Nobody holds your whole key, not even OISY.
It's split into pieces across a network. When you approve a transaction, the pieces sign together, so there's no single place to take it from.
@DanielGPT2022@fchollet I think a model will be always a model. Something that emulates, idealizes, describes what the underlying mechanism could be but it would never be exactly the same to the real one. And what a model miss it’s not always accessible by us, limited humans
@SnassyIcp@josephfounder yes,I agree the alignment problem has to be solved upstream. I was wondering if ICP can have a moat vs other chains in building a decentralized control/execution plane that cryptographically enforces what the agent is actually authorized to do while the AI agent remains untrusted
@SnassyIcp@josephfounder Supposing inference on chain won't be possible/valuable enough for large models, I was thinking about a potential external (but verifiable) inference layer + execution/control layer on ICP. Does ICP have a real moat in this case? I'd appreciate your opinion on this, thanks anyway