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.
> 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.
Sure you can. LLM doesn't know what those 700 bits are, but you do. You may not realize it, and may not even know it at the time of prompting, but you do by the time you're sending.
Typical case is like this: you have 500 bits of semantic information to transfer. You give 300 of them to LLM, and get back the 500 bits you knew you have, and extra 500 you can quickly confirm are correct and relevant. Some of them are just dereferences of your input - where you recalled a pointer, but not what it pointed to. Some of it is information you never had before, but are able to easily validate.
You send that to me. I likely immediately realize the message was AI-assisted, but I trust you to be a decent human being, and not an asshole that lobs unverified LLM vomit over the fence for others to deal with. End result: you communicate 1000 bits of information to me, instead of planned 500, and you yourself learn extra 500 bits.
This is the optimistic scenario, but it does happen when LLM operator is not an asshole.
(Excuse the strong language, but I spent a lot of effort every day on both dealing with inconsiderate people lobbing LLM output at me, and making sure never to act like one myself, so it's a topic close to my heart.)
> Some of it is information you never had before, but are able to easily validate.
This is exactly the use case an LLM might (huge emphasis on might, depends on workflow, agentic vs. relying on contextual which can hallucinate) be good at and yet humans are notoriously bad at, because we are swayed by emotional responses and it is easy to have an emotional response to text that is programmed to look good for you and you alone.
> you communicate 1000 bits of information to me, instead of planned 500, and you yourself learn extra 500 bits.
Extremely optimistic. If this were the ideal scenario, you would USE the LLM to garner information ABOUT those 500 bits and then reframe them in a way that you yourself would put it. If there is insight, your "word" in your mental register now expands from the original 1000 bits to 1500 or 2000, and then are "processed" by your human brain that includes subconscious choices that are meaningful to the end result. There are tons of hidden semiotic data in your diction and wording (think resource forks in classic MacOS/HFS, only visible to the filesys) that is lost when you rely on another source to put together words for you; it's as if it is a game of Telephone. These are subtleties which you may intend for your recipient to receive and which are crucially important to your recipient and are irretrievable, it is intrinsically lossy. You have an alphabet soup of words, they cannot be put together by an LLM in exactly the way your brain did. We must rely on the fact that we ourselves put this together, the "aha" moment when an LLM does it for you is illusory and does not itself provide meaningfully important confirmation that you indeed say what you mean to say. Of course, humans say things and put things in way we do not intend to all the time. I still fundamentally believe this is more honest than relying on a third party that is not capable of understanding human emotional nuance to put together language for you, when language is and always has been a manner in which to dictate human emotional nuance.
> (Excuse the strong language, but I spent a lot of effort every day on both dealing with inconsiderate people lobbing LLM output at me, and making sure never to act like one myself, so it's a topic close to my heart.)
Does not negate the fact that LLM output itself, at least when used to convey human emotions or thoughts, is lossy. A very highly compressed JPEG with added interpolation from an upscale algorithm might come up with cool details that were never present in the original, and may look cool to both you and the recipient but are not honest to the source material. You receiving 240p JPEGs on a day-to-day basis is irrelevant to this. For the purposes of communication, it is a massive error which has the potential to compound, regardless of whether or not you or the recipient believe this to be the case.
> If this were the ideal scenario, you would USE the LLM to garner information ABOUT those 500 bits and then reframe them in a way that you yourself would put it.
Yes, but at that point in practice we're getting into over-optimizing territory. In this optimistic case I presented, you could learn those 500 bits yourself and formulate a clean message yourself, with all 1000 bits in it, but since LLM already gave you the text, and you feel you vouch for, you may as well send it over and save yourself the effort.
In reality the numbers are probably lower, and writing the message yourself is IMO also a good way to be truly sure you vouch for the "extra" 500 bits, as it forces you to actually pay attention. There's a chance you'll find inconsistency in output, or in your own understanding. I don't begrudge people for eventually cutting the process off here, for practical reasons - it's the fuzzy line between accuracy and perfectionism.
