Comment by daxfohl
3 months ago
I wouldn't think it would be good for coding assistants, or things where character precision is important.
OTOH maybe the information implied by syntax coloring could make syntax patterns easier to recognize and internalize? Once internalized, perhaps it'd retain and use that syntax understanding on plaintext too if you fine tune it by gradually removing the color coding. Similar approaches have worked for improving their innate (no "thinking", no tool use) arithmetic accuracy.
It might be helpful for intuiting the structure of a program. Imagine if you had to read code all on a single line, with newlines represented with \n.
I can get the feel of a piece of code just by looking at it. Even if you blurred the image, just the shape of the lines of code conveys a lot of information.
True, but LLMs are already really good at that kind of thing. Even back in 2015, before transformers, here's a karpathy blog post showing how you could find specific neurons that tracked things like indent position, approx column location, long quotes, etc.
https://karpathy.github.io/2015/05/21/rnn-effectiveness/
That said, I do think algorithms and system designs are very visual. It's way harder to explain heaps and merge sorts and such from just text and code. Granted, it's 2025 now and modern LLMs seem to have internalized those types of concepts ~perfectly for a while now, so IDK if there's much to gain by changing approaches at that level anymore.
Another example might be the way people used to show off their Wordle scores on Twitter when the game first came out. Just posting the gray, green and yellow squares by themselves, sans text, communicates a surprising amount of information about the player's guesses.