Hacker News Reader

Comment by munk-a | Hacker News Reader

Specifically, lets imagine LLMs as compilers - you're passing your prompt through to get some pretty language at the end.

Don't send me your compiled code, send me your prompt. Let it be rude, if the wording is awkward I guarantee I can understand it just as well as an LLM, ignore the fact that my daughter just graduated and offering hallucinated platitudes.

Send me the actual question, don't make me try and decompile a big blob of empty text to the ten word prompt that contains all the actual meaning.

AI is a useful tool for a variety of purposes, what it is not useful for is expanding a short statement to an essay and reducing an essay to a short statement. Either the communication deserves to be an essay or ten words will get it done.

View on HN

yeah this is what drives me crazy about LLM writing. Most of the time the prompt has all the info you need and is like maybe a few sentences. Then the LLM expands it into a few paragraphs...

I guess if someone is writing like a big fancy email to send out in bulk, maybe using an LLM to improve would make sense... but just emailing some coworkers it seems super lazy and insulting to send an LLM output :-I

Some people send LLM generated replies in Slack chats! Now there's that.

Tangentially, yes, let's imagine LLMs as compilers.

How insane is it to advocate the usage of these non-deterministic compilers, where each time you compile may produce different semantics?

And then people resort to saving and hand-editing the compiled output.

But when they want to change the source, they recompile and have to start over hand-editing the output again.

Actually, in professional usage in a technical setting this is my prime objection to heavily LLM driven development. Were the tools in usage deterministic then I'd be a lot less objecting to the mandating of their incorporation into workflows.

I want to be reading, writing, testing and maintaining the software at the same layer. Right now extreme AI usage leads to reading, testing and maintaining happening in a less expressive language than writing and guess which of those four activities developers enjoy the least and find the hardest - it sure isn't writing.

More than non deterministic : LLMs don't have a specification to obey to in the first place, while compilers (rather, programming languages) do.

if you are prompting such that the LLM isn't pulling context that the recipient doesn't have access to, then your email is likely marginal.

ie the prompt "Send 'bob' an email with a description of why the VPN bridge isn't working so they can debug their side" is a mostly useless as a prompt for anyone, it's only useful when the LLM has all the context of some analysis of the particular issue and what is going on and then injects it into the email.

Does the recipient already have that context? If you want to share some internal context (say you're a front-end specialist and the recipient is quite unfamiliar with the limitations of your framework) then maybe that'd be helpful? If it's just regurgitating already communicated information then either

1. You are restating the information because you don't believe the recipient understood it the first time and thus you should be very precise in your expression to make sure it isn't too arcane for them 2. The recipient appears to understand the information already in which case why restate it?

I wonder if slight grammar incorrectness (like not capitalizing your sentences or using abbreviations) is going to start becoming a signal of authenticity for people subconsciously. Maybe it already has.

Either rely upon everyone else changing their behavior, or give up and use your LLM to re-compress incoming messages to be informationally dense as you see fit.

It's lossy tho. LLMs are crap at picking the "good stuff". Eg: the summary of the email covered the point about the family event but missed that the deal-terms were moving from Wyoming to Delaware.

Personally, I'm confident with my level of output so I'll continue to dutifully read all the crap that gets sent to me on company time. I'll just prefer to engage with people who communicate well and encourage that in others.

A privileged position to be in, for sure! Congrats

I prefer not communicating with people who don't respect my time.

Just because you yourself are OK with being talked to rudely, doesn't mean others are. In fact I'd wager most aren't consciously or unconsciously...

Yeah that attitude will not get you far in life unless you're Steve Jobs, and it'll sink your ship unless you're obnoxiously rich. And even if you're either/both of those things: A. you can and should act better, and B. people will always attach an asshole-asterisk to your name for the rest of your life and probably even a good while after.

Exactly. In my 25+ year career, I've encountered maybe two dozen or so people whose e-mails and chats were terse, yet admittedly succinct, one-liners and most of them were also raging assholes to work with. The ones who also didn't use capital letters or punctuation in their communications were uniformly assholes.

My experience is completely opposite. Majority of the people send short emails - sometimes badly written, sometimes well written. And when someone sends long emails, people dont read them. They kinda skim them and ignore the rest.

And no, they are not all raging assholes.

Correct, I was referring to being arrogantly blunt and insensitive, not short/efficient emails.

I'm pretty sure "rude" in this context means "brief and to the point", not insulting. Otherwise you can be rude with an LLM as well.

Most people I know are happy to receive a focused email rather than an LLM-enhanced, 6 paragraph wall of text.

But it's far less rude to just bluntly say something than to send an email generated by an LLM.

I think rude was the wrong word to use. I more meant lacking the pomp and circumstance fluff. I always appreciate considerate and polite speech and think it's requisite to being taken seriously. However, I think directness within the bounds of politeness is optimal.

Also, if it's wall-o-text or "staging must be updated before our os version is deprecated sunday" I prefer the latter.

Reply on news.ycombinator.com