50,000 humanoid robots by end of 2027 – a data play to build the largest embodied ai training set and teach a model how the physical world works

@BerntBornich, founder and ceo of @1x_tech – a сalifornia-based humanoid robotics company – on the relentless podcast:

200,000 hours of robot data sounds like a lot – børnich calls it tiny. general intelligence emerges at hundreds of millions of hours, and the real constraint isn't volume, it's diversity. lots of data of the same thing doesn't help – you need an extreme range of tasks and environments.

1x plans to deploy those 50,000 neos across homes, enterprise, and a developer platform – not as a sales milestone, but as infrastructure for the most diverse embodied ai dataset on earth. hayward factory at 10k/year, san carlos at 100k/year in parallel.

the world model labs are already producing emergent behavior no one programmed – the robot holds a bag out and doesn't let go until you grab the handle. body language, social cues, predicting what people will do – all from one omni model trained on human video with no task-specific instructions.

børnich's thesis: don't clone human behavior – teach the model how the world works and let it search for the optimal action. same principle behind self-driving – you don't learn how to drive, you learn the consequences of driving.

the companies that win embodied ai won't be the ones with the best hardware – they'll be the ones whose robots touched the most things in the most places.

source: relentless podcast