Verifying sources is a recursive problem - where do you stop? Humans have intuitive feel for it, but agents don’t or at least not yet (I wonder if intuition is just a secondary neural net which is currently being added to the agents as we speak).
Also as a human you are able to examine agents erroneous trajectory, real or imaginary, without contaminating your own. Agent have a problem with that - as soon as someone else’s thought is in the context it can lose track of provenance and veracity. Sometimes I think we need a bloom filter to retroactively assign “dirty” flag to invalidated or questionable token spans already in the context.
Arguably, intuition is primary neural net, the only thing an LLM has without CoT, it just is spiky so humans only notice where it's below-average and just dismiss the rest as normal. Of course it's going to lag behind in some areas compared to others.
I don’t mean that to say AGI is here or easy or necessarily that close but it’s likely going to feel like one thing after another until one day most of these things that make you think “how could something so capable be that dumb” are largely solved.
After I gave up (mostly because ChatGPT had given me incomplete information needed to solve it) I checked the source of the dispute. It was a site very similar to this one and someone had an AI agent working on the same problem, publishing dozens or hundreds of pages of notes. The agent found the solution page and concluded it was wrong because many of the 32 source passages supposedly didn’t contain enough text.
I dug up the PDF of the book and found the mistake - whenever a passage continued onto the next page, the agent wasn’t including that continuation. The passages weren’t actually too short.
Annoying that ChatGPT can cite sources like this without being able to properly weigh their reliability.