That's a great story.

I've been intending to add Timefold to our list of optimization modelling resources. Your post prompted me to do so: www.solvermax.com/resources/li...

Absolutely. A MILP is a very impractical way of compiling a crossword. The goal was to see if it would work, which it mostly does. The contrast with 1989 is also remarkable.

Just for fun, I created a Mixed Integer Linear Program for compiling crosswords www.solvermax.com/blog/crosswo...

While solving MIPs using AI seems like a pointless exercise, I did use an AI to find the paper you're referring to. So, AIs are good for some things.

I've found ChatGPT & Claude somewhat better at programming, but much the same as Copilot for optimization model code.
Perhaps an AI trained specifically on models would be better. @gurobioptimization.bsky.social have started looking at AI, so we'll see what they find www.gurobi.com/resources/br...

The process was a roller coaster. I was impressed by the AI's astuteness but frustrated by its strange mistakes.
I'm increasingly using AI for coding tasks - either straightforward yet time-consuming things, or to get syntax right for unfamiliar things. It mostly performs well in these situations.

Though I object to paying 39,95 € just to look at an article. But that's a different issue.

Perhaps "often" is a bit harsh. But I'd have better evidence if the code is published too.
I hope you're right about the culture changing.

I suppose a formulation is better than nothing. But published formulations are often wrong.
For example in www.solvermax.com/blog/green-m... we get a better optimal solution.
In another case, our solution was worse than in the paper. In discussion with the author it emerged that the paper is wrong.🤷

The Gomes, et al paper we summarize lists many reasons why academics may not publish code and data - along with mitigations that should encourage publication.
I approach this issue from a selfish perspective. When looking to the literature for help with a problem, a formulation is seldom sufficient.