Invited Talk - Brian Zhan: The RL Environment Land Grab: What's Worth Backing and What Won't Work

Frontier labs are budgeting billions for RL environments, and a crowded field of startups is racing to supply them. All of these startups are selling human feedback in a simulator that comes with human limitations (humans form opinions, not measurements, and environments they create are gameable by construction). The durable signal comes from the world itself. Many startups are betting on exactly this, like Periodic Labs on autonomous materials discovery with experimentally verified rewards, Skild on robotics learning from real-world feedback.