DeepMind's Genie 4 finally gives its world model a soundtrack

Good morning 👋 DeepMind just gave its world model a set of ears. Whether "native audio" survives contact with testers outside the building is the part worth watching.

In today's issue:

• 🔊 Genie 4 gives DeepMind's world model ears, with zero benchmarks to back the claim
• 🧠 Liquid AI's LFM3 crams vision, audio, and omni into a phone-sized backbone
• 🔬 Model transparency is getting worse, not better, across 41 releases
• 🚀 Hyphenate's $45M Series A bets CFOs want 90% of close automated
• 🛠️ Try this: script your scene with Gemini before you touch a world model
• 🔗 Quick links: Physical Intelligence's reported $20B talks, Oracle inside ChatGPT, and more

🔭 THE ONE THING

🔊 DeepMind's Genie 4 finally gives its world model a soundtrack

Google DeepMind shipped Genie 4, the first version of its world model that generates audio natively: ambient noise, music, dialogue that reacts to what you do inside the scene. Under the hood, Gemini Omni handles visual grounding while Gemini 3.8 Flash runs it as the guiding agent, watching the simulation and replanning as it unfolds. Right now it's a research preview for trusted testers only, with Google AI Ultra's Project Genie and API/Studio access still weeks out, and DeepMind shipped it with zero benchmarks, no resolution numbers, no latency figures. That silence is the tell. A world model you can hear is a real step past the mute, drifting physics toys we've seen so far, but "native audio" stays a claim until outside testers can poke at it, and DeepMind's own team still calls real-time audio accuracy an open research problem.

🧠 MODELS & RELEASES

• 🧬 Liquid AI shipped LFM3: a single 1.7B-parameter backbone branching into text, vision (LFM3-VL-2B), audio (LFM3-Audio-2B), and a combined omni model. The audio variant does 87x real-time speech synthesis and sub-50ms response on edge hardware, and the whole family quantizes down to roughly 1GB at 4-bit. This is a phone-and-robot bet, not a datacenter one. Hugging Face

🔬 RESEARCH HIGHLIGHTS

• Heads of Security teaches a model to split attention between trusted and untrusted text at the head level, so injected instructions in a document or tool output can't hijack the real prompt. Against RL-Hammer, a strong black-box jailbreak, it holds attack success to 10%, versus 85-97% for SecAlign and other current defenses. The catch: it requires retraining, not a wrapper you slap on someone else's API. arXiv:2610.37657

• SafeTulu answers a question most fine-tuners can't: which training rows are quietly making your model less safe, even when none of them read as harmful on their own. The team's gradient-tracing method (SimpleLoRIF) flagged and cut the worst 20% of Tulu 3, and the resulting model cut jailbreak success by up to 22% across Qwen, Llama, and Olmo. First real tool for this instead of eyeballing the dataset. arXiv:2610.37481

• Model cards are getting worse, not better. An audit of 41 releases across 17 families against the Foundation Model Transparency Index found disclosure fell in 11 generation-over-generation comparisons and improved in only 6. OLMo-3 scores identically to OLMo-2 while quietly swapping three disclosures for three different ones; Llama-3.1 looks like progress over Llama-2 but drops two data-related items along the way. Don't take a flat or rising transparency score at face value. arXiv:2610.37465

🚀 AI STARTUPS

• Hyphenate raised a $45M Series A (Nexus Venture Partners leading, with B Capital, Foundation Capital, BoldCap, and Gaia Ventures), pushing total funding to $54M. The CFO-automation startup says its agents now handle roughly 90% of routine finance work for customers, and HiBid's month-end close went from three weeks to one day running on it. Revenue is up 86x since the seed, which is either a real inflection or a tiny starting base. Worth tracking renewal rates before you buy the multiple. Press release

🛠️ TRY THIS

Prototype a reasoning-driven scene with Gemini before you touch a world model

1. Pick a short scenario (a delivery robot navigating a cluttered kitchen, say) and write it as a scene brief: setup, constraints, the one decision point you actually care about.

2. Run that brief through Gemini's reasoning mode first, not a world-sim tool. Have it produce the physically plausible sequence of events as plain text.

3. Feed that output into your simulation or video generator as the script. You're now generating from a reasoned plan instead of raw improvisation.

4. If sound matters to the result, mark where audio cues should land before you generate. Native-audio world models are closing the gap between "describe the scene" and "render it with sound." Plan for that now, not after.

Prompt: Reason step by step through what physically happens in this scene: [your scenario]. Flag every point where the outcome could plausibly go two ways, and pick one, explaining why.

Worth a look

• Andee drives your phone or Mac the way you would, taps and app switches included, no root and no per-app integration required. Memory and passwords stay on-device. 3,200+ stars on GitHub already. If you're building agents that act in real environments instead of simulated ones, it's a cheap way to test that loop this week.
• AWS's SOCI-indexed images take the container download off SageMaker's critical path. Jobs lazy-load the image and start in seconds instead of minutes, no changes to your job definition. Small plumbing fix, but at world-model training scale, every cut job-start adds up fast.

🔗 QUICK LINKS

• Physical Intelligence is reportedly in talks to raise $1B at roughly $20B, double its last round. Nobody's confirmed the deal yet, so treat the number as a trial balloon. Bloomberg
• Simplismart landed Nvidia as a backer in a $14M Series B, chasing the same problem everyone with a GPU bill has: running models on fewer of them. TechCrunch
• Oracle apps and agents are moving into ChatGPT under an expanded OpenAI tie-up, another front door into enterprise software that isn't Oracle's own. OpenAI
• Anthropic cut Claude pricing 40% for every tier of the US government through a widened GSA deal. Anthropic
• AllenAI built a benchmark, NoveltyBench, to test whether LLMs can actually tell a novel research idea from a dressed-up old one. Hugging Face
• Anthropic is carving off a slice of its own equity for a new Public Benefit Fund, structured to grow its stake as the company does. Anthropic
• Exowatt fine-tuned its way past frontier-model accuracy using SageMaker's serverless customization, at up to 49x lower cost. AWS
• cBRAIN is using open-source NVIDIA tooling to synthesize dementia research across 57 million patient records, work no single team could read alone. NVIDIA
• Check Point is using OpenAI's Daybreak to shrink the gap between finding a vulnerability and shipping a fix. OpenAI
• Property Finder discovered 96.8% of its vision-AI eval labels were just inherited from the old model, which means the "accuracy" it was measuring against was mostly the thing it was trying to replace. AWS
• Retell now routes 100M monthly calls through OpenAI's voice stack, at a reported $1.6M revenue run rate per employee. OpenAI

Enjoy the quiet before Genie 4's audio goes live outside Google's walls.

Pradeep Perugu