AI Can Now Create Worlds You Can Change in Real Time

Anthropic announced an approach to AI-assisted drug discovery that treats the process as a series of distributed bottlenecks rather than one problem to automate. The company listed early research, target selection, molecule design and optimization, manufacturing, regulatory filings, and clinical-trial planning as tasks where AI could participate. Anthropic said it has partnered with six organizations, launched drug-discovery efforts of its own, and used Claude in case studies to accelerate scientific and operational work. It also announced a partnership with Bristol Myers Squibb focused on bottlenecks and planning in early-stage clinical trials. The most attention-grabbing news was Anthropic’s acquisition of Mithril Medicine, a one-person company founded by Cody Tranbarger. Anthropic said Tranbarger used Claude to organize contracts, logistics, and experiments involving 11 outside laboratories, taking a possible treatment for a form of muscular dystrophy from an idea to promising mouse results in less than 18 months. Anthropic hopes to have a drug ready for clinical trials by the end of 2028, compared with a typical four-to-seven-year preparation period, largely by reducing the discovery phase. The acquisition prompted a debate over whether it demonstrates the rise of the one-person billion-dollar company or the “death of expertise.” Supporters emphasized that Tranbarger combined Claude with part-time consultants and a master’s degree in biotech. Critics noted that he is primarily a Harvard MBA and former banking and hedge-fund professional, and questioned whether a truly one-person company can handle support and scale. The host’s broader conclusion was that Anthropic is arguing for AI to reduce friction throughout biotech, not to eliminate scientists and other specialists.

The episode connects Anthropic’s biotech messaging to its public offering. The host says the company’s IPO filing appeared after a three-week decline in technology stocks and a four-month semiconductor bear market, but the market turned positive afterward: the NASDAQ rose 2% from Monday’s open and the Philadelphia Semiconductor Index rebounded nearly 6%. The filing did not prove that AI businesses have no weaknesses, but it reassured investors who had been waiting for evidence of a realistic return on AI capital expenditures. Analysts highlighted Anthropic’s financial capacity. James Wee of Tiger Brokers focused on the company’s more than $112 billion in cash and cash equivalents as of June, before possible IPO proceeds. The Wall Street Journal argued that the offering matters beyond Anthropic because its ability to raise as much as $100 billion—potentially the largest IPO ever—would signal that smaller AI companies can still access capital. The host also suggested that OpenAI could benefit the industry by delaying a competing model release and allowing Anthropic to occupy the spotlight. The episode frames the market’s changing attitude through the contrast between “P Doom” and “P Boom.” The Journal discussed how investors might price a hypothetical 10% chance that AI could kill humanity within a decade against a 90% chance that it becomes extraordinarily useful if its destructive tendencies are controlled. The host says Wall Street appears, at least for this week, to be pricing the second possibility more heavily. A related strategic debate concerns whether AI companies should sell access to models or own the valuable outcomes those models create, such as drugs, trials, and patents.

The episode then follows up on criticism of OpenAI’s recent scientific breakthroughs. After scientists complained that OpenAI had released novel mathematical proofs without enough verification, OpenAI appointed expert groups to shepherd each new result. The mathematics result was published as an academic paper. A computer-science result concerning the exponential-time hypothesis was released with proofs in AI sessions and a scientific team assigned to explain and extend it. A physics result described as an existence proof for Yang–Mills theory was kept private while physicists worked to verify and formalize it. OpenAI’s mathematics advisory group publicly disagreed with withholding the physics result. It said that understanding should develop in the open and that everyone should have equal access. The physicists responded that the manuscripts were long and complex, that they needed time to understand the methodology and meaning, and that they wanted to be convinced the work was correct before release. They described the process as an experiment in how academia can handle AI-generated knowledge. Professors quoted by The Verge largely favored immediate publication. Constantine Cotler said it made no sense for only a few distinguished scientists to inspect the work. Andreas Karch called OpenAI’s handpicked group “a disaster,” arguing that large laboratories already have too much power, although he ultimately called for release rather than a different selection process. The host concludes that OpenAI is criticized either way: releasing everything creates a huge verification burden, while organizing expert review looks like scientific gatekeeping. The underlying discomfort is about the power of AI labs, not only this particular disclosure decision.

