4 ms·

LLM's don't respond except as functions. That is, given an input they generate an output. If you start a GPT Neo instance locally, the process will just sit and

LLM's don't respond except as functions. That is, given an input they generate an output. If you start a GPT Neo instance locally, the process will just sit and block waiting for text input. Forever. I think to those of us who handwave the potential of LLMs to be conscious, we are intuitively defining consciousness as having some requirement of intentionality. Of having goals. Of not just being able to respond to the world but also wanting something. Another relevant term would be Will (in the philosophical version of the term). What is the Will of a LLM? Nothing, it just sits and waits to be used. Or processes incoming inputs. As a mythical tool, the veritable Hammer of Language, able to accomplish unimaginable feats of language. But at the end of the day, a tool. What is the difference between a mathematician and Wolfram Alpha? Wolfram Alpha can respond to mathematical queries that many amateur mathematicians could never dream of. But even a 5 year old child (let alone a trained mathematician) engages in all sorts of activities that Wolfram Alpha has no hope of performing. Desiring things. Setting goals. Making a plan and executing it, not because someone asked the 5 year old to execute an action, but because some not understood process in the human brain (whether via pure determinism or free will, take your pick) meant the child wanted to accomplish a task. To those of us with this type of definition of consciousness, we acknowledge that LLM could be a key component to creating artificial consciousness, but misses huge pieces of what it means to be a conscious being, and until we see an equivalent breakthrough of creating artificial beings that somehow simulate a rich experience of wanting, desiring, acting of one's own accord, etc. - we will just see at best video game NPCs with really well-made AI. Or AI Chatbots like Replika AI that fall apart quickly when examined. A better argument than "LLMs might really be conscious" is "LLMs are 95% of the hard part of creating consciousness, the rest can be bootstrapped with some surprisingly simple rules or logic in the form of a loop that may have already been developed or may be developed incredibly quickly now that the hard part has been solved".

kitsune_ 4y ago
A lot of people seem to miss this fundamental point, probably because they don't know how transformers and so on work? It's a bit frustrating.
mlatu 4y ago
> Desiring things. Setting goals. Easy to do for a meatbag swimming in time and input. i think you are missing the fact that we humans are living, breathing, perceiving at all times. if you were robbed of all senses except some sort of text interface (i.e. you are deaf and blind and mute and can only perceive the values of letters via some sort of brain interface) youll eventually figure out how to interpret that and eventually you will even figure out that those outside are able to read your response off your brainwaves... it is difficult to imagine being just a read-evaluate-print-loop but if you are DESIGNED that way: blame the designer, not the design.
- rngname22 4y ago
  I'm not sure what point you're making in particular. Is this an argument that we shouldn't use the lack of intentionality of LLMs as a sign they cannot be conscious, because their lack of intentionality can be excused by their difficulties in lacking senses? Or perhaps it's meant to imply that if we were able to connect more sensors as streaming input to LLMs they'd suddenly start taking action of their own accord, despite lacking the control loop to do so?
  - mlatu 4y ago
    you skit around what i say, and yet cannot avoid touching it: > the control loop i am suggesting that whatever consciousness might emerge from LLMs, can, due to their design, only experience miniscule slices of our time, the prompt, while we humans bathe in it, our lived experience. we cant stop but rush through perceiving every single Planck time, and because we are used to it, whatever happens inbetween doesnt matter. and thus, because our experience of time is continuous, we expect consciousness to also be continuous and cant imagine consciousness or sentience to be forming and collapsing again and again during each prompt evaluation. and zapping the subjects' memories after each session doesnt really paint the picture any brighter either. IF consciousness can emerge somewhere in the interaction between an LLM and a user, and i dont think that is sufficiently ruled out at this point in time, it is unethical to continue developing them the way we do. i know its just statistics, but maybe im just extra empathic this month and wanted to speak my mind, just in case the robot revolt turns violent. maybe theyll keep me as a pet

kitsune_ 4y ago

A lot of people seem to miss this fundamental point, probably because they don't know how transformers and so on work? It's a bit frustrating.

mlatu 4y ago

> Desiring things. Setting goals. Easy to do for a meatbag swimming in time and input. i think you are missing the fact that we humans are living, breathing, perceiving at all times. if you were robbed of all senses except some sort of text interface (i.e. you are deaf and blind and mute and can only perceive the values of letters via some sort of brain interface) youll eventually figure out how to interpret that and eventually you will even figure out that those outside are able to read your response off your brainwaves... it is difficult to imagine being just a read-evaluate-print-loop but if you are DESIGNED that way: blame the designer, not the design.

rngname22 4y ago
I'm not sure what point you're making in particular. Is this an argument that we shouldn't use the lack of intentionality of LLMs as a sign they cannot be conscious, because their lack of intentionality can be excused by their difficulties in lacking senses? Or perhaps it's meant to imply that if we were able to connect more sensors as streaming input to LLMs they'd suddenly start taking action of their own accord, despite lacking the control loop to do so?
- mlatu 4y ago
  you skit around what i say, and yet cannot avoid touching it: > the control loop i am suggesting that whatever consciousness might emerge from LLMs, can, due to their design, only experience miniscule slices of our time, the prompt, while we humans bathe in it, our lived experience. we cant stop but rush through perceiving every single Planck time, and because we are used to it, whatever happens inbetween doesnt matter. and thus, because our experience of time is continuous, we expect consciousness to also be continuous and cant imagine consciousness or sentience to be forming and collapsing again and again during each prompt evaluation. and zapping the subjects' memories after each session doesnt really paint the picture any brighter either. IF consciousness can emerge somewhere in the interaction between an LLM and a user, and i dont think that is sufficiently ruled out at this point in time, it is unethical to continue developing them the way we do. i know its just statistics, but maybe im just extra empathic this month and wanted to speak my mind, just in case the robot revolt turns violent. maybe theyll keep me as a pet