News for hackers. Stuff that matters.

A comment on The Lisp curse, now coming to your ecosystem. Parent comment

Re:The Lisp curse, now coming to your ecosystem by kragen on Thursday November 05, 2026 @03:05PM (#50383771)
Which models have you found most take advantage of this kind of capability? Opus 5.5 and GPT-6 Sol and Astra seem to like to minimize their tool calls, blatting out a single edit with dozens of lines of code in it rather than editing a few at a time. And they virtually never use things like the Python REPL, even though they use Python all the time.

It seems sort of wasteful of tokens!

[ Parent | on HN ]
Re:The Lisp curse, now coming to your ecosystem by iLemming on Thursday November 05, 2026 @03:29PM (#50384058)
[delayed]
[ Parent | on HN ]
Re:The Lisp curse, now coming to your ecosystem by kragen on Thursday November 05, 2026 @03:59PM (#50384388)
> It seems sort of wasteful of tokens!

This was referring to the LLMs' tendency to blat out hundreds of tokens into a single tool call, not to Clojure.

[ Parent | on HN ]
Re:The Lisp curse, now coming to your ecosystem by iLemming on Thursday November 05, 2026 @04:28PM (#50384698)
I don't think I'm missing the point. LLMs don't use Python's REPL because it's not that efficient, while properly built harness will eagerly utilize Clojure REPL if given a chance. Without "blating out hundreds of tokens".
[ Parent | on HN ]
Re:The Lisp curse, now coming to your ecosystem by kragen on Thursday November 05, 2026 @04:47PM (#50384915)
You replied to my remark about the token efficiency of a behavior I've observed LLMs
[ Parent | on HN ]
Re:The Lisp curse, now coming to your ecosystem by baq on Thursday November 05, 2026 @05:30PM (#50385421)
in lisp the repl is the primary mode of development, you replace functions one at a time and only reload the whole system if you break structure layouts, and even then you don’t really restart the whole process (most of the time anyway, sbcl’s immobile code space is a bit of a pain) - agents can and do things by themselves, unprompted, which look like magic even if you’ve been piloting llms for as long as anyone here. I use gpt 6.1 sol for this, I’m sure opus would handle it perfectly, too.
[ Parent | on HN ]