> A very highly compressed JPEG with added interpolation from an upscale algorithm might come up with cool details that were never present in the original, and may look cool to both you and the recipient but are not honest to the source material.
Again, I think it's a wrong take. LLMs aren't pulling the information out of their asses, and you are also not able to express every information directly. LLM can "upscale" information and you can take a look and recognize, "yes, this is exactly as it was", even without being able to write out that "upscaled" version by yourself. Verification is often easier than direct recall.
How often do you modify the LLM output before sending it? If it’s less than 50% of the time, it means that by not modifying it, you have added at most one bit of information to what you originally wrote. (If you don’t understand why, think of it this way: Instead of sending the LLM response, you could send the prompt and one extra bit indicating whether the LLM response to the prompt should be modified, followed by the modifications.)
> How often do you modify the LLM output before sending it?
Me specifically, I never send anyone LLM output I haven't give at least a quick read (not skim, read) to make sure it's reasonable and there is no obvious bullshit there. And then I still mention it's LLM-sourced.
> If it’s less than 50% of the time, it means that by not modifying it, you have added at most one bit of information to what you originally wrote. (...) Instead of sending the LLM response, you could send the prompt and one extra bit indicating whether the LLM response to the prompt should be modified, followed by the modifications.
It's not the case, though. Prompts are not interchangeable with output. There is no guarantee that if you send a prompt, and recipient passes it to their LLM, they'll receive anything similar to what you did. It may have mistakes - different mistakes - or just spend focus differently.
The extra bits I claim LLMs can add to the message hinge strictly on you vouching for the response. Of course, you can just prompt an LLM, learn from the response, and then write your message clean, containing both the bits you originally had, and the bits you gained. But at that point, the LLM already gave you text containing all those bits - if you can vouch for it, you may as well copy it over and save yourself the trouble.
> It's not the case, though. Prompts are not interchangeable with output. There is no guarantee that if you send a prompt, and recipient passes it to their LLM, they'll receive anything similar to what you did. It may have mistakes - different mistakes - or just spend focus differently.
I’m not saying that the response is interchangeable, but that due to the data processing inequality, it cannot convey strictly more information than the prompt.
> The extra bits I claim LLMs can add to the message hinge strictly on you vouching for the response.
My argument is that if you vouch at least 50% of the time, the vouching only adds one bit of useful information – either you vouch or not.
> due to the data processing inequality, it cannot convey strictly more information than the prompt.
Only in the case where the LLM message is not reviewed before sending, and only if we assume reliable LLM (so that the receiver could recreate the same output if given the original prompt). This is not a realistic scenario.
> My argument is that if you vouch at least 50% of the time, the vouching only adds one bit of useful information – either you vouch or not.
The alternative to vouching isn't "not vouching", but "correcting and cutting out wrong bits and vouching for the rest", which means the single "vouched for it" adds all the bits that are in final message but weren't there in the prompt.
I think you missunderstand. If the llm can "add back" 200 bits of information, these 200 bits are superfluous by definition.
They can quiet literally be inferred from the starting 300 bits. And the llm has proved it.
If the receiver wanted these 200 added bits, she could infer them either herself or even use an llm to do it.
> If the llm can "add back" 200 bits of information, these 200 bits are superfluous by definition.
No, they're not. Getting those bits takes energy.
SOTA LLMs know way more than any individual on approximately anything there is to know (and what they don't, they can look up faster than people can). It's very easy for them to make the "missing" 200 bits explicit, rather than implicit, which in practical terms is the same as adding 200 bits that weren't there before.
Theoretically, an idealized omnipotent mind / AGI could derive the unifying theory from reading your HN comment on a phone screen. There is enough information there, if you were able to extract every bit of evidence available from it. But you are not. Neither am I. It would take us practically infinite work to try, solving this most cruel mathematical riddle.
In principle, all of mathematics can be inferred from the axioms. The field of mathematics is not superfluous and something a receiver could infer themselves if they wanted it.
My brain does not contain all the information that can be added by an LLM; a human brain could contain it, even the biggest LLMs are about 1% of the (if you approximate synaptic count ~= parameters) parameter count of a human brain, but none actually will.
