Ben Lorica changes his mind on world models: if AI builds the robot, where's the value?
bigdata · x · 2026-10-27
Ben Lorica revises his earlier stance on robot world models. The pointed question: if coding agents can assemble robot programs from reusable skills and even help build the simulators to test them, where does a learned world model add something a foundation model, existing robotics software, and a physics engine can't?
The engineering bottleneck: Much "variational automation" in warehouses, factories, and food processing stays expensive—robots lack the physical experience LLMs have in text, and each application needs engineers to assemble perception, grasping, motion planning, and recovery. Simulation cuts physical testing, but building a simulator that captures what determines success is itself labor-intensive, and the commercial bar (fast and reliable, not cherry-picked demos) is high.
Four ways to build a robot: the traditional composable stack, VLA models, and routes where AI automates robot engineering itself—leaving world-model companies facing a sharp challenge to their value proposition.
More from Embodied
- Robot learning startup Machenta exits stealth with $40M seed co-led by Creandum and USV — ZeeshanZiaML · 2026-10-27
- Machenta raises $40M to teach robots from a single demonstration — Ed__Johns · 2026-10-27
- Walden Robotics teaches humanoids to call for help, cutting uncaught failures 10x — CyberRobooo · 2026-10-27
- Robot ducks face off in MicroDuck Arena's new Tug of War cup, live Nov 6 — huggingface · 2026-10-27
- Apple's 14-inch M6 MacBook Pro: dual 16-core Neural Engine, 1.9x M5 performance — whurley · 2026-10-27
- Luxonis launches OAK 4 FFC: five-camera modular vision board, $849 dev kit — Scobleizer · 2026-10-27