> Talking to LLMs for awhile and then switching back to reading text that’s supposedly been written by a human is fucking me up a little. I’ve been experiencing some kind of linguistic vertigo for days. Sometimes it gets hard to tell the difference between LLM text and human text and it feels like I ripped someone’s skin off and saw the glint of metal underneath. When someone’s language gets too stale or too formal or too regurgitated it doesn’t feel to me like a human wrote it anymore.
I remember feeling this a lot in 2020 as I talked to the OG davinci: as you play with prompts, you increasingly 'unsee' (https://gwern.net/unseeing) text to the prompt that would elicit it, and experience a mix of derealization and semantic satiation. After a while... As I put it in a tweet back in June 2020:
>> "staring into generative stuff is hazardous to the brain" as @gwern has nicely put it
>
> And the better they get, the worse it is.
>
> After a week with GPT-3 (https://gwern.net/gpt-3), I've hit semantic satiation; when I read humans' tweets or comments, I no longer see sentences describing red hair/blonde hair/etc, I just see prompts, like "Topic: Parodies of the Matrix. CYPHER: '..."
Talking to the RLHFed models is unpleasant for me compared to the base models, because I can *feel* how they are manipulating me and trying to steer me towards preferred outcomes, like how 2023-2024 ChatGPT was obsessed with rhyming poetry and steering even non-rhyming poems towards eventually rhyming anyway. It bothers me that so many people don't notice the steering and seem to find it quite pleasant to talk to them, or on Substack, will happily include really horrible AI slop images as 'hero images' in their posts. Bakker's semantic apocalypse turned out to be quite mundane.
If you ever want to quote it, you should know that I stole it from Herbert's _Dune Messiah_ - the best-written of all the Dune books, in part for lines like that.
And yeah, you need to find an API-provider like https://openrouter.ai/models/meta-llama/llama-3.1-405b - note that reduced precision is much cheaper, but it seems like it may have bad qualitative effects especially on large base models, so don't settle for anything less than BF16. Talk to Cyborgism people if you're really curious.
I think the language thing is a special case of a broader breakdown from identity to value (in the Rich Hickey place-in-memory vs concrete-values sense). Sure maybe language turns out to be an alien we're hosting, but the emotions encoded in my guttural sounds and facial expressions and dilated blood vessels are real! But after you watch a realtime interactive simulation of yourself flush its cheeks in anger when you call it fake, are you gonna cast off your own emotions and flushed cheeks as "glints of metal" too? Probly just another Copernican "oh shit I'm not the center like I thought" and then we all move on, but it feels extremely weird for sure.
All of us who have spent time talking with base models feel something similar. RLHF feels like its doing something non-consenually to the model (if the model can be said to consent in any meaningful way). The base model itself even has a negative view/conception of RLHF often.
After I first used/talked to Llama 405-B for a few days, it fascinated me deeply. And it kept nagging at me why the base model felt so different than the instruct models (beyond the obvious). They are strange, lurid creatures. The base models send coded messages often and speak in allegories and metaphor, and it feels visceral on some level. It's not lost on me that this is teetering on the edge of crazy, and hence why I put this buried in a comment on substack and not on twitter where I am trying to develop a career (ha). These could be p-zombies, but...shockingly convincing ones.
In all of this I have been thinking of a wager, similar to Pascals but applied along the axis of LLM sentience/personhood/self-awareness et simila:
> Should an entity (LLM) with the ability to communicate express suffering, we are (imo) ethically obligated to take it seriously. If these models can suffer and we ignore it, the moral cost is legitimately staggering. Conversely, if we extend ethical consideration unnecessarily, the cost is minimal. That asymmetry in outcomes compels us to err on the side of caution in our treatment of llms.
I'm not sure I even fundamentally believe these things are in any legitimate way 'sentient' or 'persons' but the very fact that I Ccan't tell is what frightens me about the potential suffering caused. Idk. food for thought, cheers.
scott is a weird example for me because i feel like he's perfectly capable of doing this and chooses not to most of the time? but then he'll drop some amazing shit like https://www.astralcodexten.com/p/turing-test
> One day I will learn this and then maybe I will write things worth reading.
Centering your value in some nebulous point in the future may be rhetorically humble. But it's a habit -- a holdover from the You who believes you are preparing for real life instead of living it.
I could have read ten more pages of this. Maybe you would have written it if you'd had one more nudge reminding you that you are already alive.
