Damek Davis, Wharton associate professor, asks if solving major math conjectures using ChatGPT prompts would secure a top-tier academic job

@damekdavis @aaswaminathan01 Can that student understand and explain the proofs to others? Can the student use the ideas from the proof to tackle other related problems?

These are important questions, for sure. But I think a student who can do those two things is not necessarily on par with a student who could have written those three papers by themselves (or even with coauthors).

There will be a period of time, could be many years, where hiring committees will encounter both kinds of researcher: the expert prompter and the old-school researcher. And it might be hard to distinguish between the two, especially if the prompters lack scruples.

@damekdavis If they came up with the problems, and the problems didn't exist before, then yes.

@mathandcobb @damekdavis @aaswaminathan01 matbe, but not necessarily better that chatgpt. What world would that be anyway? Paying graduate students to spread the word of the llm oracle to other grad students and professors? You must realise this is not sustainable.

@damekdavis Yes but the title is an assistant prompt professor obviously

@noesis89474724 @damekdavis @aaswaminathan01 I'm not saying they would get the job simply if they can explain what an LLM outputs! I'm only saying that if they *can't* digest and explain the output, then they are definitely not getting any job out there.

@kfountou @damekdavis but the nature of many fields is that people try to solve well-known open problem, and a problem is open if lots of people tried to solve it and failed.

@damekdavis my prediction is in coming years, we are going to have a track in the conferences where people submit a link to chatgpt and their interaction as a research paper.

@damekdavis No

@damekdavis if none of the other grad students in their application cohort were able to do this, then yes!

@damekdavis @aaswaminathan01 There is still a huge value in extracting new tools and ideas from generated proofs for human consumption and using them to make new advances or coming up with new interesting conjectures. Models still have little, if any, capabilities to ask interesting questions.

@mathandcobb @aaswaminathan01 @damekdavis Understanding and explaining a proof is several orders of magnitude easier than coming up with it.

@damekdavis No, I think ChatGPT gets the offer, i.e., the top tier math department buys ChatGPT Pro subscription for all the researchers in the department :)

@aaswaminathan01 @damekdavis There is no universe where an "expert prompter" would get a math prof job unless they can actually digest and explain these proofs to others. Even in that case, I assume a person who was able to prove those same results by themselves would be ahead.

@damekdavis Unsure. On one hand, the results are impressive and someone who is a pioneer with the technology could help identify use cases. On the other hand, how do you advise a student like that?

@damekdavis If they are able to explain the results successfully to the community, then yes.

@mathandcobb @damekdavis @aaswaminathan01 Well I think it's easier to explain a proof in retrospect which is a skill most prospective grad students would have, and the part on using it to tackle related problems might just loop back to the original question.

@damekdavis If they can prompt it to get the answer, and you can’t, then yes. Super intelligence is already here, but only as a synergistic collaboration of human and AI. Some people intuitively view problems through a different point of view, that enables AI to do more.

@damekdavis Read this until RACOON twit, that marks the end of the thread

ALL IS REAL, but you will ignore me

BECAUSE IT IS TOO MUCH AMAZING to be easily believable

@damekdavis If the graduate student is smart, he wouldn’t admit to its use. (But I wonder if the AI companies would respect user privacy in such a case).