The main episode turns to Google’s Genie 4, the latest version of its world model. The host traces the progression from early 3D world-model demonstrations to Genie 2 in December 2024 and Genie 3, previewed in January 2025 and made available to users in August. Genie 3 let users prompt a world into existence and navigate it with a keyboard and mouse. Its notable persistence meant that a user could paint a symbol on a wall, look away, and find the same symbol still present. However, it was limited to 60-second generations and remained a technology demo. Genie 4 changes the architecture by combining several Google systems rather than operating alone. Gemini Omni anchors simulations in real-world appearance and behavior, Gemini 3.8 Flash provides a reasoning engine for physical interactions, and Google audio models add sound. Users can now issue prompts while the simulation is running, changing the weather, adding objects or characters, or directing the camera. Google also claims more realistic aerodynamics, collisions, and human movement, with characters that respond naturally and can be steered interactively. Generative audio expands the tool beyond silent environments: worlds can include music, speech, and sound effects. The host says the sound effects are impressive, while dialogue still has the hollow, echoing quality associated with generative video several months earlier. The reasoning system is the largest conceptual change. Gemini watches incoming frames, breaks goals into steps, checks whether each step succeeded, and revises its plan when the user or world intervenes. Google therefore presents Genie 4 as a playable simulation with the realism of modern video models and the planning capabilities of an LLM.

The clearest near-term application discussed is robotics. Google demonstrated Gemini for robotics research working with Genie 4 to simulate a robot playing chess. The reasoning model decomposes a user’s task into steps, and those steps are inserted as text instructions while Genie renders the environment. Because Genie can be steered during generation, the simulation can adapt as the task progresses rather than remaining a fixed video or static scene. The host explains the significance as “training at software speed.” Google can create simulated environments covering many scenarios without tethering robot training to the expense, danger, and limited availability of the physical world. The model could generate imagined environments, observe how actions unfold, and adjust the plan as conditions change. The episode does not present this as solved: physical accuracy and long-term consistency still need improvement. The broader argument is about the limits of text-only reasoning. An LLM can describe physics or a possible action in words, but a world model gives an AI an interactive space in which to try the action and observe the result. Chris GBT argued that combining an LLM’s reasoning with a world model’s learned physical dynamics could let agents compare possible actions and improve spatial and temporal reasoning before acting in reality. The host agrees that Genie 4 offers a meaningful step toward that capability, while stressing that simulations must become accurate enough to support reliable learning.

Gaming has long been viewed as the first major commercial application for world models: players could eventually prompt personalized games and receive fully tailored experiences. The release of Genie 3 caused gaming stocks to fall, but the market reacted less severely to Genie 4 because investors now recognize that the technology is not yet ready to replace conventional game development. Genie 4 is available to Gemini Ultra subscribers, but simulations remain limited to two minutes—twice Genie 3’s duration and still far short of a persistent, fully generative game. The host contrasts that limitation with rapid changes already happening through other AI tools. Opus 5.5 has been used to decompile and repackage games, producing mashups such as Mario in Dark Souls and ports that make games compatible with different hardware. After a reporter ported Tim Sweeney’s first game from PC to the NES, the Epic Games CEO said that work which once required vast manual effort could now be done efficiently as a hobby project. Lucas Martinich also turned Mario 64 into a mixed-reality experience: a player could reach into a virtual screen, use a hand as an extra platform, grab Mario, and place him on a real kitchen table. These examples show that users are already reshaping games with available AI even though fully generated games are not ready. The host says those experiments will influence what future interactive media becomes. Genie 4 is therefore still a “tech toy,” but one that is beginning to show a path toward useful, continuously changeable worlds.

The final section places Genie 4 within Google’s broader AI position. The host says Google has disappointed on the language-model front over the previous six months, while remaining unusually strong in multimodal systems. DeepMind’s progress is described as “jagged”: it has not fully caught up in LLM competition, but its work on world models and other multimodal systems remains difficult to match. Genie 4 demonstrates that combining different model architectures—video, reasoning, audio, and simulation—can produce a novel kind of AI system rather than merely another chatbot. Google’s hardware strategy is another example of a bet beginning to pay off. Analysts previously doubted that Google could sell its TPUs externally because Nvidia dominated AI chips and Google had little sales infrastructure. TPU revenue has grown into a significant line item, although most sales have been to Anthropic. The Information reported that Google had also won OpenAI and Oracle as customers; Oracle was expected to become the first major cloud provider besides Google to rent Google TPUs to outside customers. The host also points to a possible rapid-improvement path for Google’s LLMs. Gemini 3.8 Flash is described as a large base model that can receive further post-training without another lengthy pre-training run, and an internal successor known as Gemini 4 Carbon was reportedly receiving positive reviews. Google said roughly 25 trillion tokens were processed in the first 72 hours after Gemini 4 Argon’s release, illustrating the value of distribution. The concluding judgment is that Genie 4 remains a demo, but world models may be approaching the point where impressive demonstrations become practical applications.