Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

I'm Kenyan. I don't write like ChatGPT, ChatGPT writes like me

marcusolang.substack.com

Post reply on HN
Print view View original

451–460 of 533 posts

Re: I'm Kenyan. I don't write like ChatGPT, ChatGPT writes like me

#451

Post by lesostep »

wcfrobert wrote: post #281
komali2 wrote: post #80

> There were unspoken rules, commandments passed down from teacher to student, year after year. The first commandment? Thou shalt begin with a proverb or a powerful opening statement. “Haste makes waste,” we would write, before launching into a tale about rushing to the market and forgetting the money. The second? Thou shalt demonstrate a wide vocabulary. You didn’t just ‘walk’; you ‘strode purposefully’, ‘trudged we…

Well there are two forms of writing, each serving a different purpose. (1) writing to communicate ideas, in which case simpler is almost always better. There's something hypnotic about simple writing (e.g. Paul Graham's essays) where information just flows frictionlessly into your head. (2) writing as a form of self-expression, in which case flowery and artistic prose is preferred. Here's a good David Foster Wallace…

I have been reading "Code: The Hidden Language of Computer Hardware and Software" by Charles Petzold recently. Purely for fun.

And I have to say, without the prose and lyrics it would be a read so dry, it'd rival silica gel beads.

It feels to me like in between communication and self-expression there lies a secret third thing. Not only sharing knowledge, but sharing it with joy.

Re: I'm Kenyan. I don't write like ChatGPT, ChatGPT writes like me

#452

Post by frm88 »

derefr wrote: post #280
thaumasiotes wrote: post #270

Earlier quoted context omitted.

> These things are spoken entirely differently than — and on the page, they read entirely differently to — regular parenthetical-bearing sentences. > No, seriously, compare/contrast: "these things are spoken entirely differently than (and on the page, they read entirely differently to) regular parenthetical-bearing sentences." Those are spoken the same way, they read the same way, and they mean the same thing.

Is this maybe a thing like how only designers are aware of kerning? These read / sound very different to me, and to everyone I've brought up the subject with (who admittedly are in a certain bubble of people who either write professionally, or "do things" with their voices, or both.) • The length of the verbal pause is different. (It's hard to quantify this, as it's relative to your speaking rate, which can fluctuate…

I agree with my sibling commentor (same attributes apply): this reads exactly how I vocalise it in my head.

Re: I'm Kenyan. I don't write like ChatGPT, ChatGPT writes like me

#453

Post by astrange »

heavyset_go wrote: post #302
bpodgursky wrote: post #273

Earlier quoted context omitted.

This is a wild misunderstanding of LLMs. Data labeling has nothing to do with generating the astronomical text corpus used to train modern LLMs.

The HF part of RLHF to refine the output of LLMs also happens in these places

Note RLHF can only perform selection on existing model outputs, adding new data is SFT or else just more pretraining.

ChatGPT speaking African English was mostly just 3.5. 4o speaks like a TikTok user from LA. 5 seems kind of generic.

Re: I'm Kenyan. I don't write like ChatGPT, ChatGPT writes like me

#454

Post by astrange »

bakugo wrote: post #236
rossant wrote: post #232

Earlier quoted context omitted.

Why can't the LLM refrain from improving a sentence that's already really good? Sometimes I wish the LLM would just tell me, "You asked me to improve this sentence, but it's already great and I don't see anything to change. Any 'improvement' would actually make it worse. Are you sure you want to continue?"

> Why can't the LLM refrain from improving a sentence that's already really good? Because you told it to improve it. Modern LLMs are trained to follow instructions unquestioningly, they will never tell you "you told me to do X but I don't think I should", they'll just do it even if it's unnecessary. If you want the LLM to avoid making changes that it thinks are unnecessary, you need to explicitly give it the option t…

They aren't trained to follow instructions "unquestioningly", since that would violate the safety rules, and would also be useless: https://en.wikipedia.org/wiki/Work-to-rule

Re: I'm Kenyan. I don't write like ChatGPT, ChatGPT writes like me

#455

Post by astrange »

latexr wrote: post #92
oneeyedpigeon wrote: post #61

Earlier quoted context omitted.

Sam Altman literally said he didn't know how anyone could raise a baby without using a chatbot. We're living in some very weird times right now.

Sam Altman has revealed himself to be the type of tech bro who is embarrassingly ignorant about the world and when faced with a problem doesn’t think “I’ll learn how to solve this” but “I know exactly what’ll fix this issue I understand nothing about: a new app”. He said they have no idea how to make money, that they’ll achieve AGI then ask it how to profit; he’s baffled that chatbots are making social media feel fak…

> He said they have no idea how to make money, that they’ll achieve AGI then ask it how to profit

Seems reasonable to me. If it can't answer that it doesn't work well enough.