What my brain may actually contain is the information necessary to verify the (in this example) more than 200 bits the LLM claims to have added and trim out the parts which are false, retaining the (in this example) 200 "new" bits of new information added by the LLM.*
Concrete example: I am a software developer by training, though not a web developer. If someone who does not have any developer experience asks me to make a web app, I am forced to use an LLM as I do not know enough JS etc syntax to get it done myself. But as we all know, LLMs are only "ok" but not "good" at making software, so there are a lot of rough edges and outright mistakes. My experience as a software developer extends to detecting such failures and I can usually correct them.
The original person, someone who has no developer experience, can also prompt the LLM. Right now, this would result in something that retains all the errors, because they didn't have someone like me intermediating between them and the LLM.
I add bits by removing noise, the LLM adds bits but they contain noise.
I do not know for how long this will remain true, but today it is true.
* Feels like P versus NP to me. The answers AI generate are at their best when they're easy to verify. Then again, when they're easy to verify, they can be RLed to get good at this quickly and the need to verify goes down, leaving them still pretty bad at things that are hard to verify.
> I spent a lot of effort every day on both dealing with inconsiderate people lobbing LLM output at me, and making sure never to act like one myself
So why are you so eager to defend your fantastical scenario? It doesn’t matter how considerate you are, truth is the overwhelming majority of people aren’t and won’t be. We’re discussing reality here, not “what could be if we lived in a utopia which will never come to pass”.
The optimistic case is "having 500 bits, giving LLM 300, getting back 1000, and learning extra 500 in the process". Real numbers are lower. People don't vouch thoroughly and don't catch all mistakes.
But reasonable people don't send every output from LLMs to others without giving it a cursory glance (obvious hallucinations or nonsense would paint the sender as incompetent or inconsiderate), and that alone eliminates the worst levels of noise. A cursory read and cutting out obvious bullshit before sending is enough to make the message carry more bits of information than the propmpt.
> the overwhelming majority of people aren’t and won’t be.
In my experience, the "overwhelming majority" are giving something between a cursory glance and cursory edit; whether the resulting message has more or less information than prompt then depends on how much noise LLM added on top. The inconsiderate people I deal with, they often send "net more bits than in prompt" outputs, but those outputs are also verbose and not fully filtered for bullshit, thus it's effortful to tease out the signal from noise.
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.
TeMPOraL · · focus · HN ↗
Sure you can. LLM doesn't know what those 700 bits are, but you do. You may not realize it, and may not even know it at the time of prompting, but you do by the time you're sending.
Typical case is like this: you have 500 bits of semantic information to transfer. You give 300 of them to LLM, and get back the 500 bits you knew you have, and extra 500 you can quickly confirm are correct and relevant. Some of them are just dereferences of your input - where you recalled a pointer, but not what it pointed to. Some of it is information you never had before, but are able to easily validate.
You send that to me. I likely immediately realize the message was AI-assisted, but I trust you to be a decent human being, and not an asshole that lobs unverified LLM vomit over the fence for others to deal with. End result: you communicate 1000 bits of information to me, instead of planned 500, and you yourself learn extra 500 bits.
This is the optimistic scenario, but it does happen when LLM operator is not an asshole.
(Excuse the strong language, but I spent a lot of effort every day on both dealing with inconsiderate people lobbing LLM output at me, and making sure never to act like one myself, so it's a topic close to my heart.)
kunai · · focus · HN ↗
This is exactly the use case an LLM might (huge emphasis on might, depends on workflow, agentic vs. relying on contextual which can hallucinate) be good at and yet humans are notoriously bad at, because we are swayed by emotional responses and it is easy to have an emotional response to text that is programmed to look good for you and you alone.
> you communicate 1000 bits of information to me, instead of planned 500, and you yourself learn extra 500 bits.
