Login
You're viewing the assemblag.es public feed.
  • ahistorical immaterialistjimbob@aus.social
    Oct 26, 2026, 7:00 AM

    Wondering if as an experiment...

    1) One were to develop a new programming language, let's call it Greg.
    2) One were also to write some excellent, complete documentation of that language.
    3) One were to absolutely NOT put any example programs of that language anywhere online. Just don't use the language at all.

    If you were then to say Claude for example - "Write me a program to perform task X in the new programming language Greg, here is the language's documentation", I suggest that there is absolutely no chance in hell it would output a viable program

    LLMs can't code because they "understand" programming from first principles, they can only code because of the library of code in their training data.

    💬 6🔄 3⭐ 21

Replies

  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 2
  • 💬 0🔄 0⭐ 0
  • Theophile EscargotTheoEsc@mastodon.social
    Oct 26, 2026, 8:53 AM

    @jimbob Not sure. Could maybe do it using multiple agents. One agent writes code in a language it knows. Another agent compares docs for both languages, converts the code to Greg. Other agents run the Greg code iteratively, checks output. Change conversions, Greg code iteratively until it works.

    💬 1🔄 0⭐ 0
  • ahistorical immaterialistjimbob@aus.social
    Oct 26, 2026, 8:56 AM

    @TheoEsc yes, maybe. But I would argue "iteratively throwing shit against the wall and checking whether it sticks" is not in the same ballpark, or indeed continent, as "intelligence". That's basically an evolutionary algorithm - evolution also, famously, is not intelligent.

    💬 0🔄 0⭐ 2
  • Scaliest Pinon 👻nonspecialist@aus.social
    Oct 26, 2026, 12:42 PM

    @jimbob would be fun to confirm. I presume Greg has to have at least some kind of process flow termination symbol called “STOP”?

    💬 1🔄 0⭐ 0
  • 💬 1🔄 0⭐ 0
  • Scaliest Pinon 👻nonspecialist@aus.social
    Oct 26, 2026, 8:46 PM

    @jimbob there’s an opportunity for some hilarity here too. Err, an actor-based language that “drives” but is directionless and sometimes misses the stop.

    💬 0🔄 0⭐ 0
  • Bakunin Boysbakuninboys@aus.social
    Oct 26, 2026, 9:33 PM

    @jimbob there's an Electroboom video about AI electronics, and because there's no corpus, it cannot understand it at all. Even with a corpus, it would need tooling to go between HDL / EDA and the image representation. Some people have noted that AI only works in software because you can launder the commons with floss.

    💬 1🔄 0⭐ 0
  • ahistorical immaterialistjimbob@aus.social
    Oct 26, 2026, 9:42 PM

    @bakuninboys I had wondered about electronics, but I wasn't sure about the existence of some kind of code/language to describe circuits that would be used.

    See also, assembly language. Why are people getting LLMs to output Python, why not just cut to the chase and have them produce highly optimised machine code? The same reason as you say - there is no corpus of highly annotated machine code to train them on.

    💬 1🔄 0⭐ 1
  • Bakunin Boysbakuninboys@aus.social
    Oct 26, 2026, 9:47 PM

    @jimbob there are but HDL is the closest to being "open". EDA tools will come with a language too, but I dont know if there's a standard. Generally it's used for simulation or "debugging". From what I've seen, LLMs only really know JavaScript. They'll write other languages, but really they're trying to write as JavaScriptey code as possible. They don't abstract or modularise. Someone I know called them a "word calculator" and that's pretty close. A word calculator which makes mistakes.

    💬 0🔄 0⭐ 0