A fully NN solution (as opposed to using a light sprinkling like all competitive solutions do) requires big binaries for the most part, while an arbitrary program can easily memorize exact sequences to find repetition while using a tiny NN for some finetuning of predictions. A pure NN solution like a Transformer-XL
does turn in the record-setting performance on natural language datasets including WP... unfortunately, the required NN model sizes alone tend to be larger than the entire uncompressed WP corpus here, and so have zero chance of ever being the minimum of model+compression. (An observation which I think supports the idea that the Hutter Prize has long outlived its intended use for measuring progress towards AGI; it's now just sort of a 'demo scene' version of AI benchmarking, testing intelligence within extreme constraints of sample efficiency/compute, rather than a true useful benchmark.)
The fact that small compressors like gzip or zpaq do so well at small data/compute but then can't compete as you scale up to tens or hundreds of gigabytes of text (amortizing the cost of those fancy NN models) can be considered a version of the Bitter Lesson: http://www.incompleteideas.net/IncIdeas/BitterLesson.html