AI Can Now Create Worlds You Can Change in Real Time

October 28, 2026 · Episode Links & Takeaways

HEADLINES

Anthropic Buys a One-Person Drug Company

Anthropic laid out its drug development strategy: get AI involved in every stage, because the bottlenecks are distributed across the whole process rather than concentrated in one problem. Alongside work with six partner organizations, its own discovery efforts and a new Bristol Myers Squibb partnership on early-stage trials, it acquired Mithril Medicine, a one-person company whose founder Cody Tranbarger used Claude to coordinate 11 external labs and take a muscular dystrophy treatment from idea to promising mouse results in under 18 months. Anthropic is targeting clinical-trial readiness by end of 2028, well ahead of the typical four-to-seven-year timeline. Supporters called it Sam Altman's one-person company prediction come true; skeptics, from short seller Scorpion Capital to Lenny Rachitsky, questioned the founder's background and the premise itself. It's a Rorschach test for death of expertise versus democratization, a shrewd pre-IPO move, and another sign of labs looking to own outcomes rather than just sell tokens.

Anthropic Early progress on accelerating drug development with Claude
The Information Anthropic Buys a Small Drug Startup as It Pushes Deeper Into Biotech
Andrew Curran (X) Altman's one-person company prediction is coming true, and Anthropic just bought one
Yuchen Jin (X) The real one-person billion-dollar company: a high-agency solo CEO plus AI
Scorpion Capital (X) Mithril's founder is a Harvard MBA and ex-Goldman analyst, not a scientist
VerdiK (X) The founder holds a master's in biotech
Peter Yang (X) Some of these "one-person companies" look highly suspicious
Lenny Rachitsky (X) Support load alone makes massive-scale one-person companies impossible
Chubby (X) Drug discovery as a future revenue stream, timed to signal IPO investors
Martin Varsavsky (X) Selling intelligence is a commodity; owning what it discovers is not

Anthropic's S1 Revives the AI Trade

The S1 landed after a three-week tech drawdown and a four-month semiconductor bear market, and this week the Nasdaq is up 2% while the Philadelphia Semiconductor Index has rebounded almost 6%. Investors bracing for confirmation that AI capex has no path to returns instead saw a booming business with a realistic path forward; Tiger Brokers' James Ooi pointed to more than $112 billion in cash before any IPO proceeds. The Wall Street Journal argues the whole industry needs this IPO to go well, since a $100 billion raise would signal capital access for everyone else. For this week at least, Wall Street has stopped focusing on p(doom) and started pricing in p(boom).

OpenAI Can't Win on the Yang-Mills Disclosure

After criticism that its earlier math-proof dump wasn't verified, OpenAI appointed expert groups to shepherd each new result, but keeping the physics result (a claimed Yang-Mills existence proof) private for vetting drew objections from its own mathematics advisory group and from professors who told The Verge that handpicking the group is "a disaster." The physicists say they want to be convinced it's correct before release. Release everything and labs create a verification crisis; hold results back and they're gatekeeping science. The discomfort with lab power is understandable, but it stems from where labs sit relative to this new intelligence, not from any one disclosure decision.

MAIN STORY

AI Can Now Create Worlds You Can Change in Real Time

Google's Genie 4 is no longer a standalone model: it stitches together Gemini Omni, Gemini 3.8 Flash and Google's audio models to combine the realism of modern video generation with the reasoning of modern LLMs inside a fully playable world simulator. It remains a tech demo, capped at two-minute simulations, but it's tantalizingly close to the inflection where demo becomes real-world application, with robotics training the most immediate implication. It's also a reminder that DeepMind's frontier may be jagged, but it still has another act.

GENIE 4 ARRIVES

The Road From Genie 2
Breathtaking demos, but far too limited for practical use.
Genie 3's standout was persistence: paint a symbol on a wall, look away, and it's still there, an emergent property of autoregressive frame generation. But a 60-second limit kept it a toy.

Four Big Advances
Google's entire AI ecosystem pointed at world simulation.
Real-time prompting to steer camera, weather, objects and characters; a big jump in physics and human realism; native audio including music, effects and speech (though the dialogue still sounds hollow); and Gemini 3.8 Flash as an in-the-loop agent that plans steps and verifies each one for reliable causality.

Robotics
Robot training at software speed, untethered from the real world.
With Gemini 4 Argon breaking tasks into steps fed to Genie 4 as live captions, Google can build entirely simulated training environments, demonstrated with a robot playing checkers.

The Reaction
Jaws on floors, and this time the demos are the point.
Unlike LLM launches judged on viral 3D renders, interactive worlds are the entire purpose of this model.

Gaming
Not ready for generative games, but the weirdness has started.
Genie 3 rattled gaming stocks; Genie 4 barely moved them. Meanwhile Opus 5.5's ability to decompile and repackage games has produced mashups, ports and a mixed-reality Mario 64.

DeepMind's Next Act
Behind on LLMs, unrivaled on multimodal, and selling chips.
Anthropic's S1 names Google as one of only two major competitors. TPU sales have added OpenAI and Oracle as customers, and Gemini 4 Argon gives Google a new base model for rapid post-training, with Gemini 4 Carbon already in internal testing. Distribution may also matter more than the AI-native crowd appreciates.


ALSO REFERENCED