Code golf is anecdotal, at least, evidence that proving lower bounds on the Kolgomorov complexity of a particular datum requires intelligence.
One of the best contenders is the open source cmix which, among other tricks, uses LSTM neural networks to predict the next character/bit: http://www.byronknoll.com/cmix.html
I wonder what the memory/speed/efficiency tradeoffs looks like for that kind of design.
Not just any chunk though. It (presumably) contains a lot of structure that encodes information about [English representations of] the real world.
It's not random, nor is it simple, nor is it an image of a fractal (maybe?) etc.
A system that is able to compress it well makes good predictions about what the data contain which indicates "intelligence" about the structure of the (encoded) world.
Unless the idea is that whatever methodology is required to build this particular program could then be adapted to do other things that seem intelligent. That seems possible, but not at all likely given my reading of the challenge requirements.
> But would you expect that program to itself be useful for anything else?
That exact program, no, but the process of creating that program would be educational, both the compressor that makes it and the process of making the compressor.
Another way of putting it is in terms of the "Twenty Questions" game. Which pre-selected twenty questions allows you to "cover" the most of the Universe of phenomenon? If one set of 20 allows you to distinguish more of the Universe than some other set, then the first set is somehow a better model of the Universe. Does that help?
Is Akinator’s 20q database available as a dataset for non-commercial research purposes? It’d be beat to attempt to compress it in a way where training with one half predicts the other half at the smallest trained model size.