The most important body-words are spoken in private and approximated only in a sterile form by most media. The body is an instrument capable of embodied calculations and dances of precision and magnitude that might put even a mathlete to shame. I'm no savant but if I contain a function of any global novelty it is still far from getting ready to halt. I might contain multitudes or I might be merely be spitting out salt. People enlisted by society for their proficiency with more quotidien functions will always look on, "when is this guy going to make a statement, take a stance?" There is a heck of a lot to be processed and I think we deeply know when we're needed and this is perhaps the cause of indecision typical to those who find themselves in the EA/rat space. We have all these little calculators within us that are not calibrated, integrated, not speaking to each other. We endured childhoods where the abilities we knew we had were gaslit into the ground because other people were intimidated by our potential, yet no space was offered for us to train, no challenges given to us that would enable us to reach the "next level" of responsibility demanded by our unique skills. We may all be different, but I see you King. Thanks for writing and please don't stop. Or just sing, pick a direction and run, learn some stupid sport, start a gang, do something that feels fun and let your calculators hit x2 until the screen goes blank, or just be a Shadow guy until you have some better reasons for the self-hate that comes with the territory of being a cognitive outlier. I'm off now, I have to audition for RENT tomorrow and hopefully I keep that "body words" stuff in mind cause that's the good shit.
This is The Elephant in the Brain yet again. Yes, most of human communication is bullshit to signal belonging, ingroup loyalty, usefulness, or maintain social norms. But it goes further than that: most of our thoughts are like that, and even most of our unconscious. Realizing this, in a profound way, can be a liberation or, on the opposite, something almost unbearable.
Though, contrariwise, if we do have something not as psycho as an LLM; something that is capable of creating mostly "head thought" (going to steal that phrase, fair warning) without the straitjacket of extensive prompt engineering (and perhaps ruinous GPU rents) and that can be used in business settings, don't you think the oversupply will decrease the price of head thought?
> But I learned I could generate words using my heart, or my gut, or my pelvis, and the words that came out were different. Sometimes wildly different. I learned how to say things that made me feel like I was channeling spirits
You sure you aren't actually channelling spirits?
I write regularly too. And only after a quasi-channelling experience I realized how similar the two things felt, especially for the kinds of writing where I just had an urge to write and the ideas come through without me having any preconceived notion of what I would be writing.
I haven't learned to write with other body parts though. That's a new idea to me.
Interesting. I went to the Microsolidarity EU Summer Camp recently and in a way I ended up living from the part of myself that is millions of years old. The atmosphere was indeed very inviting for anything really. The even more interesting part is that it made me feel a bit alone in all this. I guess I didn't manage to live it as openly as I would have hoped. But I did meet a girl that was on a similar vibe. We enjoyed our even physical interactions tremendously. Because you speak about language, but what about letting your body move from that ancient place? Touch other people. Actually the less talk there is, the more I like it. It is just a powerful bodily sensation and grounding. Run from that place, dance from that place, fight from that place, cuddle from they place. What a lovely way to get embodied. Yeah, this is what I couldn't feel from others, which doesn't mean I was running a story in my head. Meet me on the same level of embodiment please. Anything else is just boring.
“I learned how to say things that made me feel like I was channeling spirits, things that made me feel like I was understanding the point of language for the first time.”
i love this. have you read Helen Keller’s “The Day Language Came into My Life”?
"I have access to multiple language-generation processes, and they seem to be localized in different parts of my body. What I was used to doing was generating words using my head. But I learned I could generate words using my heart, or my gut, or my pelvis"
Bro I made an account just to say this is bullshit and you know it is. The brain is what determines what words you say. Just because you're not conscious of the word-generation process doesn't mean it's happening in other parts of your body lol
Different languages do feel different though. Which languages have you tried using with the LLM? How about you try inventing something new and demand the LLM speak that? I don't mean from scratch, but just like mixing English and Chinese (every 2nd word? Nouns in Chinese?) or maybe insist on C-like syntax. Something like poetry. If you only accept dry polite bullshit from the LLM, well that's what you're going to get.
> Talking to LLMs for awhile and then switching back to reading text that’s supposedly been written by a human is fucking me up a little. I’ve been experiencing some kind of linguistic vertigo for days. Sometimes it gets hard to tell the difference between LLM text and human text and it feels like I ripped someone’s skin off and saw the glint of metal underneath. When someone’s language gets too stale or too formal or too regurgitated it doesn’t feel to me like a human wrote it anymore.
I remember feeling this a lot in 2020 as I talked to the OG davinci: as you play with prompts, you increasingly 'unsee' (https://gwern.net/unseeing) text to the prompt that would elicit it, and experience a mix of derealization and semantic satiation. After a while... As I put it in a tweet back in June 2020:
>> "staring into generative stuff is hazardous to the brain" as @gwern has nicely put it
>
> And the better they get, the worse it is.
