds editable Markdown notes, then offers to finish tasks by using your computer. It can run fully offline and is open source (AGPL-3.0). 'Understands your entire life' is marketing, and it's Mac only. Spotted by @dr_cintas.

Why it mattersA good talking point for privacy-conscious clients who won't put data in the cloud. Grant it access folder by folder, not to everything at once.

Sourcesx.com post ↗ · x.com post ↗ · sentient-os.ai ↗

Concurrence: agents that take patient histories before appointments

Concurrence builds clinical agents that collect histories, symptoms and medications by voice or text before a visit and plug into electronic health records, with a managed rollout in six to eight weeks. The '80% less time on intake' figure is a customer quote shared by the company, and 'Superintelligent Care' is its marketing. Spotted by @kimmonismus.

Why it mattersIntake is a sweet spot for agents in any service business: a pre-call questionnaire agent for an accountant, lawyer or clinic saves the first 15 minutes of every meeting.

Sourcesx.com post ↗ · concurrence.com ↗

Axiom's agents formally checked one of maths' biggest proofs in 24 days

Axiom says its agents, with human guidance, formalised the Classification of Finite Simple Groups (about 15,000 pages of proofs by more than 100 mathematicians) in the Lean proof checker in 24 days, and flagged issues in one published route. That's the company's own account. Spotted by @kimmonismus.

Why it mattersIt's the pattern behind trustworthy agent work: output that a machine can check. Ask the same of your own agent workflows (tests, totals that reconcile, before-and-after comparisons).

Sourcesx.com post ↗ · x.com post ↗ · axiommath.ai ↗

New tools

Supabase scoped access tokens: give your AI agent one project, not your whole account

Supabase personal access tokens can now be limited to specific organisations and projects, with per-area permissions and an expiry of up to a year, plus a low, medium or high risk banner when you create one. They work with the Supabase MCP server, the Management API and the CLI; with only database read access, the MCP server's SQL tool becomes read-only. Fact vs hype: real and documented, but not brand new: it's been generally available since 28 September and is now the default, so the post is a reminder. Gotchas: the 'Read-only' preset still includes API-key secrets, so start from 'No access' and add only what you need; logging in to the CLI through the browser still creates a full-access token; and a database password set in the environment bypasses token permissions. It's included on all plans.

Why it mattersThe simplest security win for anyone letting Claude Code, Cursor or an MCP server near a database: create a token for one project with only database read access and a one-week expiry, and give the agent that.

Sourcesx.com post ↗ · supabase.com ↗Fact Hype or Fact? check

Cursor's design mode gets Annotate: draw on the page, the agent fixes it

In Cursor's design mode you can now select elements, draw on the live page, edit text and check different screen sizes, batch up your notes and send them all to the agent at once. Notes stay attached to the page's address.

Why it mattersPoint-and-draw feedback instead of describing a layout problem in words: ideal when a non-developer is giving design notes on a site an agent is building.

Sourcesx.com post ↗ · cursor.com ↗

Requesty Auto Routing: let a router pick the right model for each Claude Code session

Requesty's beta Auto Routing lets you describe the kinds of work your team does and assign a model to each; it picks a model from the first message and keeps it for the whole session. There's a free tier on free models (200 requests a day); otherwise you pay the model price plus 5%. Any savings are the company's claim, not something we've tested. Spotted by @CompleteSkeptic.

Why it mattersNot every task needs the most expensive model. If your Claude Code bill is growing, routing simple jobs to cheaper models is the obvious lever; compare a week of costs before and after.

Sourcesx.com post ↗ · x.com post ↗ · requesty.ai ↗

Skills & prompting

Matt Pocock: what makes a task good for an agent

Matt Pocock (@mattpocockuk) is looking for a word for the tasks agents are really good at: ones where the agent has all the data, can run experiments, can check its own results independently and can keep improving step by step. In a second post he shows the other half: he knew enough about phonetics to tell his agent exactly what to fix in his video editor. Karpathy and the Stanford lecture in Worth testing below make the same point about verification.

Why it mattersPick agent tasks whose results can be checked (tests, numbers, before and after), and bring your own expertise for the rest. That's also the best way to explain to clients where AI will and won't save them time.

Sourcesx.com post ↗ · x.com post ↗

Worth testing

AI search tip

Find the AI traffic you already get in Google Analytics

In Google Analytics 4, open Reports > Acquisition > Traffic acquisition and look for the 'AI Assistant' channel: Google now groups visits from ChatGPT, Gemini, DeepSeek, Copilot and Grok there, so you can compare their conversion rate with Organic Search. Two catches: clicks from Google's AI Overviews and AI Mode are not in that channel (they count as Organic Search), and Claude and Perplexity aren't named in Google's list, so add a custom channel group if they matter to you. Set up at least one key event (an enquiry form or booking) first, or you can't tell what any of this traffic is worth. For a sense of scale, SEO consultant Marie Haynes reports that across 100 sites ChatGPT sent only 0.62% as many sessions as Google organic but converted at 13.8% against 3.02%; that's a single, small study and her method isn't fully disclosed, so check your own numbers rather than relying on hers.

That's it for today.

Sourcessupport.google.com ↗ · mariehaynes.com ↗ · x.com post ↗

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.