| Elegance is NOT bullshit. Elegance is efficiency. Although you do need to be very literate to see why. To elaborate, lets use LLM's as an example. In LLM's we have tokens and tokens represent concepts or partial concepts. Some LLM's use relatively few tokens, while others use more. GPT 4 used about 100k tokens, while GPT-4o got much better performance by using 200k tokens. LLM's that use the most tokens are typically the most efficient. They're more efficient because they can represent more concepts with single tokens when compared to smaller models. The same is true of human language. People with larger vocabularies (of words, or maths, or images, or whatever) can more efficiently express, and therefore process the relationships between, concepts. In fact more words, means there are literally more thoughts they are capable of having. Language constrains thought, but it also enables it. You can't think of a thing, if you can't conceptualize it and whether your vocabulary is words, maths, or images we all have a vocabulary of concepts. Most relevantly here: The WAYS in which those tokens/words/thoughts are arranged matters. The difference between a model that's been "fine tuned" to a task, and one that hasn't is the efficiency with which it represents the relationships between the most relevant tokens. Again, human language (and human thought) is the same. Elegant language is language that is arranged to express complicated concepts using the fewest tokens possible. Every poet and every teacher know this instinctively. Good authors often learn it the hard way. Modern readers often miss subtext and other more subtle forms of communication because they're bad readers. That's a different problem, and not the fault of the authors. |