I’m not a programming language expert, but watching GenAI models stitch together code often feels unnatural. This raises a few questions:<p>1. Should we rethink programming languages for AI-assisted coding?
2. Is there a path toward a language that is both more readable for humans and easier for GenAI models to generate? Has anyone explored this idea, or would it require fundamentally new design principles (as there won't be a good stash of human written code already for the models to learn from) ?
A food for thought. If one combines multiple individual baking/cooking receipts not designed to be 'combined' will that result in same kind of result of one recipt with multiple parts (aka designed as one part, with multiple groups of steps). Taste testing the results, might lead to ideas on how to correct the first 'concatination' of receipts to 'taste' more like the 'composite' recipt.<p>Why would an AI-assisted coding design for the big picture if only has knowledge of one part without concept 'fits into other parts' to make a whole design.<p>Generating algorithm descriptions and combining the algoithmic descriptions (to smooth/normalize the algorithm as a whole) before generating code from algorithm would make things bit less fragmented. (same concept of compiling to intermediate code, then on each individual type of hardware architechture, translating the intermediate code into something runable on particular hardware architechture.)<p>Incomplete specifications cause problems for humans. AI's learn from humans. Why would an AI learing from 'incomplete specifications' infer what the 'correct specifications' are (without being requested to do such an analysis)?
Main problem is not language, but that AI models still not creators but imitators (I mean only deep AI, like LLM, will explain).<p>Yet I read about very simple test - just ask AI to draw glass of wine - it could draw only half-filled glass if trained on half-filled glass and could not grow to higher abstraction, abstract glass from fullness.<p>I don't know, what we really need to overcome this limitation, but while it exist it will limit all AI usage.<p>To be more constructive, I think, most practical languages for AI could be something like SQL, mean fully declarative, without any imperative part, so AI will just translate human text to language.<p>So may be solution, some hybrid approach, where human define limits on solution; deep AI is just translator; and some classic optimization technique will generate solution (on HN appeared link on query compiler book, where you could see huge list in contents, and later I'll check, before I seen link with list of just all existing optimization approaches).<p>Other possible solution I see, may be some hybrid of deep AI and semantic AI. For semantic AI exists problem that it need to define any possible way, it cannot even hallucinate to create something new.
It's not clear to me <i>how</i> a language would be designed to be easier for GenAI to get right. Python seems like a pretty good target already, given that there's a lot of such code in the training data already plus how it reads (with idioms like "for... in..." etc.).
Hypertokens define a spreadsheet-esque label-based referential coding environment. HTs can be used in any LLM context to create arbitrary operations as will be shown in the paper. Current expected release is later this month.
I am using ChatGPT to help me build a language that has a domain-specific use case with non-LLM models. The interpreter is written in Rust and the language resembles a mix of Lisp and SQL.