Reacción a "An AI system could win a medal at the International Mathematical Olympiad, according to a study"

The general impression is that it's an article with a certain degree of scientific rigor, which will likely lead many people outside of Computer Engineering to remain at the surface level, that is, at the applications. Regarding its quality, I think it's a commendable piece of work given the current state of AI, where you either need virtually unlimited computing resources or you won't be able to develop a prototype from start to finish. And this is precisely what happened to this group of researchers, who used pre-trained models and subsequently developed an adaptation layer for their specific problem. Although it might sound somewhat fraudulent, if this is clarified—as the authors honestly do—there's no problem at all. The scientific literature on this subject is full of this type of work with this approach.

Regarding the implications, I think this will force examiners to be more imaginative when designing problems. Let me explain: the law of least effort leads us to set exercises based on previous ones with only very slight variations. And this is where AI comes in, leveraging a vast database of exercises and, by making small modifications, arriving at solutions that might be acceptable, even surprising at first glance. However, if the solution to a proposed exercise doesn't resemble anything in its database at all, or if there is a resemblance but not quite enough, the generative AI is very likely to fail.

Leaving aside the sensationalism surrounding the Mathematical Olympiad, I believe this is what's really behind this work, which I could summarize as acceptable, much like much of the generative AI being offered commercially in very specific fields such as law, insurance, medical reports, etc.