Writing is fundamentally the transfer of information from your brain to my brain. If you have 1000 bits of semantic information you want to transfer, you can't give 300 bits of semantic information to an LLM and have it fill in the remaining 700, because it doesn't know what those 700 bits are. If it's able to guess those 700 bits correctly, then they aren't true semantic information, and you really only have 300 bits you want to transfer. You might as well transfer those bits to me directly, rather than having the LLM add on an extra superfluous 700 bits that I then have to filter out.
It isn't the transfer of information at all. What's actually happening is you're prompting experiences in a human instead of an AI using text.
Communication only works if you have multiple levels of representation and abstraction, including but not limited to - letter shapes, grammatical structures, style and register, stylometry, and subtext.
All of that is learned, and writers usually assume they can rely on that learning as the context for the text.
So you don't write to 'transfer information' like a network cable, you write to trigger experiences in the human version of latent space.
Factual information is one kind of experience. But even when that's the goal, there are always layers of implied relationship, social register, role, status, and other implications in everything that's written.
In normal communications the context - business emails, personal messages, mainstream journalism, fiction, and the rest - defines what acceptable language looks like.
The content fits inside that. But it has to fit the context, otherwise it lands in a semantic and psychological uncanny valley - like sending LinkedIn speak to a spouse on a wedding anniversary.
The real problem with LLM writing is that it's good at the technical layer - the grammar and spelling - and has some insights into the rest.
But the default content style is marketing and ad speak. And recently it's developed a weird and unique hybrid style which applies marketing fluff and pretension to technical content like code comments.
So you get one register instead of all of them. It can attempt others, but it's still too limited to generate them fluently. Sometimes the results are outstanding, but often it defaults to mechanical clichés.
So that's why it sucks and sounds so hollow.
Can it be fixed? Yes, but it's very hard work, most people don't have the skills, and it takes time - often too much time to be worth the effort.
I'm definitely being slightly too emphatic when I say it's "fundamentally the transfer of information", but I don't think that nuance is important.
When LLMs eventually get good at writing in the correct style for a given context, I'll admit that they have value in that way. But they aren't good at that yet. And even when they do get that good, I'll still dislike it for reasons that are more emotional than rational.
I don't care how good the LLM gets. If I know some text was written by LLM I'd much rather know the prompt - the seed of intent.
If the seed of intent is "convey XYZ details so they know them" then I can choose to go and learn those details any way I see fit - maybe even ask an LLM to summarise some data for me! - rather than having to ingest whatever their LLM use poops out and trying to digest the intent and content and figure it out.
Output is not interchangeable with the prompt. In many cases, the prompt does not have the information the sender wanted to give you, and there is no guarantee that your LLM will give those information - or do it correctly - if you use the prompt yourself.
The entire value of here is that sender read the output and is vouching for it. This is where "bits of information" come from. If the sender cannot be trusted to verify and vouch for the LLM text they're sending to you, well, they're an asshole and you should rebuke them or find someone more considerate of others to talk with. Them giving you their prompt doesn't help you with anything.
This is what's so amusing to me. People say they're not interested in the output of an LLM, only what a human has to say. But then when a human says "These words from the LLM are good, I vouch for them" the very same people say "If I wanted those words, I'd get them from the LLM myself, what's the point of this human at all?"
I am focusing on the thing that is usually hidden. You are correct that for a full picture I'd need the information the LLM operates on.
But my point is that the prompt is the nearest encapsulation of the intent of the originator. Give me the data and the prompt. Vouch for the output if you like, but I want the source.
hatthew · · focus · HN ↗
Writing is fundamentally the transfer of information from your brain to my brain. If you have 1000 bits of semantic information you want to transfer, you can't give 300 bits of semantic information to an LLM and have it fill in the remaining 700, because it doesn't know what those 700 bits are. If it's able to guess those 700 bits correctly, then they aren't true semantic information, and you really only have 300 bits you want to transfer. You might as well transfer those bits to me directly, rather than having the LLM add on an extra superfluous 700 bits that I then have to filter out.
TheOtherHobbes · · focus · HN ↗
Communication only works if you have multiple levels of representation and abstraction, including but not limited to - letter shapes, grammatical structures, style and register, stylometry, and subtext.
All of that is learned, and writers usually assume they can rely on that learning as the context for the text.
So you don't write to 'transfer information' like a network cable, you write to trigger experiences in the human version of latent space.
Factual information is one kind of experience. But even when that's the goal, there are always layers of implied relationship, social register, role, status, and other implications in everything that's written.
In normal communications the context - business emails, personal messages, mainstream journalism, fiction, and the rest - defines what acceptable language looks like.
The content fits inside that. But it has to fit the context, otherwise it lands in a semantic and psychological uncanny valley - like sending LinkedIn speak to a spouse on a wedding anniversary.
The real problem with LLM writing is that it's good at the technical layer - the grammar and spelling - and has some insights into the rest.
But the default content style is marketing and ad speak. And recently it's developed a weird and unique hybrid style which applies marketing fluff and pretension to technical content like code comments.
So you get one register instead of all of them. It can attempt others, but it's still too limited to generate them fluently. Sometimes the results are outstanding, but often it defaults to mechanical clichés.
So that's why it sucks and sounds so hollow.
Can it be fixed? Yes, but it's very hard work, most people don't have the skills, and it takes time - often too much time to be worth the effort.
hatthew · · focus · HN ↗
When LLMs eventually get good at writing in the correct style for a given context, I'll admit that they have value in that way. But they aren't good at that yet. And even when they do get that good, I'll still dislike it for reasons that are more emotional than rational.
_carbyau_ · · focus · HN ↗
If the seed of intent is "convey XYZ details so they know them" then I can choose to go and learn those details any way I see fit - maybe even ask an LLM to summarise some data for me! - rather than having to ingest whatever their LLM use poops out and trying to digest the intent and content and figure it out.
It is about empowerment, rather than eating shit.
TeMPOraL · · focus · HN ↗
Output is not interchangeable with the prompt. In many cases, the prompt does not have the information the sender wanted to give you, and there is no guarantee that your LLM will give those information - or do it correctly - if you use the prompt yourself.
The entire value of here is that sender read the output and is vouching for it. This is where "bits of information" come from. If the sender cannot be trusted to verify and vouch for the LLM text they're sending to you, well, they're an asshole and you should rebuke them or find someone more considerate of others to talk with. Them giving you their prompt doesn't help you with anything.
ModernMech · · focus · HN ↗
TeMPOraL · · focus · HN ↗
_carbyau_ · · focus · HN ↗
But my point is that the prompt is the nearest encapsulation of the intent of the originator. Give me the data and the prompt. Vouch for the output if you like, but I want the source.