AGENTS.md on a standard protocol. Sorry but no. You either understand the thing or you don’t.

It’s a chat protocol. Something that is supposed to be very private and secure. Seeing an AGENT.md file is a huge turn-off there, sorry to say. Feel free to burn my karma down to oblivion.

▲ 0x1ch a day ago | parent | next [-]

I'd actually feel pretty safe with a LLM project if spec'd against the protocol documentation as opposed to a rookie's handcoded implementation. If security is the concern, I don't think using random reimplements is very safe to begin with.

	▲	jayd16 a day ago \| parent \| next [-]
		This is a false choice. You can pick neither and for all you know you're picking both.
	▲	Bender a day ago \| parent \| prev \| next [-]
		I'd actually feel pretty safe with a LLM project Same. I had a couple LLM's rewrite, harden an anonymous image board. It's not popular but exposed to the internet and the Tor network. Not worried at all.
	▲	rootnod3 a day ago \| parent \| prev [-]
		Unless the model changes and messes half the code base on next iteration. Or Anthropic decides to A/B test your agent. Or when shit goes sideways and someone actually needs to take a human look at the BF s. I’d rather have the guy who manually with human eyes read the specs.

▲ jjice a day ago | parent | prev | next [-]

> Something that is supposed to be very private and secure

Not sure how using an LLM for development of the tool has any affect on this. From a security perspective, I think that would've had more validity a year or two ago, but these days, I'd argue LLMs practice better security in generated code on average than a normal developer.

From the privacy angle, I have no idea how having an AGENTS.md (for development purposes) has any impact on privacy.

I don't want to "burn [your] karma to the ground", I just want to understand this angle, because I'm not sure I see how it applies.

▲ tcfhgj 9 hours ago | parent | prev | next [-]

The security comes from E2EE - which happens in the clients not the server

▲ 013 a day ago | parent | prev [-]

Do you not use AI for coding?

▲ solid_fuel a day ago | parent | next [-]

> Do you not use AI for coding?

A lot of people prefer to understand the things they ship. I don't understand why this remains surprising to some.

▲ lcnPylGDnU4H9OF 21 hours ago | parent [-]

Indeed, Linus Torvalds is infamous for not caring what he ships (https://news.ycombinator.com/item?id=50111350). That's why he's never been known to go on, let's say impassioned tirades against people who submit bad patches to the linux kernel. After all, it's not like he's ever apologized for inflammatory behavior before (https://news.ycombinator.com/item?id=18000698).

On a more serious note, it's not a dichotomy that one either uses AI to code or they care about what they ship. Some people use AI to code and care what they ship.

▲ solid_fuel 13 hours ago | parent | next [-]

Linus, like anyone else, has been wrong about things before, I'm not really interested in the appeal to authority here.

Two simple facts:

    1. Reading & understanding code takes longer than writing it the first time.
    2. Caring about what you ship requires *understanding* what you ship.

Therefore

    Using LLMs to write code cannot save you time, because it takes longer to understand what it has written than it would to write it in the first place.

That leaves simple changes like refactorings as the main use of LLMs in responsible development. Hardly a massive accelerator.

▲

simonra 11 hours ago | parent | next [-]

Point 1 is dependent on the individual and often not the case. Especially in cases where the part that takes effort is figuring out the poorly documented syntax/api you’re stuck with for the task (extra fun with new components iterating rapidly and pinning the next version requires new scaffolding), you can read the invocation and the accompanying couple of lines of test code way faster than you’d mould them yourself. Which also translates neatly to point 2 (how efficiently can you in many cases ascertain the outcome of smearing the chosen bits on the cpu with easily digestible tests).

Also the emotionally laden appeal to "responsibility" is weak, as it glosses over the many cases where overspending resources on creating the thing would be the irresponsible choice.

	▲	solid_fuel 11 hours ago \| parent [-]
		> Point 1 is dependent on the individual and often not the case. Not really. To properly understand code you must have a mental model of what it is doing. You need that model whether you are reading or writing the code. It is easier to build that model while writing the code than while simply reading it. To build the model while reading the code requires having the same understanding of the problem and the same approach already in mind, which is almost never the case for any moderately complex problem. > Especially in cases where the part that takes effort is figuring out the poorly documented syntax/api you’re stuck with for the task Reading LLM generated code does not avoid this, and in fact it is often worse since LLM comments are terrible at getting to the point.

▲

LoganDark 5 hours ago | parent | prev [-]

> I'm not really interested in the appeal to authority here.

Citing Linus isn't an appeal to authority, it's an example. A prominent example, but no one is saying to follow what he does or what he believes.

▲ bigstrat2003 19 hours ago | parent | prev [-]

> Some people use AI to code and care what they ship.

The vast, vast majority of people using LLMs to build software don't care at all about quality. I'll grant you Torvalds, but he's very much not the norm (and he's wrong: by the time you review the code the model wrote, you spent as much time reviewing it as you would've writing it, so no time was saved).

▲ stephenr 14 hours ago | parent | prev [-]

Believe it or not, some people actually know how to write code, and <shock horror> understand what it fucking does too!

I will never understand software dev's obsession with LLM generated slop. Compare it with the response from the likes of professional authors, artists.

I've said it before I'll say it again: a computer generating shit en masse for you is great if you don't know how to do it yourself, and don't care what the result is.