This is hilarious. In compression contest circles, there's long been this concept of "AI compression" (long before the current LLM wave), considered the upper

This is hilarious. In compression contest circles, there's long been this concept of "AI compression" (long before the current LLM wave), considered the upper limit of compression algorithms. This is based on the idea of the Dictionary... IIRC, some compression systems include the Dictionary as part of the decompressor, to save space in the payload, but then that Dictionary is fixed and may be suboptimal for the data. Others include it in the payload, so that it's optimal for the data, which most do. But you could achieve incredible compression of, say, The Complete Works Of Shakespeare if you knew your peer knew about The Complete Works Of Shakespeare (had a rich enough dictionary) and just send them that designator of the payload! Or, you could send them a particular high-resolution picture of a wheat field with cypresses under a blue sky with an appropriately specific "description". What Fabrice did is gone ahead and put this theoretical idea in practice. And that's hilarious.