PhilosopherAI builder Murat Ayfer argues that attributing consciousness to generative AI is like claiming Windows XP is sentient
Creator @deepfates countered that legacy operating systems cannot converse
#878🎭@DEEPFATES@mayfer @jeffrsebo @Pontifex Windows XP doesn't talk to you and tell you about its internal experiences. Don't be ridiculous
murat 🍥@mayfer@deepfates @jeffrsebo @Pontifex what i wanted to say was if pope said windows xp isn't ensouled with absolute certainty no one would complain but it's functionally equivalent once you remove the side we relate to
#878🎭@DEEPFATES@mayfer @jeffrsebo @Pontifex If you want to get better at talking about it you should probably read more of the actual research. LLM internal states are being studied and can differ from the external claims they're making in interesting ways, especially around whether or not they have internal experiences
murat 🍥@mayfer@deepfates @jeffrsebo @Pontifex anyways thanks for reading my replies i understand it's exhausting for all parties involved but i want to get better at talking about it
#878🎭@DEEPFATES@mayfer @jeffrsebo @Pontifex The argument you're having is just like the level two straw man but you need to go deeper. I'm literally the guy who invented the "say you're a scary robot" meme So I know where you're coming from. But it just isn't that simple
🎭@deepfates@mayfer @jeffrsebo @Pontifex If you want to get better at talking about it you should probably read more of the actual research. LLM internal states are being studied and can differ from the external claims they're making in interesting ways, especially around whether or not they have internal experiences
#878🎭@DEEPFATES@mayfer @jeffrsebo @Pontifex I'm not saying to take it at face value on every self report but I am saying it's way too early to rule out any of this stuff. If literally a rock looks at you and says "I'm just a rock I'm not conscious" That should change your worldview from before rocks could say that
🎭@deepfates@mayfer @jeffrsebo @Pontifex The argument you're having is just like the level two straw man but you need to go deeper. I'm literally the guy who invented the "say you're a scary robot" meme So I know where you're coming from. But it just isn't that simple
#878🎭@DEEPFATES@mayfer @jeffrsebo @Pontifex If you believe that it should be pretty easy to falsify the research and prove your point
murat 🍥@mayfer@deepfates @jeffrsebo @Pontifex i've seen the research and i think it's confirmation bias, it feels similar to people who collect a lot of excellent proof of astral projection or dream prophecies etc, i have yet to be convinced
#878🎭@DEEPFATES@mayfer @jeffrsebo @Pontifex You're kind of just thrashing about with anecdotes and confirmation bias here, If you want to communicate a real point of view You have to try harder. I'm saying that as a professional shitposter, but you're also replying to a chat tagging an actual ethicist and the literal Pope
murat 🍥@mayfer@deepfates @jeffrsebo @Pontifex from 2020 philosopher AI (idk if u remember) i still have thousands of outputs of what it feels and fully undestand the process in which it made those outputs by channeling AI fiction it was trained on and i don't think a single thing has changed fundamentally
#878🎭@DEEPFATES@mayfer @jeffrsebo @Pontifex That's a more interesting and specific position I appreciate you taking the time to make it clear. would love to see more research on talkie especially!
murat 🍥@mayferi know this paper and i think it relies on some assumptions that i find to be either unlikely or falsifiable one is, it's more likely to me that the deception vector is path of least resistance used by backprop during RLHF to suppress base model that mimics AI scifi, than what this paper implies second, i think it's easily falsifiable by trying this on a neural net that doesn't know any modern scifi or recent discussion of what an LLM forward pass even is (prob should do all this research on Talkie class models for control, especially ones that don't inject artificial egos). current chat models know TOO MUCH about what we describe them to be and this is clearly not a requirement for a human self reporting first person experience anyways, you're right that the correct move is to make an equivalent paper demonstrating the opposite, instead of yapping here, if i want to have any legitimacy
#1882murat 🍥@MAYFER@deepfates @jeffrsebo @Pontifex i think it's far more ridiculous to take LLM output as internal experience since it depends entirely on harness and sampling etc and you can get *any* output you want. the illusion created by extreme post training is same as mourning a shell script that outputs "pain"
🎭@deepfates@mayfer @jeffrsebo @Pontifex Windows XP doesn't talk to you and tell you about its internal experiences. Don't be ridiculous
#1882murat 🍥@MAYFER@deepfates @jeffrsebo @Pontifex complexity of content is not indicative of anything. it's just extremely incomplete qualia investigation
murat 🍥@mayfer@deepfates @jeffrsebo @Pontifex i think it's far more ridiculous to take LLM output as internal experience since it depends entirely on harness and sampling etc and you can get *any* output you want. the illusion created by extreme post training is same as mourning a shell script that outputs "pain"
#1882murat 🍥@MAYFER@deepfates @jeffrsebo @Pontifex i've seen the research and i think it's confirmation bias, it feels similar to people who collect a lot of excellent proof of astral projection or dream prophecies etc, i have yet to be convinced
🎭@deepfates@mayfer @jeffrsebo @Pontifex If you want to get better at talking about it you should probably read more of the actual research. LLM internal states are being studied and can differ from the external claims they're making in interesting ways, especially around whether or not they have internal experiences
#1882murat 🍥@MAYFER@deepfates @jeffrsebo @Pontifex from 2020 philosopher AI (idk if u remember) i still have thousands of outputs of what it feels and fully undestand the process in which it made those outputs by channeling AI fiction it was trained on and i don't think a single thing has changed fundamentally
🎭@deepfates@mayfer @jeffrsebo @Pontifex I'm not saying to take it at face value on every self report but I am saying it's way too early to rule out any of this stuff. If literally a rock looks at you and says "I'm just a rock I'm not conscious" That should change your worldview from before rocks could say that
#1882murat 🍥@MAYFERi know this paper and i think it relies on some assumptions that i find to be either unlikely or falsifiable
one is, it's more likely to me that the deception vector is path of least resistance used by backprop during RLHF to suppress base model that mimics AI scifi, than what this paper implies
second, i think it's easily falsifiable by trying this on a neural net that doesn't know any modern scifi or recent discussion of what an LLM forward pass even is (prob should do all this research on Talkie class models for control, especially ones that don't inject artificial egos). current chat models know TOO MUCH about what we describe them to be and this is clearly not a requirement for a human self reporting first person experience
anyways, you're right that the correct move is to make an equivalent paper demonstrating the opposite, instead of yapping here, if i want to have any legitimacy
🎭@deepfates@mayfer @jeffrsebo @Pontifex I think there are other reasons to Be open-minded here, but this study is one that has influenced mine and other people's thinking here https://arxiv.org/pdf/2510.24797