DailySubscribe
OCT 28AI News10 stories

Google's Genie 4: AI Worlds Where the People Talk Back!

Google DeepMind's Genie 4 builds worlds you can walk around in real time, and now the people inside them react and talk back 🤯

Plus: Axiom says the biggest proof in math history is now fully computer-checked, 43.1 million lines of Lean.

Also: a Roblox-labeled crawler grabbed 145 GB from 3D model sites, Valve on AI slop in Steam Next Fest, who's holding OpenAI's physics proof, and Anthropic buying a one-person drug company.

10 stories in this episode

  1. 01

    Genie 4: real-time worlds where people talk back

    Genie 4 is DeepMind's first real-time world model with native audio: people in the scene react and talk. It's a research preview for trusted testers, with AI Ultra access promised but undated.

  2. 02

    Echo-3 turns a phone video into a 3D world

    SpAItial's Echo-3 turns a phone capture into a photorealistic, metric-scale 3D world, from one room to whole buildings, and it's available now.

  3. 03

    A Roblox-labeled bot took 145 GB of 3D models

    Open3DLab says a crawler carrying a Roblox researcher's email in its user agent took 145.5 GB from its three 3D sites, adult content included; Roblox hasn't replied.

  4. 04

    Valve: "fully aware" of slop, but no AI filter

    At Steam Indie Connect in Korea, Valve admitted Next Fest's random first days surface low-quality games, but said it won't police which tools developers use.

  5. 05

    Axiom: math's biggest proof, fully computer-checked

    Axiom says it has fully formalized the Classification of Finite Simple Groups in Lean: about 15,000 pages of human proof became 43.1 million machine-checked lines.

  6. 06

    OpenAI's physics proof: who holds it, who wants it

    A committee chaired by David Gross now holds OpenAI's claimed Yang–Mills proof and will manage its release; a math advisory group says it should be public now.

  7. 07

    New AI ranking: Opus 5.5 leads by a single point

    Artificial Analysis' Intelligence Index v5 leans 60% on private tests; Claude Opus 5.5 (58) edges GPT-6 Astra (57) and GPT-6.1 Sol (55).

  8. 08

    Anthropic buys a one-person drug company

    Anthropic acquired Mithril Medicine, a one-person muscular dystrophy startup where Claude ran much of the work; it says human trials could start as soon as late 2028.

  9. 09

    Stanford authors opt out and sue over pirated books

    Stanford authors Tobias Wolff and Chang-rae Lee left Anthropic's $1.5 billion book settlement to sue on their own, seeking up to $150,000 per work.

  10. 10

    AI plate readers flag "suspicious" road trips

    Automated plate readers now flag 'suspicious' travel patterns that can lead to traffic stops; in an August YouGov poll, 46% opposed police using them.