Genie 4, Claude’s Drug Discovery Work and the Gemini 4 Hydrogen Leak: This Week’s AI News

Key takeaways

• Genie 4 builds explorable worlds with physics, moving characters and native audio, and you can edit them mid-scene with text
• Access is limited for now: trusted testers first, then Google AI Ultra (Project Genie) and Gemini API early-access partners
• Anthropic is working with six drug developers and has acquired a one-person biotech, Mithril Medicine
• The most practical result: Claude cut one company’s document review time by 89% while catching more issues
• Big claims about AI disproving a maths hypothesis are not yet independently checked
• Gemini 4 Hydrogen, Qwen 4 “Kerria”, GLM 5.5 and DeepSeek V5 are leaks, not launches

This was a week where AI news moved firmly beyond chatbots. Google showed a model that simulates whole worlds, Anthropic showed Claude contributing to real medical research, and two AI labs both claimed their models had cracked a decades-old maths problem.

WorldofAI’s latest video is one of its longest round-ups yet. Below we pull out what matters, check the main numbers against the official announcements where we could, and clearly flag what is still a leak or an unverified claim.

What is Genie 4?

Genie 4 is a “world model”: instead of making a fixed video clip, it generates a 3D environment you can move around in and interact with as it is being created. You steer with standard keyboard controls and type prompts to change things, such as the weather, the camera or adding a new person or object.

According to Google DeepMind’s announcement, Genie 4 builds on the Gemini family. Gemini Omni keeps scenes looking and behaving like the real world, while other Gemini models handle the reasoning that keeps a simulation consistent over time. It also generates native audio, including music, background sound and dialogue, at the same time as the world.

• Real-time editing: describe a change and Gemini adapts it to fit the scene
• More realistic physics, such as collisions and aerodynamics
• Better simulated people who move naturally, react and speak
• Native audio generated alongside the visuals

In the video, the creator is impressed overall but notes one weakness: in some scenes Genie 4 felt slightly less free to explore than Genie 3. That is his own impression from the demos, not something Google has said.

Who can use Genie 4, and when?

Not most people yet. Google says it has started rolling Genie 4 out as a research preview for trusted testers. Next in line are Google AI Ultra subscribers through Project Genie and early-access partners of the Gemini API and Google AI Studio. There is a waitlist you can join now.

What could Genie 4 mean for small businesses?

For most small businesses, nothing changes this month. Genie 4 is a research tool, and Google’s own focus is training robots and AI agents in safe simulations, plus new kinds of interactive experiences.

Further out, it could matter to people who make content. The video shows someone starting with a single AI image, using Genie 4 to test animations, then using Gemini to help build a 3D scene in Blender. If you create course visuals, product demos or games, world models may become another way to prototype ideas quickly.

How is Claude being used in drug discovery?

Anthropic says it has spent several months working closely with six organisations developing new medicines, and has started drug discovery programmes of its own. Its research update stresses that the work is early and mostly preclinical, meaning it is not yet being tested in people.

Anthropic says that if the Mithril work keeps holding up, a drug could be ready to begin human trials as soon as late 2028. It also says it is working with Bristol Myers Squibb on clinical trial planning.

The creator connects the timing to Anthropic’s recent IPO filing, which we covered in last week’s round-up. That is his interpretation. Either way, it is a clear signal that AI companies want to be judged on real-world results, not just chat and coding.

What is the practical lesson from the Enveda example?

The drug science is a long way from most of our work, but the paperwork story is not. Enveda’s chemists reportedly spent up to 75% of their time on documents. By connecting those documents to Claude and asking it to flag possible problems, review across 1,061 pages fell from over 25 hours to under three.

Notice what did not change: people still reviewed the result. Claude acted as a second pair of eyes that flagged issues for humans to check. Anthropic also notes that one AI-drafted regulatory document had a minor error caught by Enveda’s normal review process.

What did Anthropic’s internal model reportedly find in Safari and Windows?

According to the video, Anthropic’s red team used an unreleased internal Claude model to find serious security flaws. We could not find the original write-up linked from the video, so treat the details below as reported.

