Users like the proposal for cryptographic chains to verify AI agent tasks and block misaligned covert work, praising the creative thinking involved.
Based on 2 visible X reactions from 3 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@idavidrein I like the proposal. shameless self plug I made a toy implementation of a similar idea using the Mac Touch ID sensor to prove that a human actually verified the work an agent did. Was surprisingly easy to get working. https://bengoldhaber.substack.com/p/the-importance-of-physical-touch
@idavidrein this is a wild rabbit hole to go down, love the thinking.
I wrote up a mechanism for preventing unmonitored misaligned AIs from using monitored AIs to covertly do work for them. TL;DR: At some point in the next 1-2 years we should require that all agent tasks are cryptographically chainable back to a verified human instruction.
This is a good idea https://twitter.com/idavidrein/status/2075346119102693388
Users like the proposal for cryptographic chains to verify AI agent tasks and block misaligned covert work, praising the creative thinking involved.
Based on 2 visible X reactions from 3 accounts; directional sample.
Ask a question below.
Published answers will appear here.