[barnabee]

Probably the most fun I’ve had with LLMs has been slowly making a programming language as a side project.

I used to give up somewhere around the type system, too, but this time I’m approaching something vaguely useful. It even has a basic LSP.

It’s been both enjoyable and enlightening, and LLMs turn out to be an excellent pair designer as (in addition to implementation) they’re really good at summarising the impact of various decisions.

> the reason you will fail for sure, is the inability to restrict the scope of the project

This will be the reason, for sure. But then the scope of every project like this tends towards building an OS with it then replacing every piece of software, including all embedded devices :)

[sieve]

> slowly

I cannot do slow. It is either burn the candle at both ends, or do nothing at all.

I am using LLMs this time as well, but I spent close to 400 hours over a period of 6-7 weeks on my project before I put it to the side temporarily (got bored once the thinking part was done). About 300 of those were spent on iterating over the language and VM specs and eliminating all ambiguities and needless features. The remaining 100 were used to produce the code --- the VM, the assembler and the compiler --- and to repeated rewrite it to conform to my way of doing things.

LLMs have let me become extremely choosy about which code I am willing to keep.

[barnabee]

I've taken the approach of writing and even directly reviewing almost no code for this, otherwise I'd simply not have time for it as another side project. It's also interesting to see how far I can push this "vibe engineering" approach, and although it's not perfect, the answer is much further than I'd have expected going in.

I've managed to get OpenCode setup such that I can have a productive discussion about the design or an issue / change then leave the LLM iterating for long periods while I do other work. It's instructed to maintain test coverage and treat quality very seriously - as a result there are over 5000 tests (some I suspect are useless...) and it's pretty rare to get a regression.

I'm pretty sure there are plenty of significant bugs and gaps, but also that once found it seems like all of them will be fixed pretty quickly by the LLM.

I just have to avoid looking at the code...