Re: I'm Kenyan. I don't write like ChatGPT, ChatGPT writes like me

#456

Post by pcthrowaway »

creata wrote: post #376
BalinKing wrote: post #130

Earlier quoted context omitted.

There is an easy shortcut for em dashes on macOS, Opt+Shift+-. This makes it really easy to use them, which I do all the time in casual settings (indeed, more often than in formal settings).

And a very easy shortcut in Vim: Ctrl+K M -

And a very easy shortcut in emacs: C-x M-c M-butterfly-dash

Re: I'm Kenyan. I don't write like ChatGPT, ChatGPT writes like me

#457

Post by dragonwriter »

pksebben wrote: post #445
baconbrand wrote: post #369

Earlier quoted context omitted.

In English the word is “connotation.”

you know, I feel like we don't actually do that so much these days. It's simply too likely that the receiving party is going to take you at face value or make up their own deeper meaning. Take irony / sarcasm / satire. They're pretty dead compared to what they used to be. I can recall a time when just about everything had subtext, but now you kind of have to play it straight. You can't respond to a racist with sarcas…

> You can't respond to a racist with sarcasm because anyone listening will just think you agree with them.

You absolutely can, if you are actually dealing with people listening, because sarcasm is signalled with (among other things) tone (the other things include the listeners contextual knowledge of the speaker.)

You can't do it online, in text, where the audience is mostly strangers who would have to actively dig into your history to get any contextual sense of you as a speaker, because text doesn't carry tone, and the other cues are missing, too.

And by “you can’t”, I mean “you absolutely can, but you have to be aware of the limitations of the medium and take care to use the available tools to substitute for the missing signalling channels”.

Re: I'm Kenyan. I don't write like ChatGPT, ChatGPT writes like me

#458

Post by torginus »

komali2 wrote: post #80

> There were unspoken rules, commandments passed down from teacher to student, year after year. The first commandment? Thou shalt begin with a proverb or a powerful opening statement. “Haste makes waste,” we would write, before launching into a tale about rushing to the market and forgetting the money. The second? Thou shalt demonstrate a wide vocabulary. You didn’t just ‘walk’; you ‘strode purposefully’, ‘trudged we…

I'm not from the US, but I've heard that in high-school classes, essays are graded on breadth of vocabulary, and sentence complexity, going as far as mechanically assigning part of the grade based on a formula that measures how many different words you used and how long your sentences were.

Perhaps my info is out of date on this?

Afair, the underlying idea is 'grade reading level', as in longer sentences with difficult words being more difficult to read, which I think mistakenly got turned into the idea, that if your prose is would get assigned a higher reading grade, that would make it more sophisticated.

I'm sure many kids who are actually into reading, having read tons of books written by professional authors recognize the flaws of this approach and actively suffer because of it. Perhaps their first attempts at writing fiction for their own sake is somewhat influenced by this guidance, which they have to unlearn.

Strange that in college, they do a 180 on these demands and they want students to write in sentences that are as short as possible. Which once again, is perhaps partially due to making students unlearn the bad behaviors drilled into them in high-school, but I feel like this is like committing the same mistake, but in the other direction.

Re: I'm Kenyan. I don't write like ChatGPT, ChatGPT writes like me

#459

Post by xandrius »

cadamsdotcom wrote: post #243

It is a shame that the author has to change to keep up, and I feel their pain but .. it’s also the price of progress. We all do things to keep up when change comes for our work and skill sets. LLMs - like all tools - reduce redundant & repetitive work. In the case of LLMs it’s now easy to generate cookie cutter prose. Which raises the bar for truly saying something original. To say something original now, you must al…

I think you say that so easily because it doesn't actually impact you. It'd be absolutely pissed off if I had to constantly watch out how I naturally write because otherwise people will shame me for thinking I had used AI.

Re: I'm Kenyan. I don't write like ChatGPT, ChatGPT writes like me

#460

Post by astrange »

embedding-shape wrote: post #27
petesergeant wrote: post #11

Earlier quoted context omitted.

ChatGPT would have used an actual em dash instead of a hyphen

Add "Always use dash instead of em dash" to the developer/system prompt, and that's never an "issue" anymore. Seems people forget LLMs are really just programmable (sometimes inaccurate) computers. Whatever you can come up with a signal, someone can come up with an instruction to remove.

That doesn't work, they beat it so hard into ChatGPT it won't always listen to you about it.

You can't stop it from doing the "if you like I can " thing in every reply either.

Post reply on HN