Extremely optimistic. If this were the ideal scenario, you would USE the LLM to garner information ABOUT those 500 bits and then reframe them in a way that you yourself would put it. If there is insight, your "word" in your mental register now expands from the original 1000 bits to 1500 or 2000, and then are "processed" by your human brain that includes subconscious choices that are meaningful to the end result. There are tons of hidden semiotic data in your diction and wording (think resource forks in classic MacOS/HFS, only visible to the filesys) that is lost when you rely on another source to put together words for you; it's as if it is a game of Telephone. These are subtleties which you may intend for your recipient to receive and which are crucially important to your recipient and are irretrievable, it is intrinsically lossy. You have an alphabet soup of words, they cannot be put together by an LLM in exactly the way your brain did. We must rely on the fact that we ourselves put this together, the "aha" moment when an LLM does it for you is illusory and does not itself provide meaningfully important confirmation that you indeed say what you mean to say. Of course, humans say things and put things in way we do not intend to all the time. I still fundamentally believe this is more honest than relying on a third party that is not capable of understanding human emotional nuance to put together language for you, when language is and always has been a manner in which to dictate human emotional nuance.
> (Excuse the strong language, but I spent a lot of effort every day on both dealing with inconsiderate people lobbing LLM output at me, and making sure never to act like one myself, so it's a topic close to my heart.)
Does not negate the fact that LLM output itself, at least when used to convey human emotions or thoughts, is lossy. A very highly compressed JPEG with added interpolation from an upscale algorithm might come up with cool details that were never present in the original, and may look cool to both you and the recipient but are not honest to the source material. You receiving 240p JPEGs on a day-to-day basis is irrelevant to this. For the purposes of communication, it is a massive error which has the potential to compound, regardless of whether or not you or the recipient believe this to be the case.
kunai · · focus · HN ↗
[dead]
TeMPOraL · · focus · HN ↗
Yes, but at that point in practice we're getting into over-optimizing territory. In this optimistic case I presented, you could learn those 500 bits yourself and formulate a clean message yourself, with all 1000 bits in it, but since LLM already gave you the text, and you feel you vouch for, you may as well send it over and save yourself the effort.
In reality the numbers are probably lower, and writing the message yourself is IMO also a good way to be truly sure you vouch for the "extra" 500 bits, as it forces you to actually pay attention. There's a chance you'll find inconsistency in output, or in your own understanding. I don't begrudge people for eventually cutting the process off here, for practical reasons - it's the fuzzy line between accuracy and perfectionism.
> A very highly compressed JPEG with added interpolation from an upscale algorithm might come up with cool details that were never present in the original, and may look cool to both you and the recipient but are not honest to the source material.
Again, I think it's a wrong take. LLMs aren't pulling the information out of their asses, and you are also not able to express every information directly. LLM can "upscale" information and you can take a look and recognize, "yes, this is exactly as it was", even without being able to write out that "upscaled" version by yourself. Verification is often easier than direct recall.
xigoi · · focus · HN ↗
TeMPOraL · · focus · HN ↗
Me specifically, I never send anyone LLM output I haven't give at least a quick read (not skim, read) to make sure it's reasonable and there is no obvious bullshit there. And then I still mention it's LLM-sourced.
> If it’s less than 50% of the time, it means that by not modifying it, you have added at most one bit of information to what you originally wrote. (...) Instead of sending the LLM response, you could send the prompt and one extra bit indicating whether the LLM response to the prompt should be modified, followed by the modifications.
It's not the case, though. Prompts are not interchangeable with output. There is no guarantee that if you send a prompt, and recipient passes it to their LLM, they'll receive anything similar to what you did. It may have mistakes - different mistakes - or just spend focus differently.
The extra bits I claim LLMs can add to the message hinge strictly on you vouching for the response. Of course, you can just prompt an LLM, learn from the response, and then write your message clean, containing both the bits you originally had, and the bits you gained. But at that point, the LLM already gave you text containing all those bits - if you can vouch for it, you may as well copy it over and save yourself the trouble.
xigoi · · focus · HN ↗
I’m not saying that the response is interchangeable, but that due to the data processing inequality, it cannot convey strictly more information than the prompt.
> The extra bits I claim LLMs can add to the message hinge strictly on you vouching for the response.
