AI Can Now Create Worlds You Can Change in Real Time

Google’s Genie 4 is designed to generate interactive worlds that users can change while a simulation is running. Google says it combines Gemini Omni for real-world appearance and behavior, Gemini 3.8 Flash for reasoning about physical interactions, and Google audio models for music, speech, and sound effects. Users can steer the camera or change weather, objects, and characters; the stated simulation limit has grown from Genie 3’s 60 seconds to two minutes.

Gemini 3.8 Flash can observe a scene, break a goal into steps, check subsequent frames, and revise its plan as the world or user input changes. In one demonstration, Google used Gemini 4 Argon as a reasoning backend for a robot playing chess inside a Genie simulation. The combination could help researchers train robots in imagined settings without real-world constraints, though the demonstration does not establish that the simulated skills will transfer reliably to physical robots.

Genie 4 extends a line of work that Google first revealed with Genie 2 in December 2024. Genie 3, previewed in January 2025 and made public in August, let users create and navigate worlds and could preserve changes such as a painted symbol after the user looked away. Google attributed that persistence to autoregressive frame generation, while high compute requirements limited Genie 3 demonstrations to 60 seconds. Genie 4’s longer runtime and steering make worlds more interactive, but scenes can still be glitchy, generated speech can sound hollow and echoey, and physics and long-term consistency need improvement.

At two minutes, Genie 4 remains too constrained for fully generative games and is closer to a tech toy than a practical application; it is available to Gemini Ultra subscribers. The host sees personalized games and more capable robotics as possible longer-term uses, but said gaming disruption is not imminent. Genie 3’s release sent gaming stocks plunging, while Genie 4 had less market impact. World models have also attracted interest as a possible route toward AGI: Yann LeCun and Fei-Fei Li have argued that general intelligence requires physical understanding, not just reading.

Anthropic said AI could reduce bottlenecks throughout drug development, from target identification and molecule design to trials and regulatory filings, without fully automating the process. It acquired Mithril Medicine, whose founder used Claude and part-time consultants at 11 external labs to take a muscular dystrophy treatment from an idea to promising mouse results in under 18 months. Anthropic hopes to prepare the treatment for clinical trials by the end of 2028, against a typical preparation period of four to seven years. Mithril was described as a one-person company, but the founder relied on consultants; Scorpion Capital noted his Harvard MBA and former Goldman Sachs and hedge-fund work, as well as his master’s degree in biotech. Peter Yang questioned the credibility of some one-person-company claims, while Lenny Rzicicki argued that support demands constrain massive scale unless contractors are excluded from the count.

OpenAI took different approaches to disclosing three scientific results: it published a mathematics paper, released computer-science proofs alongside further scientific work, and held back a physics result while it was reviewed. The withheld result reportedly included an existence proof for Yang-Mills theory. OpenAI’s mathematics and artificial intelligence advisory group opposed keeping it private, while appointed physicists said the lengthy, complex manuscripts needed more assessment before they could be convinced. The host described criticism in both directions: OpenAI has faced questions about releasing results before academic verification and about relying on its own review groups.

The host said Anthropic’s IPO filing appeared to reassure investors about the AI business’s path forward, and that the offering’s reception could affect industry access to capital. James We said Anthropic had more than $112 billion in cash and cash equivalents as of June; the host said it could raise $100 billion “with room to spare,” potentially making it the largest IPO in history. The host also noted that the Nasdaq was up 2% since Monday’s market open and the Philadelphia Semiconductor Index had rebounded almost 6% during the week.