I find the MMIX code less relevant to my line of work now as getting down to machine-level optimization requires a whole new different suite of tools. The ensuing optimizations on modern CPUs seem deviated from what the MMIX code can help.
Knuth's a good, engaging writer, but TAOCP's content and typography are definitely better thought out than the prose.
EDIT: Just remembered a BBC interview with a philosophy professor about Kant. Apparently Kant is criticized for being really verbose. The professor's retort was he sensed Kant had so much to say and so little time to say it that he didn't edit very carefully, giving it a similar kind of bloated, meandering quality. Even so, Kant is held up as one of the GOATs, because in the end it's the content that counts.
And maybe even more importantly, a great collection of problems to work through. I think it is actually an underappreciated text nowadays.
Meandering through topics is good, but he meanders through style as well: sometimes he was in a mathematical mood, sometimes he was feeling more practical. I think it's better to intuit first then formalize, Knuth likes to do these in groups of pairs rather than a pair of groups.
Knuth is not the worst at this, but the lost potential is the greatest in his writing.
It’s a good short story if you’ve got 20min.
My issue with the books is that they're actually quite long winded even by what you'd expect from the tone.
There's some really cool stuff in them, obviously, but I think they're objectively not very good textbooks for any purpose.
Then again I'm coming from a background of physics rather than mathematics so I'm not set out for a real battle of wits when it comes to constructing proofs.