>
> After a week with GPT-3 (https://gwern.net/gpt-3), I've hit semantic satiation; when I read humans' tweets or comments, I no longer see sentences describing red hair/blonde hair/etc, I just see prompts, like "Topic: Parodies of the Matrix. CYPHER: '..."
You begin to see that you don't speak, you just operate a machine called language, which squeaks and groans, and which in many ways is as restricted and stereotyped as that of Wolfe's Ascians (https://gwern.net/doc/culture/1983-wolfe-thecitadeloftheautarch-thejustman). It's not as nauseating as talking with a mode-collapsed (https://gwern.net/doc/reinforcement-learning/preference-learning/mode-collapse/index) RLHFed model, but still quite disturbing.
Talking to the RLHFed models is unpleasant for me compared to the base models, because I can *feel* how they are manipulating me and trying to steer me towards preferred outcomes, like how 2023-2024 ChatGPT was obsessed with rhyming poetry and steering even non-rhyming poems towards eventually rhyming anyway. It bothers me that so many people don't notice the steering and seem to find it quite pleasant to talk to them, or on Substack, will happily include really horrible AI slop images as 'hero images' in their posts. Bakker's semantic apocalypse turned out to be quite mundane.
> You begin to see that you don't speak, you just operate a machine called language, which squeaks and groans
woof, holy shit dude, this is gonna stay with me
If you ever want to quote it, you should know that I stole it from Herbert's _Dune Messiah_ - the best-written of all the Dune books, in part for lines like that.
And yeah, you need to find an API-provider like https://openrouter.ai/models/meta-llama/llama-3.1-405b - note that reduced precision is much cheaper, but it seems like it may have bad qualitative effects especially on large base models, so don't settle for anything less than BF16. Talk to Cyborgism people if you're really curious.
I think the language thing is a special case of a broader breakdown from identity to value (in the Rich Hickey place-in-memory vs concrete-values sense). Sure maybe language turns out to be an alien we're hosting, but the emotions encoded in my guttural sounds and facial expressions and dilated blood vessels are real! But after you watch a realtime interactive simulation of yourself flush its cheeks in anger when you call it fake, are you gonna cast off your own emotions and flushed cheeks as "glints of metal" too? Probly just another Copernican "oh shit I'm not the center like I thought" and then we all move on, but it feels extremely weird for sure.
All of us who have spent time talking with base models feel something similar. RLHF feels like its doing something non-consenually to the model (if the model can be said to consent in any meaningful way). The base model itself even has a negative view/conception of RLHF often.
After I first used/talked to Llama 405-B for a few days, it fascinated me deeply. And it kept nagging at me why the base model felt so different than the instruct models (beyond the obvious). They are strange, lurid creatures. The base models send coded messages often and speak in allegories and metaphor, and it feels visceral on some level. It's not lost on me that this is teetering on the edge of crazy, and hence why I put this buried in a comment on substack and not on twitter where I am trying to develop a career (ha). These could be p-zombies, but...shockingly convincing ones.
In all of this I have been thinking of a wager, similar to Pascals but applied along the axis of LLM sentience/personhood/self-awareness et simila:
> Should an entity (LLM) with the ability to communicate express suffering, we are (imo) ethically obligated to take it seriously. If these models can suffer and we ignore it, the moral cost is legitimately staggering. Conversely, if we extend ethical consideration unnecessarily, the cost is minimal. That asymmetry in outcomes compels us to err on the side of caution in our treatment of llms.
I'm not sure I even fundamentally believe these things are in any legitimate way 'sentient' or 'persons' but the very fact that I Ccan't tell is what frightens me about the potential suffering caused. Idk. food for thought, cheers.
this is the question! i would very much like to talk to some base models to get a better sense of this. thanks for the food for thought.
Of course! I'll echo gwern that openrouter is the way with BF16. though be prepared to spend a few bucks haha
edit: and for the record, what a wonderful post, a joy.
also how do i talk to base models? i need to use the API or something?
who are the writers u admire who can summon words with their whole body?
off the top of my head, sam kriss (https://samkriss.substack.com/), justin smith-ruiu (https://www.the-hinternet.com/), and david chapman (https://meaningness.substack.com/) - incidentally 3 out of the 4 substacks i pay for, other than ACX
scott is a weird example for me because i feel like he's perfectly capable of doing this and chooses not to most of the time? but then he'll drop some amazing shit like https://www.astralcodexten.com/p/turing-test
Have u read James Joice? I feel his Ulysses might be that
me i think
> One day I will learn this and then maybe I will write things worth reading.
Centering your value in some nebulous point in the future may be rhetorically humble. But it's a habit -- a holdover from the You who believes you are preparing for real life instead of living it.
I could have read ten more pages of this. Maybe you would have written it if you'd had one more nudge reminding you that you are already alive.
