AI vendors don't lock you in, they lock you out: value piles up in verified harness data

pablooliva · reddit · 2026-09-05

Synthesizing Karpathy's point that LLMs automate only what you can verify with Reyes's argument that value settles in the harness rather than the model, the author argues verification turns usage into an asset — and the harness is where it accumulates. Drawing on a concrete vendor workshop experience, he urges anyone building or buying agent systems to ask early: who ends up holding the verified pairs? That determines whether vendors are moating themselves with your usage data.

Original post →

More from AGI Musings

AGI Musings channel →