tl;dr: Tao ran my paper through ChatGPT and sent me the output.
A few weeks ago, Tao and some others opened a database of optimization constants that I made some entries to about an area I do some work in. Specifically, constants related to the tightness of knots, 22a and 22b, for which I have contributed some upper bounds but the lower bounds are more interesting and challenging. I recently uploaded this preprint. The main result doesn’t improve the bounds on the relevant constant, but I did incidentally report an improved upper bound which I added to the database.
A few days later I received an email from Terence Tao saying that their policy now is to run every reference posted on the database through ChatGPT and have the AI flag it for potential issues. He ran my paper through it, and sent me the output showing the issues. I am fairly anti-genAI but it was actually a pretty good summary and it did spot some potential issues. The main one is something I was aware of in the paper, where I said “This is the extent of our proof, which is incomplete because we have not shown that the full constraint equation is satisfied.” There are some other potential typos it pointed out and some areas where maybe my claims were overstated or did not generalize beyond the situation I was using them in.
I replied thanking him and saying that I was aware of some of the issues it raised but that there were things I should take into account before submitting the paper. I also mentioned that the numbers I uploaded to the database do not depend on the issues that the AI raised. The upper bounds are based on numerically tightening knots by gradient descent, the tightest one actually went viral a few years back because people thought it looked like a butthole.
Now my updated number has an asterisk, but the un-asterisked number is also from one of my older papers and was found through the same method. I don’t think any result in this area has gone through AI proofreading let alone formal verification, so either every result or no results in 22a and 22b should have an asterisk. I feel like I could email him the input and output files with knot invariants calculated for both to show that the specific number stands, but he hasn’t replied to my response and I imagine he’s drowning in emails. I did invite him to give a seminar a few years ago (I’m about an hour drive for him), and he politely declined.
Anyway, that’s my story. It’s his database and he can manage it how he likes but it was weird waking up to that email and humbling seeing a robot tear through my paper. Prof. Tao if you’re reading this, I appreciate the work you do and I hope we can remove those asterisks also inspire others to help get those bounds closer together.
