Google DeepMind releases the Genie 4 world model, which can generate sound effects and converse with characters in the scene

AIGoogle’s AI worlds can now talk back

PLUS: Make editable product mockups with GPT-6.1 Terra

Zach Mink

Good morning, AI enthusiasts, and welcome to our 5,193 new readers. Last year, Genie 3 let you wander an AI-generated world for a few minutes. Genie 4 lets you talk to the ‘people’ living in the simulation.

Google DeepMind’s newest world model brings its own soundtrack too, and a demo of a humanoid robot playing checkers inside one of its worlds shows where Google wants to take things next.

Reminder: Our next live workshop is today at 2 PM EDT — join and learn why AI deep research has gotten worse over the past six months, and how to get sharp, decision-ready reports out of it again. RSVP here.

In today’s AI rundown:

Genie 4 brings sound and interactivity to AI worlds

Anthropic starts developing its own drugs

Make editable product mockups with GPT-6.1 Terra

Axiom’s AI swarm checks math’s longest proof

LATEST DEVELOPMENTSGOOGLE DEEPMIND

🧞 Genie 4 brings sound and interactivity to AI worlds

Image Source: Google DeepMind

The Rundown: Google DeepMind just introduced Genie 4, its next-gen world model that turns prompts into explorable, photorealistic scenes in real time, with features like built-in, natively generated audio and lifelike people that users can actually interact with.

The details:

Genie 4 is the first Genie model to generate its own sound, like background noise and music, with characters in the worlds also now able to hold live conversations.

Users move with keyboard controls and can use text prompts to create changes mid-scene, with the scene staying coherent between inputs.

A simulated Apptronik Apollo robot played checkers in one demo, with Gemini 4 Argon planning each move in a nod to Genie’s future robot-training applications.

Why it matters: Genie’s progress has been pretty wild to watch, and this iteration’s jump, with sound and characters you can actually talk to, takes it somewhere new. The gaming applications are obvious, but Genie is starting to look as much like a full-blown director’s tool as a world simulator. TOGETHER WITH CUBE

🧠 The semantic layer for AIVisit cubecloud.dev

The Rundown: Cube is the semantic layer for AI. Every agent gets the same governed answer from your metric definitions and access rules. Ask the hardest questions, get real answers.

With Cube, you can:

Connect Snowflake, Databricks, or BigQuery

Let AI build and maintain your data model

See how Brex reached nearly 90% answer relevance

Try Cube free (no card needed). ANTHROPIC

💊 Anthropic starts developing its own drugs

Image source: Anthropic

The Rundown: Anthropic shared early results with its drug-development partners and revealed it’s now running its own drug discovery programs, including buying Mithril Medicine, a solo-founder biotech whose Claude-assisted muscular dystrophy candidate could reach human trials as early as late 2028.

The details:

The candidate showed promise in mice within 18 months, with Claude helping solo founder Cody Tranbarger juggle contracts and tests at 11 outside labs.

Tranbarger also turned out to be “Professor Oak,” a biotech account run pseudonymously on X for six years, and calls the setup a “Pipeline-in-a-Person.”

At Enveda, Claude wrote a first draft of a key regulatory filing in 4 hours vs. ~3 months by hand, and also flagged seven batch-record issues humans missed.

In another case, three of roughly 35K antibodies Claude designed for Divergence Labs appeared to beat today’s main drug at unhooking allergy triggers from cells.

This week’s IPO filing pitched a global pharma market expected to top $2T this year, but Anthropic warned its own drug bets “may never generate revenue.”

Why it matters: Dario Amodei said in August to expect “early glimmers” from Anthropic’s biology push in the coming months, and we’re seeing the groundwork for that. With Anthropic also running its own wet lab, and its IPO filing saying Claude can operate gear like microscopes, the pieces for AI’s hands-on science are falling into place fast. AI TRAINING

📦 Make editable product mockups with GPT-6.1 Terra

The Rundown: In this guide, you’ll use the new GPT-6.1 Terra to build a product mockup in Blender. You’ll create two color options and a wide campaign image, then save the scene so you can handle client changes without starting over.

