4 ms·

> Rather, the LLM is generating (authoring) text about a fictional entity, Sydney the Artificial Intelligence. Maybe we are doing the same. We have a mental mo

> Rather, the LLM is generating (authoring) text about a fictional entity, Sydney the Artificial Intelligence. Maybe we are doing the same. We have a mental model of our Self and generate language from its perspective.

Jensson 4y ago
What humans say tend to be related to what the human body the mind is attached to has done or experienced. That sort of relation doesn't exist for todays AI, what they say aren't related to anything at all, its just fiction.
- visarga 4y ago
  But there is something they can relate to - it is our replies and questions. We know how easy it is to gaslight an AI. For AI we are the external world, they get to perceive and act in pure text format.
  - Jensson 4y ago
    But that AI just lives for a single conversation. Then you refresh and now it is dead, instead you get to interact with a new AI and see it birth and then die a few seconds/minute slater. There is so little there that it is hard to say much at all about it.
    - pixl97 4y ago
      Philosophically you keep arguing more terrible points... if this is a lifeform (which I'm not saying it is) we're playing genocide with it by murdering it a few billion times a day.
      - klipt 4y ago
        Some people can't form long term memories due to brain injury. Are they killing themselves every time they forget their short term memories?
aflukasz 4y ago
Since the whole GPT3 thing blown up, I'm thinking from time to time... how I am generating what I say. I'm sure many smart people wrote papers on that. I did not read any of them, mind you, will just share a short thought of my own here, hopefully providing some intellectual entertainment for someone. It seems from my point of view that, broadly speaking, I maintain four things at the moment of talking to someone: 1. A graph of concepts that were used / are potentially going to be used by me or by my interlocutor. 2. Some emotional state. 3. Some fuzzy picture of where I'm going with what I'm saying in the short term of say 20 seconds. 4. Extra short term focused process of making sure that the next 2-3 words fit to the one I just said and are going to fulfill requirements stemming from (3) and (1); this happens with some influence form (2), ideally not too much, if I consider current state of (2) not helping to be constructive. GPT3 obviously lacks (2). My limited understanding of LLMs is that it does (4), maybe (3) and probably not (1) (?). So I'm just wondering - are those LLMs really that far from a "human being"? Again, not an expert. Happy to be corrected.

Jensson 4y ago

What humans say tend to be related to what the human body the mind is attached to has done or experienced. That sort of relation doesn't exist for todays AI, what they say aren't related to anything at all, its just fiction.

visarga 4y ago
But there is something they can relate to - it is our replies and questions. We know how easy it is to gaslight an AI. For AI we are the external world, they get to perceive and act in pure text format.
- Jensson 4y ago
  But that AI just lives for a single conversation. Then you refresh and now it is dead, instead you get to interact with a new AI and see it birth and then die a few seconds/minute slater. There is so little there that it is hard to say much at all about it.
  - pixl97 4y ago
    Philosophically you keep arguing more terrible points... if this is a lifeform (which I'm not saying it is) we're playing genocide with it by murdering it a few billion times a day.
    - klipt 4y ago
      Some people can't form long term memories due to brain injury. Are they killing themselves every time they forget their short term memories?

aflukasz 4y ago

Since the whole GPT3 thing blown up, I'm thinking from time to time... how I am generating what I say. I'm sure many smart people wrote papers on that. I did not read any of them, mind you, will just share a short thought of my own here, hopefully providing some intellectual entertainment for someone. It seems from my point of view that, broadly speaking, I maintain four things at the moment of talking to someone: 1. A graph of concepts that were used / are potentially going to be used by me or by my interlocutor. 2. Some emotional state. 3. Some fuzzy picture of where I'm going with what I'm saying in the short term of say 20 seconds. 4. Extra short term focused process of making sure that the next 2-3 words fit to the one I just said and are going to fulfill requirements stemming from (3) and (1); this happens with some influence form (2), ideally not too much, if I consider current state of (2) not helping to be constructive. GPT3 obviously lacks (2). My limited understanding of LLMs is that it does (4), maybe (3) and probably not (1) (?). So I'm just wondering - are those LLMs really that far from a "human being"? Again, not an expert. Happy to be corrected.