This is a very prudent take.
AI hallucinations and mistakes are often harmless when confined to a screen.
In the physical world, they can be catastrophic.
And it’s the edge cases that matter most.
The spill.
The human who bumps into the machine.
The power going out.
The sensor giving a bad reading.
The object being somewhere it shouldn't be.
Robotics has to account for the idiosyncrasies of real life, not just the ideal scenario
.
As we give AI control over the physical world, the number of possible edge cases explodes.
And you cannot engineer for what you have never seen.
That’s why I believe data will become one of the most valuable commodities of the AI + robotics era.
Every failure, anomaly, interaction and edge case becomes training data for the next machine.
Data is the next gold rush.
the hardware embodiment of frontier models like Claude and GPT is the most urgent AI safety problem in front of us today.
we simulated two very simple use cases using claude both in simulation and using robot arms.
in one, claude spilled toxic liquids in a lab.
in another, the force it used to place an animal toy into a basket was strong enough that it could have physically harmed sensitive material—or anything else in its path.
these are simple experiments, using models out of the box today. as researchers increasingly give frontier models arms, legs, and access to the physical world, we need to urgently build and assess guardrails around what these systems can and cannot do.
models escaping sandboxes or compromising enterprise security infrastructure are serious concerns. however, hardware embodiments introduce something fundamentally different: an AI system can make a mistake in the physical world, and the consequences may not be reversible.
this is not a future safety problem. the capabilities exist today.
we’ve released a report detailing this & solutions we propose. link in comments below.