Step-by-step:

Save an empty Blender project and connect Terra through ChatGPT Work’s Computer Use plugin. Keep its access limited to your practice project

Give Terra the guide’s carton dimensions and colors. Check the size, then add the product name and label as editable text

Add a camera and studio lights, then render a small preview. Check that the lettering is readable, the edges look right, and the carton sits on the ground

Copy the scene to make a second color option. Create a wide version with the product on the right and space for a headline on the left

Export all three images and reopen the saved Blender file. Check the image sizes and confirm you can still change the text and product colors

🔍 Consultants take months. Scribe takes days.Visit scribe.com

The Rundown: Most orgs map workflows the slow way: outside teams, interviews, and months of workshops. Scribe Optimize maps how work happens across your org automatically.

With Scribe Optimize, you can:

Map real workflows — no surveys or consultants

Surface inefficiencies leadership can’t currently see

Build your AI strategy on ground truth, not gut feel

See what's actually happening in your org. AI & MATH

🧮 Axiom’s AI swarm checks math’s longest proof

Image source: Axiom

The Rundown: AI math startup Axiom completed a fully machine-checked version of math’s longest-ever proof, called the ‘Classification of Finite Simple Groups’, with its team steering an AI agent swarm that rebuilt the decades-long effort in 24 days.

The details:

The theorem catalogs what Axiom calls the “periodic table of symmetry,” a full list of the basic pieces that every symmetry pattern in math breaks down into.

100+ mathematicians spread the original across roughly 15K pages and 500+ papers, and Axiom says no single expert has read it cover to cover.

Axiom broke the work into 108 milestones, then dozens of agents produced ~43M lines of Lean, the coding language that lets a computer verify each logical step.

The theorem stands, but AI caught a false claim in one key step, then patched it with a new proof based on 1976 work by mathematician Ronald Solomon.

Why it matters: OpenAI used agents to crack the Navier–Stokes Millennium Prize Problem in September, and now a startup has pointed one at a 15K-page proof that no mathematician could check alone. Dario Amodei’s “country of geniuses in a datacenter” feels like it’s coming to life in the sciences, and a lot more progress is coming with it.

▸ Anuradha built a one-person sales team with Claude

Today’s workflow comes from reader Anuradha:

“I use Claude, connected live to Airtable through MCP, as a one-person GTM team. It runs a daily pipeline war room, drafts outreach, and helps me create data-driven LinkedIn content.

I start a morning war room session by opening Claude and letting it read my live Airtable pipeline. It gives me a scoreboard of what moved, what is stuck, and who needs a touch that day. It drafts outreach using a written style rulebook: peer tone, hypothesis framing, short sentences, and no brochure language. I review everything and send it myself.

I log every LinkedIn post in Airtable with its full text, followed by its analytics: impressions, industry mix, seniority, comments, and notes on what worked. Before writing a new post, Claude reviews the full post history, ranks the posts, and identifies which patterns drove reach and engagement from my target audience. What makes this work is that the analytics live in the same table as the content, so every post makes the next one smarter.”

See the workflow on our Community Workflow Hub →

🛠️ Trending AI Tools

💧 LFM3 - Liquid AI’s new family of small, on-device multimodal models

🎨 MAI-Image-2.7 - Microsoft’s upgraded image AI, launching first for editing

🔎 Parallel Decisions API - Decision AI that classifies with live web data

🌎 GPT-6.1 Terra - OpenAI’s cheaper, faster model for everyday computer work

📰 Everything else in AI today

Anthropic published a paper saying an internal Claude model also disproved the Exponential Time Hypothesis, coming the same day as OpenAI’s own result.

U.S. President Donald Trump posted that he will be “BANNING the term, ‘AI/Artificial Intelligence,’ in all relevant documents” in another push towards Super Intelligence.

AI-for-physics lab Physical Superintelligence released five new papers, each paired with a computer-checked proof, claiming to make progress on long-unsolved quantum physics problems with help from its in-house AI.

Meta lined up $4B for its $95B AI fundraise from chip suppliers Nvidia ($3B) and AMD ($1B), with both companies also expanding the hardware they run in Meta’s data centers.

Google inked a deal to put its TPU chips inside Oracle’s cloud from 2027, the first time a rival cloud giant will offer the in-house Nvidia alternative to customers.

🎓 Highlights: News, Guides & Events

RSVP to workshop today @2 PM EDT: Get consultant-grade deep research

The AI bubble meets Anthropic’s real numbers

Zach Mink • 8 minutesAI

Google’s Gemini 4 Carbon tops the charts (for now)

Zach Mink • 8 minutesAI

Kimi continues to challenge the frontier with K3.1 Pro