But, high level languages don't really lose precision in any way that matters (except when you need extreme performance, resistance to timing attacks, or niche stuff like that). Natural language on the other hand loses precision that's usually important, so I don't see how your comparison is valid.
Heck, OpenAI can't even get their models to reliably do what they intend, so yeah.
If you don't get the difference, I get there isn't much that you know about how instruction selection works, across all those approaches.
Do any of the optimizations you mentioned affect the underlying behavior of the run code? Semantics...
If not, it doesn't fit to LLM use, where this is very much the case. We can talk about bugs in compilers and undefined behaviours and so on and so forth all the livelong day, but you know as well as I that this is not of the same scope of impact.
I could and in the past would have written a long diatribe about how pure compilation to assembly is deterministic, thus making this not really suitable and that while yes, there are things like JVM that aren't 100%, we have contracts, docs, specs and, because of them, an expectation of consistency that LLMs do not deliver. I did not because I thought that was not necessary, but see my error now, that is my mistake.
Which for example, unless you can exactly replicate the world every single time a PGO data gathering session is done, there are no guarantees that the compiler will generate exactly the same machine code pattern, when given various PGO files generated at different occurrences.
https://static0.cbrimages.com/wordpress/wp-content/uploads/2...
Gleam also runs on BEAM, which has important runtime characteristics vs other languages. That will impact your system and how it scales even if you never look at the code.
I've seen even the best models try to build & debug parallel and distributed systems. The results are always a bit uneven at best, even with a good harness and a lot of guidance.
The disadvantage of niche languages is that the LLM won't quite know the language as well due to less training data being available. But if you actually need to build something distributed, then this might be a good trade-off.
Moreover, to me coding carefully is still a hobby, a pleasure, a craft. Gleam is exactly the kind of language that makes this such a pleasant activity.
Because, we do?
You may not, which is fine. A lot of developers may not... that's fine. I guess you're all employed and don't risk your livelihood when things go wrong.
I do though. I take responsibility.
And just like with any junior developer I guide AI and I do review what it creates in the project that matter. (Personal tools I've entirely vibe-coded, never any client projects though)
1. In a lot of cases, even if you're pure vibecoding and not reviewing any code, its still useful to be able to read it - you can then verify and reason about various aspects. LLMs are often flat out wrong on niche issues of performance etc, so for cases like this, not being able to read code is a big risk.
2. "Languages optimized for LLMs" sounds awesome, but so far, to me at least, it seems like LLMs work best in the same situations as humans- good test coverage, low-complexity, clean architecture, etc.
On the "languages for LLMs front", isn't that a bit "cart-before-horse"? Doubt we are anywhere near the limit of what a model could do with a static verifiable language, even more so since stuff like TS has massive training data, any new language wouldn't benefit in such a way.
Add to that, I'd rather be able to understand the output, partly for aforementioned reasons, partly because we have seen with post Spud GPT reasoning/summarisation issues what happens when being able to read model output becomes even slightly harder. Having model-focused languages would only exacerbate such issues.
Also, absolutely not a Maths expert so take this for an utter layman regurgitating of what I have read, but apparently the Lean proofs by OpenAI are very hard to parse even for the experts in the field, lovingly called "Clanker Proofs", even when they are valid.
Having not just correctness but also human understanding as a goal has merit, so there is that to consider too. Rather models get more efficient at using what we have (and can understand+maintain as humans if need be) then make it more opaque.