Google's Genie 4 stitches Gemini reasoning, Omni video, and native audio into a world model you can steer live — still a two-minute tech demo, but one that points at robots trained at software speed and generative games. Plus: Anthropic buys a one-person biotech, Wall Street swaps P(doom) for P(boom), and OpenAI can't win on science disclosure.
World models are getting tantalizingly close to the jump from tech demo to real-world application.
Genie 4 isn't a standalone model. It's Google's whole AI stack pointed at one problem: Omni anchors realism, Gemini 3.8 Flash runs in the loop as a reasoning agent, and audio models add sound for the first time. The result is a world you can change in real time with prompts. It's still capped at two minutes and still glitchy, so it remains a toy for now. But the path is visible: simulated training grounds for robots, personalized generative games, and a reminder that DeepMind's jagged frontier is unrivaled on multimodal even while it trails on LLMs.
World models drew attention because figures like Yann LeCun and Fei-Fei Li saw them as a more viable path to AGI than LLMs. Genie 3 let users prompt worlds and navigate them with keyboard and mouse. Its standout trick was emergent persistence: paint a symbol on a wall, look away, and it's still there when you look back. Compute limits capped it at 60 seconds.
DeepMind dropped the standalone design. Gemini Omni anchors simulations in real-world appearance, Gemini 3.8 Flash supplies reasoning for physical interactions and causality, and Google's audio models bring sound to generated worlds for the first time.
Earlier Genie versions let you prompt once and then explore. Genie 4 takes live prompts mid-simulation to steer the camera, change the weather, or introduce new objects and characters. Google also highlights a jump in physics realism, covering aerodynamics and collisions, and in simulating people who move naturally and can be steered in real time.
Generated worlds now carry music, speech, and sound effects, a new level of tooling for creators. The dialogue still has the hollow, echoey quality that marked generative video until a few months ago.
Gemini 3.8 Flash acts as a guiding agent in the loop. It observes the scene, breaks goals into steps, checks incoming frames to verify that each step succeeded, and adapts when the user intervenes. That unlocks multi-step actions with reliable causality.
Using Gemini 4 Argon as the backend, Google showed a robot playing chess inside a Genie simulation. Argon breaks tasks into steps fed to Genie as text captions while the scene renders. The upshot is fully simulated training environments, untethered from real-world limits.
Usually the rush to share flashy renders at a model launch is a distraction from what the model is actually for. Genie 4 is different, because building and sharing visual, interactive worlds is the entire purpose of the model.
Ethan Mollick typed in a neo-gothic brutalist city of robots and horses, then added dragons and a friendly otter NPC, all in real time. Justine Moore built a Lovecraftian world and a GTA riff, quipping "we got GTA Waymo before GTA VI."
The AGI case: pair an LLM's reasoning with a world model's learned physics, and agents can test actions and learn from simulated outcomes before acting in the real world, much as humans imagine a scenario before trying it. That could help with spatial and physical reasoning, provided the simulations are accurate enough.
Genie 3 sent gaming stocks plummeting, while Genie 4 hit them less hard as investors came to understand how far off the disruption is. Meanwhile, Opus 5.5's ability to decompile and repackage games has produced mashups, ports to almost any hardware, and a mixed-reality Mario 64 you can pull out onto your kitchen table.
— Tim Sweeney, Epic Games CEO. The Epic Games CEO was reacting to a tech reporter porting his first game from PC to the NES. What once took vast manual effort can now be a hobby project.
Genie 4 is available to Gemini Ultra subscribers. Simulations are capped at two minutes, double Genie 3's limit but well short of a fully generative game. It's still a tech toy, though one showing glimmers of real usefulness.
After six disappointing months on LLMs, with Gemini 4 Argon earning only a passing grade, Google's multimodal work remains unrivaled. Anthropic's S-1 still named Google as one of just two major competitors. Genie 4 shows that combining disparate model architectures can produce entirely new kinds of AI systems.
Analysts scoffed when Google moved to sell TPUs externally. Sales are now a significant revenue line, though almost entirely from Anthropic. The Information reports two new customers: OpenAI, with terms undisclosed, and Oracle, which would become the first large cloud provider other than Google to rent the chips to outside customers.
Follow-up model Gemini 4 Carbon is already getting positive internal reviews. Because Argon is also a new ultra-large base model, Google can skip another long pre-training run and iterate through post-training. Meta used a similar path to bootstrap Mu Spark into a competitive model across three versions in five months.
AI insiders may overweight raw capability. Logan Kilpatrick reported Google processed around 25 trillion tokens in Gemini 4 Argon's first 72 hours. For Google, existing reach can matter as much as leaderboard position.
Anthropic's view is that there's no single problem to solve, because the bottlenecks are spread across the whole drug development process, from target identification to regulatory filings to trials. Its answer is to get involved in real programs. That means six partner organizations, its own discovery efforts, and a new Bristol Myers Squibb partnership on early-stage clinical trials.
Mythril Medicine's founder, Cody Tranberger, used Claude to run contracts, logistics, and experiments across 11 external labs. He took a muscular dystrophy treatment from idea to promising mouse results in under 18 months. Anthropic hopes to have a drug ready for clinical trials by the end of 2028, against a typical four-to-seven-year runway.
Fans called it Sam Altman's one-person-company prediction coming true. Skeptics pointed out that the founder is a Harvard MBA and ex-Goldman analyst, though he also holds a master's in biotech. Others, including Lenny Rachitsky, doubt any solo company can reach massive scale, if only because of support load. Whether you see the death of expertise or democratization depends on where you started.
Many read the deal as shrewd pre-IPO signaling that AI goes beyond writing code. It also fits a broader diversification trend. One argument holds that selling intelligence is becoming a commodity business, and that labs paid for value will have to discover the drug, run the trial, and hold the patent, much like Alphabet's Isomorphic Labs.
The filing landed after a three-week tech drawdown and a four-month semiconductor bear market. This week the Nasdaq is up 2% and the SOX has rebounded almost 6%. Investors who feared the filing would show no path to returns on AI capex instead saw a booming business, and strategists pointed to Anthropic's $112B cash pile.
The Wall Street Journal echoed Monday's argument that a successful Anthropic raise of $100B, the largest IPO ever, would signal access to capital for smaller companies with more pressing fundraising needs. That's also a reason OpenAI might hold its next release and let Anthropic have the limelight for a couple of weeks.
A Journal column ran the logic half-seriously: an AI smart enough to kill us all but which prefers to help us is supremely useful and should generate healthy gains. Tongue in cheek or not, this week the market is betting on the boom.
After criticism that its earlier math-proof dump wasn't verified, OpenAI appointed expert groups to shepherd each result. The math proof went out as a paper. The computer science result disproving the exponential time hypothesis shipped with its proofs and AI sessions. The physics result, said to include a Yang-Mills existence proof, is being held privately for verification.
— OpenAI's advisory group on mathematics and artificial intelligence. OpenAI's own mathematics advisory group publicly broke with the decision to keep the physics result private. The physicists replied that the manuscripts are formidable and they want to be convinced the result is correct before release, calling it an experiment in how academia handles AI-generated knowledge.
Professors interviewed by The Verge called for immediate release. One said it makes no sense that only a few distinguished scientists can look at the results. UT Austin's Andreas Kark called the handpicked group "a disaster," arguing that labs already have enough power and shouldn't be picking winners and losers.
If OpenAI releases everything at once, it creates a giant verification burden. If it holds results back for vetting, it's accused of gatekeeping. The underlying discomfort is understandable, but it comes from where labs sit relative to this new intelligence, not from any single decision.