My argument is that if you vouch at least 50% of the time, the vouching only adds one bit of useful information – either you vouch or not.
TeMPOraL · · focus · HN ↗
Only in the case where the LLM message is not reviewed before sending, and only if we assume reliable LLM (so that the receiver could recreate the same output if given the original prompt). This is not a realistic scenario.
> My argument is that if you vouch at least 50% of the time, the vouching only adds one bit of useful information – either you vouch or not.
The alternative to vouching isn't "not vouching", but "correcting and cutting out wrong bits and vouching for the rest", which means the single "vouched for it" adds all the bits that are in final message but weren't there in the prompt.
cwassert · · focus · HN ↗
If the receiver wanted these 200 added bits, she could infer them either herself or even use an llm to do it.
TeMPOraL · · focus · HN ↗
No, they're not. Getting those bits takes energy.
SOTA LLMs know way more than any individual on approximately anything there is to know (and what they don't, they can look up faster than people can). It's very easy for them to make the "missing" 200 bits explicit, rather than implicit, which in practical terms is the same as adding 200 bits that weren't there before.
Theoretically, an idealized omnipotent mind / AGI could derive the unifying theory from reading your HN comment on a phone screen. There is enough information there, if you were able to extract every bit of evidence available from it. But you are not. Neither am I. It would take us practically infinite work to try, solving this most cruel mathematical riddle.
ben_w · · focus · HN ↗
My brain does not contain all the information that can be added by an LLM; a human brain could contain it, even the biggest LLMs are about 1% of the (if you approximate synaptic count ~= parameters) parameter count of a human brain, but none actually will.
What my brain may actually contain is the information necessary to verify the (in this example) more than 200 bits the LLM claims to have added and trim out the parts which are false, retaining the (in this example) 200 "new" bits of new information added by the LLM.*
Concrete example: I am a software developer by training, though not a web developer. If someone who does not have any developer experience asks me to make a web app, I am forced to use an LLM as I do not know enough JS etc syntax to get it done myself. But as we all know, LLMs are only "ok" but not "good" at making software, so there are a lot of rough edges and outright mistakes. My experience as a software developer extends to detecting such failures and I can usually correct them.
The original person, someone who has no developer experience, can also prompt the LLM. Right now, this would result in something that retains all the errors, because they didn't have someone like me intermediating between them and the LLM.
I add bits by removing noise, the LLM adds bits but they contain noise.
I do not know for how long this will remain true, but today it is true.
* Feels like P versus NP to me. The answers AI generate are at their best when they're easy to verify. Then again, when they're easy to verify, they can be RLed to get good at this quickly and the need to verify goes down, leaving them still pretty bad at things that are hard to verify.
latexr · · focus · HN ↗
> This is the optimistic scenario
So is it typical or optimistic?
> I spent a lot of effort every day on both dealing with inconsiderate people lobbing LLM output at me, and making sure never to act like one myself
So why are you so eager to defend your fantastical scenario? It doesn’t matter how considerate you are, truth is the overwhelming majority of people aren’t and won’t be. We’re discussing reality here, not “what could be if we lived in a utopia which will never come to pass”.
TeMPOraL · · focus · HN ↗
The optimistic case is "having 500 bits, giving LLM 300, getting back 1000, and learning extra 500 in the process". Real numbers are lower. People don't vouch thoroughly and don't catch all mistakes.
But reasonable people don't send every output from LLMs to others without giving it a cursory glance (obvious hallucinations or nonsense would paint the sender as incompetent or inconsiderate), and that alone eliminates the worst levels of noise. A cursory read and cutting out obvious bullshit before sending is enough to make the message carry more bits of information than the propmpt.
> the overwhelming majority of people aren’t and won’t be.
In my experience, the "overwhelming majority" are giving something between a cursory glance and cursory edit; whether the resulting message has more or less information than prompt then depends on how much noise LLM added on top. The inconsiderate people I deal with, they often send "net more bits than in prompt" outputs, but those outputs are also verbose and not fully filtered for bullshit, thus it's effortful to tease out the signal from noise.