@RosaCtrl a program which deterministically turns a specification into machine code is literally a compiler, even if the specification is written in a natural language. Writing such specifications is programming, ergo the language used to write such specifications would be a programming language.

It's the same old "if we just make programming look like English/flowcharts/UML diagrams, we won't need programmers anymore" idea that has been around since basically forever. It never works because it's based on the fundamental misunderstanding that syntax is the hard part of programming, when it's actually creating a specification that's detailed and unambiguous enough to get you the results you want.

For the same reason, I expect such a deterministic LLM-based language to be pretty much useless. Being good at guessing what the spec intends is great for quick and dirty hacks or "write me a game like pacman but different"-type demos intended to wow people into forking over money for your tool, where the point is that you get something but exactly what you get doesn't really matter. But it's absolutely awful for building real systems where details matter. Having to write those specs in an informal language is just going to, counter-intuitively, make syntax way, way more of a problem.

@RosaCtrl does it though? In my (obviously very anecdotal) experience, LLM enthusiasm generally comes from:

Except for the first category, which generally doesn't care (or know anything about) tech at all, I've rarely heard anyone see LLMs as a way to make up for bad languages. If anything, most people I've talked to who are not outright hostile to the very concept of LLMs put more emphasis on the importance of good language tooling to be able to effectively use LLMs.

@nyancient I think the last three cases may have bad languages a source. But when I said crap above I should have clarified that it may be a language, a library, a method, whatever we find crappy.

For example someone asked me if I wanted the keep spending hours updating dependencies in big projects, or if I preferred to let an LLM to do it. My answer was I use very few dependencies, and I fight at work to get the time to update them often.

Now regarding PLs in particular, I know a former GHC contributor that told me, to my face, he doesn’t care about programming languages that much anymore. He still has to deal with them, but theres hope that that won’t be the case in a near future. That kinda broke my heart, as I do believe we would lose a lot if we achieve purely natural language programming

@RosaCtrl ouch. Are you sure they don't have a stake in some LLM business or research grant? To be fair though, I've been pretty disillusioned with the whole SML branch of the FP community after almost everyone there I knew and respected went all in on crypto, and even before then I don't really think that crowd is representative of anything in tech.

Even so, I don't think it's unreasonable to work on the problem from both ends. Better synthesis and analysis tools are a valid band-aid for the billions of lines of code that are already written in bad languages. Of course, whether LLMs will end up meaningfully contributing to that is very much an open question, and anyone considering putting all their eggs in the LLM basket is very poorly informed at best.

Considering that just with current investments the "AI" industry would require revenue of 70 EUR per human on the planet in perpetuity to deliver just a 10% ROI, according to JP Morgan, LLMs rapidly ruining the training material their own existence depends on, I'm far from convinced that they will be.

@RosaCtrl also, there definitely are tasks that are just inherently boring, no matter how good languages become. I can't imagine that fine tuning SELinux or AppArmor rules for a complex, distributed system can ever be anything but frustrating, for example.

Unfortunately, that particular example is also completely unsolvable with LLMs. Which I had to explain to our CTO at great length a few weeks back. ಠ_ಠ