[nathan_compton]

I worry about this. I happen to have done quite a lot of programming in Lisp dialects over the last decade or so but since adopting Gpt4 I tend to just code in Python because that is what the model understands best. It does seem like AI will enhance network effects by increasing the efficiency difference between technologies that the AI knows and those that it doesn't. Kind of depressing.

[kevindamm]

I wonder how much of this is syntactic familiarity (from training) and how much of this is needing to attend to balanced parentheses.

I don't use lisp often enough to have played with getting GPT to lisp with me, but I have played a bit with getting it to read and write Datalog (which I suspect is even more scarce in The Pile dumps). It's ok at recognition but misses details. I haven't seen it produce much of value yet. But it can write JavaScript for days, and has no problem balancing parentheses and brackets there, even without compiler/tree-sitter support.

If I had spare experimenting bandwidth I would look into whether fine-tuning for Lisp format and conventions would show a significant boost in performance..

[nathan_compton]

I find its Lisp generation to not just be syntactically wrong but conceptually inadequate. It's inability to generalize across languages is one of the big reasons I'm skeptical about language models' general intelligence.

[wanderingbort]

This seems more of a concern for foundational models rather than personalization.

Any pressure you feel to adopt python is not because it has detected you enjoy python, it’s because it’s global training data skewed to python.

Its a huge concern but, not this article’s concern I think.