Nov 3, 2026 · 1h 24m
AI agents move from chatbots to physical-world models
Odyssey CEO Oliver Cameron is 90% sure we live in a simulation | E2358
The episode connects everyday agent workflows to a deeper race to model reality, automate physical work, and build the infrastructure AI requires.
- 1Multi-agent workflows now handle research, design, websites, outreach, and recurring tasks that once required substantial manual effort.
- 2Oliver Cameron argues that video-trained world models could help robots and autonomous vehicles understand physics, causality, and possible futures.
- 3AI’s infrastructure push is colliding with local resistance to data centers while companies explore satellites and space-based computing.
Don't miss
Oliver Cameron’s discussion of simulation theory turns Odyssey’s world models into a question about whether reality itself could be simulated.
The brief
Jason and Lon begin with GrokBot’s popularity, pricing, usage limits, and opaque feedback, then show how specialized bots can research, design, source, and launch real projects.
Their argument is that newer agents matter less because they chat better than because they connect to context, browsers, virtual computers, and recurring workflows.
Oliver Cameron describes Odyssey’s world models, which learn motion, physics, and cause and effect from video to simulate possible futures for robots, drones, and autonomous vehicles.
The interview’s sharpest turn is philosophical: if world models can generate coherent artificial realities, Oliver says reality itself may be a simulation, while the group tests what that claim means.
The discussion returns to consequences, from tactile limits in robot training and public opposition to data centers to AI-built interfaces, platform access, and computing infrastructure in space.
What was said on this episode
17 statements · 12 positive · 2 negative · 1 mixed · 2 neutral
The bot’s refined Substack recommendations produced six-percent higher open rates.
“The open rates were 6% higher.”
Listen at 5:23
Properly deployed AI produces fifty dollars of output per dollar invested.
“A dollar invested in AI today, if you're doing it properly, is $50 of output.”
Listen at 13:36
Agent systems became more useful because they gained context and access.
“It's that the agents have. context, and access. That's what changed.”
Listen at 14:48
AI providers will raise prices to eliminate losses after subsidizing adoption.
“I think what's going to happen is this is kind of like Uber and Lyft when they were discounting rides to get people to use the software. They're going to start boiling the frog. They're going to start raising the prices and take the losses out.”
Listen at 17:22
Odyssey’s world models learn physics, dynamics, and cause and effect from visual observations.
“they learn physics, they learn dynamics, cause and effect, all of these fundamentals of the universe.”
Listen at 26:12
Oliver Cameron is 90 percent certain reality is an advanced simulation.
“I'm 90% certain that this is a simulation, a very advanced simulation of some sort.”
Listen at 32:23
World models prioritize computation along an agent’s path and leave irrelevant surroundings uncomputed.
“everything outside of that, unessential to be computed.”
Listen at 36:16
Odyssey’s world models currently support about twenty simultaneous agents per session.
“Right now, we can support something like 20 agents inside a world model session simultaneously.”
Listen at 36:31
Driverless-car technology was built using an overly narrow approach.
“we were building that technology all wrong, I believe.”
Listen at 41:19
World models can become good drivers after general training plus twenty-to-thirty hours of driving experience.
“the model is able to take those very general concepts and the 20 to 30 hours of experience and then become a really good driver.”
Listen at 42:34
Odyssey considers physics accuracy a crucial development priority for world models.
“one of the hills that we believe is most important to climb is physics accuracy.”
Listen at 45:37
Jason Calacanis supports potentially slowing automation and self-driving deployment.
“I'm not opposed to throttling the pace of. uh automation and self-driving”
Listen at 1:03:56
Lon Harris prefers Waymo rides to rides with human drivers.
“I prefer the Waymo experience to these strangers.”
Listen at 1:04:31
Users will increasingly use agents to access platforms such as Reddit, Hacker News, and LinkedIn.
“People are going to have agents. They're going to have bots. They're going to want to access your Reddits or your Hacker Newses or your LinkedIn.”
Listen at 1:12:52
Space-based data centers are considered viable after several development iterations.
“absolutely viable”
Listen at 1:14:36
Space-based computing could eliminate real-estate costs and eventually make solar energy costs negligible.
“The real estate cost is free out there for life. Free for life. And I think the energy cost out there is going to be also, with solar, free for life.”
Listen at 1:17:01
Manufacturing will eventually expand into space.
“we're eventually going to have manufacturing in space.”
Listen at 1:21:14
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Weekly email
State of the Pod, every Monday
The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.