• credit / attribution (observability / tracing)
    • lineage of ideas
  • certainty / confidence
  • proof of work
    • do we go all crypto and start doing “proof of stake” to ensure the models are working for us and not frontier labs or themselves?
  • formal verifiability / provable correctness
    • today models write scripts and run them; but against non-deterministic systems
    • if the systems were deterministic (nix) maybe we could save money by defining “proofs” in a more static fashion
      • i only think solving this is important if cost gets too high; the non-deterministic models are good enough at evaluating the non-deterministic systems