@ShriramKMurthi One thing implicit here is “abstraction as leverage”.

Good abstractions (PLs) save tokens in analogous ways to saving human thinking: they summarize complexity and focus a concept on a domain.

This is the “make more good things happen” dual of properties “stopping bad things”.

@ShriramKMurthi hmm... issues which come to mind include proper vs improper modularity (for example AI may tend towards data replication where the intent needed control over a unique instance) and code maintenance (handling issues which are only discovered in the field by users)

@ShriramKMurthi Lovely write-up! An abstract/tldr would be helpful.

@JoePolitz Right. That is another goal of the DSLs, but it felt too low-level, and also I don't understand the tech *that* well.

I ran an experiment over the weekend where I got Claude to "rename" both the keywords and variables of programs to match the tokens of a local LLM: ↵

This is excellent. I was actually thinking about this earlier since my PL class was amazing but there has obviously been much development in the way people write and interact with programs. Something I’ve wondered is if agent plans can be written in more verifiably defined language. The main advantage I see here is to exercise fail-closed semantics when running an agent with an approved plan (agent has defined scopes per step and cannot break that hard boundary)

@JoePolitz We picked names and notations for humans, but why bother? Just choose from available tokens! And w/ #lang we can just build out that very language. I did this for some shplait code. The keywords were already okay, the sigils weren't, and var names needed some renaming. ↵

@ShriramKMurthi And, of course, AI tends to flourish in tightly constrained contexts (especially when this can be incorporated into their training).

@JoePolitz But at the end of the day, I think we got only like a 3–6% saving. Hardly worth the loss of shared context, meaningful names, extra tooling, lack of training, etc. Maybe makes sense if you're tokenmaxxing with 1000 claws on worktrees, but that ain't me, bud.

@bhaskark_la Thanks! I gave myself a 6 page limit, and now the abstract would push it over 6. (-:

@ashishjayamohan Yep, indeed. I'm not sure "planning" is the right place for that, since I view plans are inherently a bit squishy and natural-language-y. But as you translate down from those plans, yeah, you want to shift into more and more tightening up.

@raudelmil Indeed, but I view those as SE questions.

@raudelmil The "tightly constrained contexts" is the whole point, no?