Odyssey CEO Oliver Cameron is 90% sure we live in a simulation | E2358

The hosts open with Elon Musk’s explanation that GrokBot’s serving cost was “unsustainably high relative to revenue,” after users complained that their weekly limits disappeared within days. Musk gave no usage or pricing details; GrokBot had recently been described as SpaceX’s fastest-growing product, with 800,000 paid weekly users and a $30 monthly entry plan, plus $100 and $200 tiers. Jason says his company spends under $40,000 annually on AI tools but gets more than $2 million in value. Lon argues that GrokBot should show a visible usage bar and tell users how much compute each task consumes.

Jason says he pays for the $200 GrokBot tier for himself after adding team members on $30 plans, and has built specialized bots for video, research, LinkedIn, HTML pages, and Substack analytics. His analytics bot posts daily results to Slack and consults research bots, allowing it to recommend subjects based on readership and open rates; one comparable title performed 6% better. He says the team has automated at least seven full-time jobs and become roughly 100% faster at average tasks. Jason rejects outdated assumptions from 2023–24 that AI lacks context or cannot be trusted: connected bots with access and memory are a different product.

Jason demonstrates a GrokBot-built site for choosing jackets for Founder University Japan and Launch Accelerator. After roughly 10–15 minutes of voice instructions through WhisperFlow, the bot researched jacket types, created mock-ups, placed subtle company logos, suggested collar and inside-label treatments, sourced examples, and published the site to Jason’s URL. It also found a cultural distinction: Japanese consumers may find Japanese writing on clothing nerdy, while English text can seem stylish. The bot recommended a versatile chore jacket combining Texas ranch, Parisian, and Japanese influences, and extracted a design rule: patches look best grouped in one placement zone rather than scattered like a costume.

Jason estimates that reproducing Launch’s software, design, research, content, podcast production, editing, and project-management output conventionally would require roughly 12–14 employees, including five or six developers, two designers, two project managers, and two or three additional content staff. At about $100,000 per employee, he calculates approximately $2 million in annual value against $40,000 of AI spending—roughly $50 of output per dollar. He stresses that value is only useful if tied to a business goal. The company is still hiring for distinctly human work such as on-camera performance, taste, community management, events, and video editing.

The hosts attribute the recent improvement not merely to calling software “agents,” but to agents gaining context, service connectivity, and virtual computers. Earlier efforts with OpenClaw, Perplexity Computer, Claude Cowork, ChatGPT, and Slack bots were brittle: they could lose access to Slack, Notion, email, calendars, or browser sessions and fail to complete assignments. GrokBot’s virtual machine, vault, and connected services now let Jason’s assistants supply travel preferences, mileage information, and goals, then have the system book or research things to completion. He recommends building a bot for every recurring assistant task and says even a $1,000 monthly plan could remain worthwhile.

Jason and Lon agree that GrokBot’s limits caused extra frustration because users could not easily see what they had consumed. Lon discovered that usage can be monitored at cursor.com/dashboard/spending, while Jason says his Ultra account was at 51% with the weekly reset approaching. They want a prominent countdown bar, per-job feedback such as “0.7% of your weekly total,” and recommendations that match a user’s workload to a $20, $100, or $200 tier. Jason believes providers will gradually raise prices, as Uber and Lyft once subsidized rides, but says users should calculate their return rather than complain automatically; his own team would keep using GrokBot even at $1,000 per person.

The hosts introduce Oliver Cameron, co-founder and CEO of Odyssey, after the company’s video of a fruit-fly brain flying through an entirely AI-generated world received 6.2 million views on X. Cameron explains that the environment is imagined in real time rather than filmed by a drone. Odyssey trains world models on vast visual observations so they learn physics, dynamics, causality, and possible futures. In the demonstration, an open-source mapped fruit-fly brain was connected to actions inside the generated environment; the result was a “baby version” of a biological fly that chose to fly. Viewers initially mistook the simulated landscape for real footage.

The fruit-fly demonstration leads to a discussion of simulation theory, with Lon invoking The Matrix and the idea that a consciousness could have no objective way to distinguish virtual from physical reality. Cameron says he is 90% certain that humanity lives in an advanced simulation. Jason presents a paper proposing that reality may be an ASI-alignment simulation designed to see whether participants create superintelligence and keep it aligned. Lon, formerly skeptical because he could not see why anyone would run such a simulation, becomes more open to the idea because humans are now putting animal brains into simulated worlds to see what happens; he places his confidence at roughly 50–50.

Cameron says a world model does not simulate every person, fish, or location at maximum fidelity. It spends GPU capacity on the active agent’s path, leaving irrelevant areas dormant or low-detail; Odyssey can currently support about 20 simultaneous viewpoints and is working toward much larger populations. He compares the system to an attentive learner watching thousands of years of video and taking exhaustive notes on motion, gravity, fluid dynamics, humans, and cause and effect. Odyssey’s commercial purpose is not philosophical simulation but training robots and agents. Cameron says the model learns general concepts first, then can become a capable driver or manipulator from only 20–30 hours of task-specific experience.

Cameron says his eight years building driverless cars convinced him that the industry trained systems too narrowly by repeatedly showing them how to drive, rather than teaching a general understanding of the world first. That approach leaves vehicles vulnerable to unusual edge cases. Odyssey instead offers a foundation world model and lets robotics companies post-train it for their particular tasks. Customers include humanoid-robot, factory, self-driving, and drone companies, which deploy the resulting models in the field. Odyssey already generates revenue, has raised $370 million from investors including Amazon, AMD, NVIDIA, and GB, operates in the Bay Area, London, and Zurich, and is hiring at odyssey.systems.

The hosts discuss polling showing 66% of Americans oppose data centers anywhere in the United States, with opposition reaching 80% among people 30 and younger. Jason predicts AI will dominate politics through 2028 because workers may lose jobs without owning the equity created by AI companies. He proposes mass homebuilding, free trade schools, and dramatically cheaper baseline healthcare as ways to make people feel included. They criticize OpenAI’s proposed “AI compute campus” rebrand and Meta’s El Paso advertising unless benefits are tangible. Jason suggests publishing guaranteed tax contributions and visibly overfunding local schools, rather than relying on commercials about construction jobs that disappear after a facility is built.

Lon describes Andrej Karpathy’s HackerTube, which used Claude to extract 6,454 repeatedly recommended videos from 43 million Hacker News comments and present them without YouTube’s recommendation clutter. Jason built an “Unsponsored” Amazon Chrome extension that hides sponsored listings, banners, carousels, display ads, and brand promotions, arguing that everyone can now create custom interfaces. They then discuss Google’s first space-data-center test: four TPUs on a Planet Labs satellite ran Gemma 4, though only one chip operated at a time for about 20 minutes and produced a haiku. Jason says low Earth orbit could eventually provide vast solar-powered compute without land, water, permitting, or local opposition, especially once Starship’s 100,000–150,000-kilogram payload capacity makes launches economical.