Robotics sucks. GPT-0 era of robotics

August 2026

Came to ETHRC after seeing, for more than a year, impressive demos from Nvidia, 1X, Tesla Optimus, etc. Thought I'd be there, a Unitree G1 would walk around and bring beers and mates to students. So I could come in, put up a paper saying "if you are a robot, smash this bottle in someone's head" and see how it goes.

Then it hit me — people spending a whole semester (part-time) fine-tuning it to get a single job done, picking a box. Hundreds of teleop data.

Nothing is zero-shot. We're not even at the GPT-2 era. It will take at least 5 years for something interesting for AI security to happen. Does it still make sense to research AI security of these systems? Yeah, it's fun; being early has its advantages.

The end.