21 ms·

> My point was that we can't find out anything about the phsycial world by just discussing it. We each have to experience it ourselves using our own senses. Yo

> My point was that we can't find out anything about the phsycial world by just discussing it. We each have to experience it ourselves using our own senses. You cite the primacy of sensation, but what more is sensation than a message from the sensory organ to the brain. Perhaps encoded in the brains internal language rather than the plain text we are used to, but messages none the less. The concept of existence-as-messages is thus not contradicted by the experience of messages from your own body. Irrespective of any underlying physical reality, I can't escape the fact that I can never test reality itself, only my perception of it. We can get a lot of mileage out of the physicalist approach, accepting as an axiom that what we perceive as reality is real. But in that ground up approach we've had terrible difficulty deriving our ourselves from physical first principles. I'm not saying its impossible or wrong to take the path from atoms on upwards to consciousness and perception. But no one has been able to make it to the end of that path. What happens if we go the other way instead, starting from the known conclusion "I'm here and conscious enough to converse" and working our way down to the perception/understanding of an inanimate reality governed by objective physical principles? Can a system only designed to represent/reason about people talking to and about people be hacked into a system for representing and reasoning in general? If we start from purely language models like GPT3 and continue to teach it the "social reasoning" of saying things we want to hear, will the language model eventually become capable of non-social reasoning as well? In the process of figuring out what we want to hear well enough to describe a non-contradictory scene to us, does GPT3 have to actually learn the rules of 3d euclidean space governing the scene? Is there any possible way to avoid scene contradictions without a full understanding of the underlying physical reality it is supposed to describe to us? If a camera feeds data to a language model, does the language model suddenly have eyes? What about pictures from the internet?

tsimionescu 4y ago
> You cite the primacy of sensation, but what more is sensation than a message from the sensory organ to the brain. This is false, it's a version of the homunculus fallacy. A sensory organ is something that connects the brain to the physical world. Even if you chose to model it as a an agent that sends messages to the brain, it is an agent of a different nature. The sensory agent doesn't receive messages from other agents, it receives raw input from the outside world (photons, electrical fields, chemical reactions etc). In contrast, if you were a brain in a vat with no ability to directly perceive the world or interact with it in any way, it would be impossible for you to know that "I can pass through walls" is fundamentally impossible. Even for your GPT-3 thought experiment - ultimately it is the effects of the physical world perceived directly by humans sense organs that shape what GPT-3 would utter. That is, even if it can learn what the world is like simply by talking to us, it's still learning about the real world from someone's direct experience with it. If we were all GPT-3s, with no cameras and pressure sensors and motors etc, we would be unable to reason in any sense about the world itself. We could perhaps come to agree upon some imagined world, but that could change arbitrarily much from one day to the next on a whim.
- IIAOPSW 4y ago
  >A sensory organ is something that connects the brain to the physical world. Even if you chose to model it as a an agent that sends messages to the brain, it is an agent of a different nature. The sensory agent doesn't receive messages from other agents, it receives raw input from the outside world (photons, electrical fields, chemical reactions etc). Where do you draw the distinction between "messages" and "information" (raw or cooked). Information theory was contrived to model messages sent in a noisy channel, but it applies just as well to data streams that have no communicative intent or origin. Its a distinction without a difference. You may as well treat all information as messages in a channel, even if the sender is nature herself. Alternatively, you may as well treat all messages as just information, and view "senders" with "intent" as just another physical process in a world of physical computation. As the cliche saying goes, "information is physical". >We could perhaps come to agree upon some imagined world, but that could change arbitrarily much from one day to the next on a whim. Ok, this is a fine basis to work with. How about this. "Reality is the set of beliefs which, if you disagree with them too much and for too long, you are eventually and permanently removed from the conversation." For example, quite recently, large swaths of people held an exquisite referendum on the existence of covid. Needless to say, rather than covid disappearing on their whim, a great deal of them are now permanently no longer participants in this conversation. Notice this isn't far off from my original postulate. "Whatever we agree upon is our reality, and we'll argue with/fight anyone who disagrees until they agree with us or us with them." Allowing for some personification I could phrase this scenario as "they disagreed with the virus and the virus won." So perhaps I do need to add one thing to the postulates. One thing which remains objectively true even in the conversational model of reality. "You can die. Dying means never being heard from in this conversation again." You're right there's no mechanism for any choice of words to win over any other choice of words without an objective consequence to losing. The things that exist so far are "you", "me", "this" and "death". Not where I was expecting this to go but good point.

tsimionescu 4y ago

> You cite the primacy of sensation, but what more is sensation than a message from the sensory organ to the brain. This is false, it's a version of the homunculus fallacy. A sensory organ is something that connects the brain to the physical world. Even if you chose to model it as a an agent that sends messages to the brain, it is an agent of a different nature. The sensory agent doesn't receive messages from other agents, it receives raw input from the outside world (photons, electrical fields, chemical reactions etc). In contrast, if you were a brain in a vat with no ability to directly perceive the world or interact with it in any way, it would be impossible for you to know that "I can pass through walls" is fundamentally impossible. Even for your GPT-3 thought experiment - ultimately it is the effects of the physical world perceived directly by humans sense organs that shape what GPT-3 would utter. That is, even if it can learn what the world is like simply by talking to us, it's still learning about the real world from someone's direct experience with it. If we were all GPT-3s, with no cameras and pressure sensors and motors etc, we would be unable to reason in any sense about the world itself. We could perhaps come to agree upon some imagined world, but that could change arbitrarily much from one day to the next on a whim.

IIAOPSW 4y ago
>A sensory organ is something that connects the brain to the physical world. Even if you chose to model it as a an agent that sends messages to the brain, it is an agent of a different nature. The sensory agent doesn't receive messages from other agents, it receives raw input from the outside world (photons, electrical fields, chemical reactions etc). Where do you draw the distinction between "messages" and "information" (raw or cooked). Information theory was contrived to model messages sent in a noisy channel, but it applies just as well to data streams that have no communicative intent or origin. Its a distinction without a difference. You may as well treat all information as messages in a channel, even if the sender is nature herself. Alternatively, you may as well treat all messages as just information, and view "senders" with "intent" as just another physical process in a world of physical computation. As the cliche saying goes, "information is physical". >We could perhaps come to agree upon some imagined world, but that could change arbitrarily much from one day to the next on a whim. Ok, this is a fine basis to work with. How about this. "Reality is the set of beliefs which, if you disagree with them too much and for too long, you are eventually and permanently removed from the conversation." For example, quite recently, large swaths of people held an exquisite referendum on the existence of covid. Needless to say, rather than covid disappearing on their whim, a great deal of them are now permanently no longer participants in this conversation. Notice this isn't far off from my original postulate. "Whatever we agree upon is our reality, and we'll argue with/fight anyone who disagrees until they agree with us or us with them." Allowing for some personification I could phrase this scenario as "they disagreed with the virus and the virus won." So perhaps I do need to add one thing to the postulates. One thing which remains objectively true even in the conversational model of reality. "You can die. Dying means never being heard from in this conversation again." You're right there's no mechanism for any choice of words to win over any other choice of words without an objective consequence to losing. The things that exist so far are "you", "me", "this" and "death". Not where I was expecting this to go but good point.