The TEE + LLM combination is clever becuz it addresses both the trust problem and the flexibility problem simultaneoulsy. Traditional MPC is too rigid for messy real-world negotiations, but you're right that LLM bias verification is gonna be the hard part. The idea of using simpler models for verifiability vs state-of-the-art capability is an interesting tradeoff most people wouldnt think of. Also the point about failed negotiations not leaking info is underrated, thats often the main reason parties avoid trying to coordinate in the first place.
*Maybe* a state-of-the-art model could be used as well, if the involved parties are given access to test it properly for bias first. But that would probably require mechanistic interpretability advancing a bit more first.
The TEE + LLM combination is clever becuz it addresses both the trust problem and the flexibility problem simultaneoulsy. Traditional MPC is too rigid for messy real-world negotiations, but you're right that LLM bias verification is gonna be the hard part. The idea of using simpler models for verifiability vs state-of-the-art capability is an interesting tradeoff most people wouldnt think of. Also the point about failed negotiations not leaking info is underrated, thats often the main reason parties avoid trying to coordinate in the first place.
Thanks.
*Maybe* a state-of-the-art model could be used as well, if the involved parties are given access to test it properly for bias first. But that would probably require mechanistic interpretability advancing a bit more first.