Users appreciate the researcher's push for AI explanation tools to target discrepancies because it simplifies defining evaluable subtasks for career proposals and provides a key insight for building real-world tools.
No Digg Deeper questions have been answered for this story yet.
@lastland0 For example, suppose an AI tool generates a specification for a program I have written, and I don't know whether my program is correct to begin with, nor whether the specification is what I want. It should not explain its specification directly
@lastland0 Rather, it should explain any discrepancies between that specification, the program, and any sources of information about my intent (natural language interaction with the tool, comments, test cases, and so on). That is the information I need to decide which of those is wrong
@lastland0 This actually makes my CAREER proposal work much easier, insofar as I can define separate subtasks that are easy to evaluate, but I also understand how to unify them with an explanation component that highlights discrepancies between what various subtasks would produce
@lastland0 But also I think it's the key insight needed to actually build tools for this IRL