Perspectives

“Most published capability gains cannot currently be distinguished from contamination, and the field has chosen not to find out.”

Until evaluations are built to survive models that may have seen them, capability claims are unfalsifiable — and the field has not paid the cost of fixing this.

“Reliability, not capability, is the binding constraint on autonomy — and it is not on the trajectory the capability curves suggest.”

Four years of shipping agents into environments where mistakes cost money. What actually breaks is never what the benchmarks measure.

“Labour markets have absorbed technological shocks before, but always over a generation. The adjustment speed, not the endpoint, is what is different here.”

The debate is about the eventual equilibrium. What determines whether this is survivable is the speed of the transition, and almost nobody is modelling it.

“Embodiment is not behind because it is under-invested. It is behind because physical interaction data cannot be scraped, and that does not change.”

Digital capability curves have made the field systematically overconfident about robotics. The data constraint is structural, and it does not respond to capital.

“Interpretability coverage is nowhere near what an oversight decision requires, and deployment is proceeding on the assumption that it is.”

Whether human review can be removed is an empirical question. It is being answered by cost curves instead.