I think whats confusing me is that they think AI is the solution, so that would tend towards a huge very clever compressor program (think something like GPT3). But presumably the competition also has limits on the size of the compressor - because you want to avoid trivial solutions like embedding the 1GB file in the compressor. So how do they manage this contradiction between wanting the compressor to be very clever while also limiting its size?
Right, but then they're not going to get a very clever compressor then. The ideal AI compressor would encompass all human knowledge.
It just seems like they've hit upon a good idea - AI/compression crossover - then shot themselves in the foot by excluding anything genuinely intelligent. I suppose the problem is how to tune the rules to allow AI but not cheating.
Those all all vague, hand-wavey concepts; open to disagreement, and in some cases might turn out not to exist or make sense. Entire research careers have been spent trying to even define these terms, let alone implement them.
Focusing on compression of a particular chunk of Wikipedia eliminates all of that, and gives us a precisely measurable quantity.
Is it a perfect defininition of intelligence? No; it was never meant to be.
Is it a runnable, measurable, comparable experiment? Yes.
The ideal AI compressor would encompass all human knowledge.
That's what they're doing (in the constraints of defining all human knowledge ~ English Wikipedia). More refined models of representing all of our knowledge will beat weaker ones in this test.
Whether this is AI or intelligence or not depends on how you define those terms - but to win this competition your compressor has to be able to anticipate the rest of the human knowledge fairly accurately based on having seen part of it.
