Data compression works by removing redundancies, so it depends on what is considered to be "redundant" in the sum of human knowledge.

Lossless compression handles redundancies in a way that can be replicated exactly on decompression. Huffman coding for example works well with repeated strings of data by reducing them to a single figure (a "word" repeated n times and in so many places). Lossy compression on the other hand rerepresents data in a way that removes the stuff that isn't very important. In a jpeg image, high frequency information in the RGB color channels is truncated because humans suck at perceiving color compared to changes in brightness.

Data can't be losslessly compressed infinitely. If you have some algorithm that "compresses" an image down to say two bits, 01, and can faithfully reproduce the original image, then you haven't actually compressed anything. You just moved the image information from being in the image container to being in the algorithm. (In this extreme case, something like a lookup table.)

If the sum of human knowledge is one great big text file, then with typical tools you can expect a 4:1 compression ratio.

Compressed files have little redundancies left to enable further compression. They look a lot like binary noise, and further compression needs to be lossy (not really possible with plain text, unless you change/reduce the source content).

Data compression works by removing redundancies, so it depends on what is considered to be "redundant" in the sum of human knowledge.

Lossless compression handles redundancies in a way that allows them to be replicated exactly on decompression. Huffman coding for example works well with repeated strings of data by reducing them to a single figure (a "word" repeated n times and in so many places). Lossy compression on the other hand rerepresents data in a way that removes the stuff that isn't very important. In a jpeg image, high frequency information in the RGB color channels is truncated because humans suck at perceiving color compared to changes in brightness.

Data can't be losslessly compressed infinitely. If you have some algorithm that "compresses" an image down to say two bits, 01, and can faithfully reproduce the original image, then you haven't actually compressed anything. You just moved the information about the image from being in the image container to being in the algorithm. (In this extreme case, something like a lookup table.)

If the sum of human knowledge is one great big text file, then with the best tools you can expect a 4:1 compression ratio.

Compressed files have little redundancies left to enable further compression. They look a lot like binary noise, and any further compression needs to be lossy. Not really possible with plain text unless you reduce the source content (maybe lots of abbreviation, using simplified wording, etc.).