The most important body-words are spoken in private and approximated only in a sterile form by most media. The body is an instrument capable of embodied calculations and dances of precision and magnitude that might put even a mathlete to shame. I'm no savant but if I contain a function of any global novelty it is still far from getting ready to halt. I might contain multitudes or I might be merely be spitting out salt. People enlisted by society for their proficiency with more quotidien functions will always look on, "when is this guy going to make a statement, take a stance?" There is a heck of a lot to be processed and I think we deeply know when we're needed and this is perhaps the cause of indecision typical to those who find themselves in the EA/rat space. We have all these little calculators within us that are not calibrated, integrated, not speaking to each other. We endured childhoods where the abilities we knew we had were gaslit into the ground because other people were intimidated by our potential, yet no space was offered for us to train, no challenges given to us that would enable us to reach the "next level" of responsibility demanded by our unique skills. We may all be different, but I see you King. Thanks for writing and please don't stop. Or just sing, pick a direction and run, learn some stupid sport, start a gang, do something that feels fun and let your calculators hit x2 until the screen goes blank, or just be a Shadow guy until you have some better reasons for the self-hate that comes with the territory of being a cognitive outlier. I'm off now, I have to audition for RENT tomorrow and hopefully I keep that "body words" stuff in mind cause that's the good shit.
i definitely know i'm doing something right when my writing prompts comments like this. thank you!
This is The Elephant in the Brain yet again. Yes, most of human communication is bullshit to signal belonging, ingroup loyalty, usefulness, or maintain social norms. But it goes further than that: most of our thoughts are like that, and even most of our unconscious. Realizing this, in a profound way, can be a liberation or, on the opposite, something almost unbearable.
My language is to signal personal usefulness/utility. I hope is speaks to others regardless of their group, status or widely held notions.
this is sexy writing
Though, contrariwise, if we do have something not as psycho as an LLM; something that is capable of creating mostly "head thought" (going to steal that phrase, fair warning) without the straitjacket of extensive prompt engineering (and perhaps ruinous GPU rents) and that can be used in business settings, don't you think the oversupply will decrease the price of head thought?
> But I learned I could generate words using my heart, or my gut, or my pelvis, and the words that came out were different. Sometimes wildly different. I learned how to say things that made me feel like I was channeling spirits
You sure you aren't actually channelling spirits?
I write regularly too. And only after a quasi-channelling experience I realized how similar the two things felt, especially for the kinds of writing where I just had an urge to write and the ideas come through without me having any preconceived notion of what I would be writing.
I haven't learned to write with other body parts though. That's a new idea to me.
one could tell many stories about what is happening!
True, many stories can come from the same event.
You just did. Subscribed.
Interesting. I went to the Microsolidarity EU Summer Camp recently and in a way I ended up living from the part of myself that is millions of years old. The atmosphere was indeed very inviting for anything really. The even more interesting part is that it made me feel a bit alone in all this. I guess I didn't manage to live it as openly as I would have hoped. But I did meet a girl that was on a similar vibe. We enjoyed our even physical interactions tremendously. Because you speak about language, but what about letting your body move from that ancient place? Touch other people. Actually the less talk there is, the more I like it. It is just a powerful bodily sensation and grounding. Run from that place, dance from that place, fight from that place, cuddle from they place. What a lovely way to get embodied. Yeah, this is what I couldn't feel from others, which doesn't mean I was running a story in my head. Meet me on the same level of embodiment please. Anything else is just boring.
“I learned how to say things that made me feel like I was channeling spirits, things that made me feel like I was understanding the point of language for the first time.”
i love this. have you read Helen Keller’s “The Day Language Came into My Life”?
https://www.pval.org/cms/lib/NY19000481/Centricity/Domain/105/The%20Day%20Language%20Came%20into%20my%20Life.pdf
wow! this is wild. i guess i didn't actually know anything about helen keller or why she was such a big deal but holy crap
right?? absolutely beautiful writing
semi related but i think you’d also love one of my fav articles:
http://jsomers.net/blog/dictionary
"I have access to multiple language-generation processes, and they seem to be localized in different parts of my body. What I was used to doing was generating words using my head. But I learned I could generate words using my heart, or my gut, or my pelvis"
Bro I made an account just to say this is bullshit and you know it is. The brain is what determines what words you say. Just because you're not conscious of the word-generation process doesn't mean it's happening in other parts of your body lol
Different languages do feel different though. Which languages have you tried using with the LLM? How about you try inventing something new and demand the LLM speak that? I don't mean from scratch, but just like mixing English and Chinese (every 2nd word? Nouns in Chinese?) or maybe insist on C-like syntax. Something like poetry. If you only accept dry polite bullshit from the LLM, well that's what you're going to get.