A Fly Brain Flew Through a Video World That Was Never Filmed
The approach studies vast video and tries to learn motion and cause and effect by guessing what comes next.
This Week in Startups · Odyssey CEO Oliver Cameron is 90% sure we live in a simulation | E2358The aerial video looks like drone footage drifting over a river town in Spain or Portugal. Every pixel was imagined by artificial intelligence, in real time, as a virtual fruit fly steered itself through it.
The pilot was a mapped brain from a fruit fly species. An open source project connects that map to actions, so the brain can decide to move, and Oliver Cameron, co-founder and chief executive of Odyssey, a company that builds world models, said that once placed inside the generated world, the fly flew.
A world model, meaning software that studies moving images to grasp how physical things behave, makes that flight possible. Odyssey feeds its models huge libraries of everyday sights, then asks them to guess what could happen next in a scene.
Cameron pictures the learner as exceptionally attentive. Picture a tireless student parked before a television for ages of footage, never looking away, filling a notebook with notes on motion, behavior, people, and how objects act.
The idea is that repetition teaches fundamentals. Cameron's aim is for that practice to build knowledge of physics, dynamics, and cause and effect. Learning fluid dynamics and gravity, he says, is part of what physics accuracy would mean, which he calls one of the most important hills to climb.
Cameron cautions that the demonstration should not be taken as a faithful account of how a real fly brain works in the physical world. He describes the effort as early and incomplete.
In Cameron's account, the payoff would be twofold. For robots, a world model would act as a general brain for judging the consequences of an action before acting. For software agents, he says such models could create endless realistic places to practice, improving work on tasks learned inside them.
That matters because Cameron thinks current language models skip important stages of how intelligence develops. Humans grow up inside a world, shaped by surroundings and other people, and he wants to place both new and mature agents inside worlds that feel utterly real so they can learn the same way.
Questions people ask
How can watching video teach an AI about touch, weight, or force?
That remains an open question in the interview. The question was put to Cameron about weight, a paper bag versus a rock, and he did not show that video alone solves it. He said Odyssey trains on vast amounts of video by predicting how the world evolves next, and sees getting physics right as one of the most important hills it is still trying to climb.
Does the virtual fly know it is flying?
Odyssey showed only that the connected brain map produced flight-like movement in its generated world. The company presented no evidence about awareness. Whether flies or machines have subjective experience remains an unsolved question.
How is a world model different from a chatbot?
A chatbot predicts the next word in text, while a world model predicts the next state of a physical scene in video. That means learning how objects move, collide, fall, splash, and respond to actions. The output is a simulated environment, not an answer in sentences.