Google adds live editing to its AI worlds
Eligible Ultra subscribers get access and a gallery for sharing and remixing worlds. Pro users get a waitlist.
By Saeed Ezzati8 min read

The audio edition
Listen to this newsletter
0:006:03
Read transcript
Google is letting some subscribers change an AI-generated world while they’re inside it. The Genie 4 update to Project Genie began rolling out on November 18 to eligible Google AI Ultra subscribers. Users can type a prompt during an exploration to change the weather, add a character or redirect the action, without leaving the environment. The change is more than visual. Genie 4 generates audio alongside the scene, including ambient sounds and character dialogue that respond to what’s happening. A new creation agent uses conversation to build the initial setting, visual style and background score. And Gemini can observe the screen during exploration, offer suggestions about where to go, or trigger events automatically. Google is also adding a community gallery. People can start with another user’s world, work with the agent to change its structure, soundtrack or rules, then publish the remix. That makes these environments less like one-off generated clips and more like works people can explore and build on together. But the experience is still experimental. Google identifies camera latency, complex physical interactions, precise real-time audio and long-term scene permanence as research challenges. In other words, the environment may look convincing, but it can still struggle to respond smoothly or stay coherent over time. Google says generated video and audio carry SynthID watermarks, and that model inputs and outputs are safety-screened. Genie 4 doesn’t replace Genie 3 for every use. Google is keeping the older model available for longer sessions and worlds based on panoramic images. Its session limit rises from one minute to three minutes per generation, and users can link worlds into a longer journey. AI Pro subscribers can join a waitlist for Project Genie, but Google hasn’t announced when they’ll get access. The notable shift is from generating a world before entering it to shaping it as you explore. Whether that feels genuinely interactive will depend on how reliably the world responds—and whether it can hold together beyond a short, impressive moment. That gap between an impressive demonstration and dependable performance also matters in chip design. OpenAI says models found up to 14 percent improvement in power, performance and area across selected AMD GPU circuits. That is a result for particular design blocks, not a claim that whole GPUs are 14 percent faster. The models rewrote SystemVerilog code describing the circuits. Synopsys software checked that each change preserved the circuit’s behavior and evaluated the tradeoffs. OpenAI says some edits are in an AMD design sent for manufacturing, while others have been added to design files for future versions. It characterizes the GPU progress as roughly a year’s worth in less than five months; that comparison comes from OpenAI, not an independently measured engineering timeline. OpenAI also reports single-digit percentage gains in selected blocks of AMD’s next-generation flagship CPU. Separately, it says model-proposed changes to its Jalapeño inference chip reduced standard-cell area in two units, with those optimizations proven during initial silicon testing. These are block-level results, not whole-processor performance figures. OpenAI and Synopsys are developing GPT-Synopsys for chip-design workflows. The limits of what can be confidently inferred from technical systems are especially clear in a separate debate over recent outages. Homeland Security Secretary Markwayne Mullin relayed an unnamed contact’s theory that self-modifying AI software might overwhelm systems unable to keep pace. He called it a possibility, not an established cause, and said he was 98 percent certain the incidents did not involve a cyberattack. The operators described different faults. Microsoft traced a nearly 24-hour disruption to a software bug introduced in July, increased network traffic and reduced server capacity; it explicitly said the outage was unrelated to AI agents or cyber activity. The FAA said a flood of data destabilized older air-traffic-control equipment, affecting radar feeds and communications at 17 of its 22 centers. The agency’s root cause remains under investigation. Chase attributed its banking disruption to a software bug in hardware systems on its own premises and said there was no evidence of a cyber incident. Transportation Secretary Sean Duffy said the FAA’s SMART AI tool played no role. Security experts have called for careful investigation of unexplained failures, but that concern is not evidence that AI caused them. The useful distinction is between a theory about possible system-wide effects and causes that operators have actually identified—or have yet to establish. That question of what the evidence really says carries into the market for AI itself. Financial outlets citing sources report that Anthropic’s IPO was priced at $83 a share, above its planned range, ahead of its scheduled Nasdaq debut on November 19. NBC News reported a target market value of $1.5 trillion and first-half 2026 revenue above $16 billion. The reported growth is striking: revenue was $41 million in the same six-month period three years earlier, and NBC reported that Claude Code reached $1 billion in revenue within six months of its February 2025 release. But the listing comes as technology shares have sold off and investors question the cost, debt and demand behind the AI buildout. Vanda Research estimated retail investors could buy about $250 million of shares when trading begins; that is an estimate, not a confirmed total. A strong debut could offer a useful signal about investor appetite, but it won’t settle the larger question of whether AI investment pays off. Across these stories, the thing to watch is the distance between a promising result and performance that holds up in the real world: responsive worlds, verified circuit gains, established outage causes and durable customer demand.



