Ben Lorica Changes His Mind on World Models as Agents Start Writing Robot Code
bigdata · x · 2026-10-28
Ben Lorica (Gradient Flow) writes that he has "changed his mind about world models." The central question: when an agent can write the robot code — and even the simulator to check it — where does a learned world model add value over existing software stacks?
He identifies a key bottleneck in "variational automation": the huge class of robotics jobs between rigid factory automation and general-purpose robots — packages arriving in different positions, groceries of varying shapes, cables bending differently — common in warehouses, factories, and food processing, yet expensive to automate. Bottlenecks include the lack of physical experience data compared to language models, heavy per-application engineering across perception, grasping, motion planning, and recovery, the labor of building physically faithful simulators, and a high commercial bar (fast and reliable, not just cherry-picked demos).
The article outlines four ways to build a robot, from traditional component stacks to VLAs, and how automated robot engineering reshapes that landscape.
Related event: Ben Lorica Reconsiders World Models as AI Agents Code Robots(3 posts)→
More from Embodied
- NVIDIA's MotionBricks generates character motion at 15,000 FPS with 2ms latency on an RTX 5090 — kimmonismus · 2026-10-28
- Machenta robot reacts and recovers in real time, trained 100% in simulation — Ed__Johns · 2026-10-28
- DeFM Depth-First Foundation Model: Why Anchor to Depth, Not Estimator Features — chris_j_paxton · 2026-10-28
- Neocambrian's from-scratch NeoSLAM tops LaMAria Long category among mono+IMU methods — AashaySachdeva · 2026-10-28
- Robotics long tail is too long: semi-humanoids will beat task-specific robots, says Paxton — chris_j_paxton · 2026-10-28
- XPeng's IRON solid-state battery enters small-batch production: 30% lighter, 30% more capacity — CyberRobooo · 2026-10-28