Theoretically, you can’t benchmaxx ARC-AGI, but I too am suspect of such a large improvement, especially since the improvement on other benchmarks is not of the same order.
It's a sort of arbitrary pattern matching thing that can't be trained on in the sense that the MMLU can be, but you can definitely generate billions of examples of this kind of task and train on it, and it will not make the model better on any other task. So in that sense, it absolutely can be.
I think it's been harder to solve because it's a visual puzzle, and we know how well today's vision encoders actually work https://arxiv.org/html/2407.06581v1
The real question is: Why are people designing benchmarks that, if a model is trained on them, it won't improve the performance of the model at any real-world tasks? Why would anyone care about such benchmarks?
Benchmark maxing could be interpreted as benchmarks actually being a design framework? I'm sure there are pitfalls to this, but it's not necessarily bad either.
He said in an interview that it doesn't count if it's explicitly targeted, only if a model generalizes to it.
He also said that the "real test of intelligence" is being unable to come up with new tests that a human can easily do that the AI can't, not in being able to pass any specific benchmark.