Anthropic: Progress on provable inference for large language models

anthropic.com

11 pointsTepix3 comments

Not my area of expertise, can someone please explain why they can’t just digitally sign the generated tokens?

If you trust the token generator and just worry about tampering in transit, signing the tokens after the fact is a fine solution.

This work is about the case where you trust nobody and therefore require even the token generator to do extra bookkeeping to prove that it's not taking any shortcuts.

The scenario is that I access Fable via openrouter and I don't want to be rerouted to an inferior model. If the vendor (Anthropic) sends a signature at the end of each response, that would solve the issue, wouldn't it?

Reply on news.ycombinator.com