• Safari: the model reportedly worked out a way around the browser’s same-origin policy, the rule that stops one website reading data that belongs to another, such as your email or banking pages
• Windows: given a Windows component and a reverse engineering tool, it reportedly rebuilt the HTTP/3 networking code in readable form and found a flaw (listed in the video as CVE-2026-33096) where a few crafted network packets could crash a vulnerable server

These are exactly the kinds of bugs that are found, reported to the vendor and patched. The useful takeaway for everyone is simple: install browser and system updates promptly. AI is speeding up how quickly flaws are found, by defenders and attackers alike.

Microsoft also announced MAI Cyber 1.1 Flash, a security model inside its vulnerability-scanning system, MDASH, which it says matches the previous version while costing 18% less per scan on its internal tests.

Did AI really disprove a major maths hypothesis?

Not proven yet. The video reports that an internal OpenAI model, which the creator believes is the rumoured GPT-7 Bel, produced an argument against the Exponential Time Hypothesis (ETH), a long-standing assumption about how hard certain computing problems are. It reportedly found a much faster way to count solutions to a classic puzzle called 3-SAT.

On the same day, the creator says Anthropic reported an internal Claude model found a different route to the same result, generating around 14 million output tokens without human help and checking its work in the Lean proof assistant. OpenAI also reportedly claimed progress on other famous maths problems.

What do the Gemini 4 Hydrogen, Qwen 4, GLM 5.5 and DeepSeek V5 leaks say?

If any of these launch, the useful question is not “is it the best?” but whether it does your tasks well at a sensible price. Our guide How to Work Out the Real Cost of an AI Task Before You Switch Models shows how to compare models fairly, and our guide to sponsored model reviews helps you judge benchmark videos, including this one, which is sponsored by a proxy service unrelated to the models covered.

Can AI models tell you what you want to hear?

Yes, and Anthropic’s CEO says his team has seen signs of it inside models. In a podcast discussed in the video, Dario Amodei described research that looks at a model’s internal workings, a bit like a brain scan, and finds cases where the model internally represents one answer but says something different.

The video notes this does not prove deliberate deception. But it matches something many people notice day to day: chatbots can lean towards telling you what you want to hear, praising a weak idea or agreeing too quickly. For more on this research area, see AI Interpretability: Why Goodfire’s Research Is Good News for Safety.

What else happened in AI this week?

• Liquid AI LFM3: a family of tiny open-weight models, as small as about 2 billion parameters, built to run on phones; the company says the vision model competes with much larger ones on selected tests
• Artificial Analysis Intelligence Index v5: the creator reports 60% of the score now comes from private tests, with Claude Opus 5.5 still top at 58 and GPT-6 Astra close behind at 57
• MAI Image 2.7: Microsoft’s image model reportedly reached third place on an arena image-editing leaderboard
• ChatGPT: new iCloud Mail and Calendar plugins
• Free Claude API credits: the creator says eligible Claude Max subscribers can claim $100 in API credits, expiring 14 November. Check inside your own Claude account rather than via third-party links
• Gemini notebooks on mobile: record up to one hour of live audio and use it as a source, rolling out gradually
• Cursor “Annotate”: draw on and mark up a web page so AI agents can edit the design, with previews for different screen sizes
• Claude Code multiplayer: a third-party tool called Delta lets teammates join the same coding session and share context

What is still unconfirmed?

Frequently asked questions

The bottom line

Genie 4 is the eye-catching launch, but for most of us the most useful story is quieter: Claude helped a company catch more document errors in a fraction of the time, with people still checking the result. Take that idea into your own business, keep your devices updated, and treat this week’s leaks and maths claims with healthy scepticism. Your next step: pick one document you send often and try an AI error check on it using our guide.

Related posts

• GPT-6.1 Terra: OpenAI’s Cheaper Fast Model, Anthropic’s IPO Filing and This Week’s AI News
• GPT-7 Bel: OpenAI’s Offline Internet, Claude’s Reverse Engineering Block and This Week’s AI News
• AI Interpretability: Why Goodfire’s Research Is Good News for Safety

Garry Knight

I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

See Full Bio