Is there an actual language implementation? Or is the repo empty because the whole idea is that you're specifying things LLMs will already understand (but in a standarized way)?
Very cool idea in general. I guess I'm just curious what the practical limits of just-a-format-spec is. Like, the original golang compiler was written (ie bootstrapped) in C. Is there some equivalent bootstrapping process here or is it all just format specification? e.g. when I say MUST, I would expect the output to really adhere to that constraint. But, in my experience, LLMs do not always comply. Is that your language's responsibility to solve or mine?
That is I think you need 3 stages: (a) Lexer + Mini LLM that transpiles HUMAN lang into structured JSON and also outputs a static prompt describing how to interpret this JSON. Both of which are then fed (b) into big LLM, (c) Output code (plus instruction prompt) is then fed back to another smallish LLM to verify that OUTPUT matches the "MUST" "SHOULD" "NEVER" directives.
[Mini LLM] = Something like gemma3-270m finetuned as a lexer / json compiler [BIG LLM] = Claude 4 or whatever [Verification LLM] = Could be something like Qwen3.
I am going to write a basic lexer in rust. eventually a formatter too. the repo is empty for now because i am still working on it. the cool thing is that it does not need a lexer, but it would be nice. it is going to have a lot of cool stuff coming soon, like a package manager.
Very cool idea in general.
Ty let me know how it works for you.
Is that your language's responsibility to solve or mine?
the entire goal is that the language will solve it. still figuring it out but the idea is that this would integrate directly into an LLM so it is actual code, not just a prompt. i also want to add ADDRESS_HALLUCINATION or something similar eventually.
also, i was going to name it huumn but that github @ was taken... by you! lol.
new project idea: create a country called Human and apply for .hmn or .human TLD -- when human/AI distinction becomes more important, that TLD may become quite valuable.
I read a tweet where some savvy guy bought walmart.ai, intel.ai, etc shortly after ChatGPT launched and raked in the mongolian beef if you know what I mean.
.human domains would be nice. I bet you could really gouge.
Ah this was a bummer. It looked good at first but then when I dig in, it looks like pre-pre-alpha with no working examples. I was hoping there was something I could try out.
Is there an actual language implementation? Or is the repo empty because the whole idea is that you're specifying things LLMs will already understand (but in a standarized way)?
Very cool idea in general. I guess I'm just curious what the practical limits of just-a-format-spec is. Like, the original golang compiler was written (ie bootstrapped) in C. Is there some equivalent bootstrapping process here or is it all just format specification? e.g. when I say
MUST, I would expect the output to really adhere to that constraint. But, in my experience, LLMs do not always comply. Is that your language's responsibility to solve or mine?My take:
[HUMAN LANG] --> [Lexer + MINI LLM] --> JSON + Prompt -> [BIG LLM] --> OUTPUT + Prompt --> [Verification LLM]
That is I think you need 3 stages: (a) Lexer + Mini LLM that transpiles HUMAN lang into structured JSON and also outputs a static prompt describing how to interpret this JSON. Both of which are then fed (b) into big LLM, (c) Output code (plus instruction prompt) is then fed back to another smallish LLM to verify that OUTPUT matches the "MUST" "SHOULD" "NEVER" directives.
[Mini LLM] = Something like gemma3-270m finetuned as a lexer / json compiler
[BIG LLM] = Claude 4 or whatever
[Verification LLM] = Could be something like Qwen3.
Json takes too many tokens, and will not compile to json. however, i have thought about it.
{
"react_dev": {
"model": "gpt-4",
"temperature": 0.2
},
"todo_rules": {
"NEVER": [
"use_class_components",
"mutate_state_directly"
],
"MUST": [
"use_hooks",
"include_add_todo",
"include_delete_todo",
"include_toggle_complete"
],
"SHOULD": [
"use_local_storage",
"handle_empty_input"
],
"AVOID": [
"inline_functions",
"magic_numbers"
]
},
"generates working component": {
"INPUT": "Create a React todo list",
"EXPECT": "contains useState and retu"
}
}
I am going to write a basic lexer in rust. eventually a formatter too. the repo is empty for now because i am still working on it. the cool thing is that it does not need a lexer, but it would be nice. it is going to have a lot of cool stuff coming soon, like a package manager.
Ty let me know how it works for you.
the entire goal is that the language will solve it. still figuring it out but the idea is that this would integrate directly into an LLM so it is actual code, not just a prompt. i also want to add
ADDRESS_HALLUCINATIONor something similar eventually.also, i was going to name it huumn but that github @ was taken... by you! lol.
I also own
huu.mn. I'll let it go for 10 BTC. Let me know when it's in escrow (if there's a platform for such things)!edit: with all the AI rage, I should probably put it up for auction. I might actually get 10 BTC from some fat cats.
stop abusing mongolia TLD haha
new project idea: create a country called Human and apply for
.hmnor.humanTLD -- when human/AI distinction becomes more important, that TLD may become quite valuable.I read a tweet where some savvy guy bought
walmart.ai,intel.ai, etc shortly after ChatGPT launched and raked in the mongolian beef if you know what I mean..humandomains would be nice. I bet you could really gouge.I like domains a lot. I can't wait to make use of https://outer.space
Wouldn't that be cybersquatting? Couldn't Walmart and Intel get those domains back via legal action?
I’m not super familiar with cybersquatting but I imagine they could claw them back that way
Nice starfield animation, brings back memories of Win 95
I suspect we're looking at a case of this #1077714
i do not care if this takes off or not, it was a side project i use personally and i thought i would share
That's awesome!
Ah this was a bummer. It looked good at first but then when I dig in, it looks like pre-pre-alpha with no working examples. I was hoping there was something I could try out.
This you?
https://njump.me/npub1wc2de6hg7xk3htekmqjpahjjakl42glaxu6zntvvmalk8tkg7nksma5jth
Got a buddy in enterprise saas who recently had a new CTO come in and tried pushing infrastructure-as-prompts
He didn't last long
deleted by author