AI Assistance Reduces Persistence and Hurts Independent Performance (researchers at Carnegie Mellon, Oxford, MIT, UCLA)

[Lucky Luciano]

4. I roughly agree with what those users are saying.

Lately it has been very good for me compared to some past uses where it kept getting some things wrong (consistently at least 😂 .

I don’t think it can do true innovation yet like the Erdos problem solution you mentioned, but I won’t be shocked if it gets there in the next couple years. What Terence Tao was using it for though - a peer reviewing sidekick….definitely useful.

I don’t use it for image creation or anything artistic that might include the written word. I do sympathize with some of the copyright related stuff…it’s blurry as hell for how to attribute things for its responses…wouldn’t even know where to begin. An impractical thing would be to retrain the whole thing while leaving out some texts to compare to the fully trained set and seeing the difference in output…combinatorially absurd and utterly impossible - sort of like something called Shapley Attribution in trying to attribute trading strategies blended together in a portfolio. At that much lower scale (trading), Shapley attribution is also somewhat infeasible.

I strongly believe people still need to learn all of the raw academic skills needed so critical thinking is not destroyed. Only with critical thinking can one know when the LLM is hallucinating. This is the big challenge of the day in education…most people will opt to take a helicopter (a kind that doesn’t exist of course) to the top of Mount Everest rather than engage in the actual climb. That needs a solution. I have some ideas around this, but I also hate typing on iPhones…so I’ll stop.