Post

I don’t think that article has any foundation on reality. I’m not even speaking from my own experience either

Jason Lengstorf

@jason.energy

I'm always willing to be proven wrong, but the research paper is using self-reported data 6000 execs across a variety of industries, so it's hard for me to believe it's not based in reality

I have this lil theory that LLMs are uniquely good at software compared to almost everything else atm for a few compounding reasons: 1. It’s made by software people, so we’re biased toward our use cases. 2. There happens to be an enormous corpus of code to train on, legally and otherwise.

19/2/26, 7:38 pm0 reposts21 likes

3. We’re the early adopters and have had the most time to get used to using it. I can absolutely understand a CEO of a paper mill might try the tech and get bad results out of it. And even within software this tech is at ⚠️proceed with caution⚠️ level. You may get eaten by a Grue.

I’m currently losing my mind over the whole skills.md file thing. Sure, they’re useful and make sense. Not arguing against that. But mfers are out there unironically saying shit like “you need to make sure you’re clear and to the point in your skills files, provide good examples”

@jurgisbekepuris.bsky.social

THEY’RE DOCUMENTATION THEN? THE SHIT YOU SHOULD HAVE BEEN WRITING THE WHOLE TIME? AND THE ADVICE FOR WRITING THEM BOILS DOWN TO THE EXACT SAME SET OF POINTS YOU SHOULD APPLY TO DOCS FOR HUMANS? HOW COOL. AND INTERESTING. TELL ME MORE. _dissolves into a frothing puddle_

Jurgis Bekepuris

21 Feb

Just ask Claude Code to write the skills.md and you don't need to write documentation^H^H skills.md anymore. You can thank me later.

“Yeah bro LLMs are great at figuring out how to use your CLI tool you just gotta write good --help output” SO ARE HUMANS. THIS IS KNOWLEDGE FROM THE 60S. I AM GOING INSANE.

Drake

@saeris.gg

yes, but you see now these devs *have the time* to write all those good docs, cuz the AI needs it and making the AI work better is their primary job function! AI freed them from having to write all the code, so now they can write all the docs! /s

dan

@danabra.mov

also 4. there's a clear feedback loop

Nick Williams

@wicky.nillia.ms

A subtle detail is that code produces concrete outputs. Those outputs can be verified independently of the code. Even if the code is a blackbox, you can poke and prod it and have some confidence the code works Whereas other purposes the only output is the LLM artifact. Which is difficult to verify

Patrick McEwen

@pmcewen.com

5) the cost of trying something is very low relative other fields 6) it’s very derivative relative to other fields (even most new innovations use existing languages, servers, web browsers, operating systems, etc and add a small new part)

Patrick McEwen

@pmcewen.com

7) there is minimal real world interaction with stuff that exists outside the computer

For some problems, yes. There’s no clear feedback loop on code quality, though, which is why I suspect the code LLMs output is so distinctively odd.

Luke

@lukeocodes.dev