5 ms·

Everyone is captivated by the current hot thing, LLM/GPT's. But LLM's are not the whole of AI research. A lot of your arguments are based on 'embodied' reason

Everyone is captivated by the current hot thing, LLM/GPT's. But LLM's are not the whole of AI research. A lot of your arguments are based on 'embodied' reasoning. Humans live in the world, they need to eat and survive. LLM's just compress what humans generated in the world. Correct, current LLM's are mostly regurgitating, but they don't "speak because we are in a shared world of social intentions". I'd say game worlds are the frontier, because they are able to simulate a lower resolution world for current AI's to learn in. And in that world, they do embody it and have purpose (rewards/goals), they need to survive. DeepMind's AlphaGo was when I switched. https://www.wired.com/2016/03/two-moves-alphago-lee-sedol-redefined-future/ https://www.wired.com/2016/03/two-moves-alphago-lee-sedol-re... Move 37, it was called alien, creative, inhuman intelligence. Now, scale that up to our world, with admittedly, thousands/millions of more variables. Put it in a robot body, with vision (the latest studies show AI vision building a world model context). Add some goals. Bam, some dangerous stuff, AI embodied in the world, with a goal to survive. It might be far away, but where we are now was supposed to take another hundred years. So who knows. The military is already running world simulations where the AI's goal function lead it to kill the soldier operating the AI in order to bypass him. It 'learned' to bypass the operator by killing them to achieve its goal. Yes, Hyperbole. But really, not by much. But back to our discussion. At that point, does the robot have an internal subjective perspective? Did AlphaGO when it was reasoning about its small low variable world?

mjburgess 3y ago
It never has an "internal subjective experience" -- AlphaGO wasnt reasoning. Reasoning is a cognitive process in which propositions, which model the world, are considered in turn and subject to ecological rationality (concerns of utilty, effort, interest, preference, etc.). At no point in the flight of an aeroplane does it ever lay eggs. You are using smoke to establish fire, these ways of measuring internal mental states of animals only work on animals. If you can produce a robot with no prior conceptual scheme of, say, a novel apartment it is thrown into; a robot which can then determine what is in that apartment, how roughly it works (eg., light switch -> lights turn on), of an account of that apartment; explain why it has explored it; show that its behaviour is moderated and caused by these stated goals; ask it for opinions about the apartment etc. -- then we are actually playing the intelligence game, at least. Rather than stupid magic laterns. Now, does this robot have a subjective experience? Well I think we need to keep going with our tests: does it have an aversion to toxic stimulous? Is this aversion moderating its goals and behaviour? Are its memories contextualised by these aversions (eg., does its process of remembering display a variety when remembering negative vs. positive experiences)? And so on. If I can ask, "Did you find my apartment fun?" and it can answer because it did, or did not -- then we're very close. That is if we can show the reason it says, "yes" or "no" had to do with a history of taste, judgement, preference, curiosity, etc. all built up by itself -- not under "supervision with the right answers" but with no problem-specific answers ever given... and so on. Questions of these kind arent even revenant to anything in AI. Any sincere AI engineer will say that they have nothing to do with the goals of the system theyre building. All AI that we can actually access, 100% has no interest, methods or ambitions to deliver any of the above. AI isnt even in the category of intelligence; it isnt even trying to produce it.
- calf 3y ago
  So I'm trying to understand your argument here, but why isn't "Reasoning is a cognitive process" circular logic? AlphaGo wasn't reasoning, how? I think reasonably AlphaGo has modeled a world, and it is by design subject to the bounded rationality of a game-theoretic optimization problem. So two of your criteria are satisfied. So - I'm just reading your definition here - AlphaGo wasn't reasoning because it is not a cognitive process. Is that your specific argument? And if so, then what is a cognitive process? I'm just focusing on this part here and I don't see the argument clearly: > It never has an "internal subjective experience" -- AlphaGO wasnt reasoning. > Reasoning is a cognitive process in which propositions, which model the world, are considered in turn and subject to ecological rationality (concerns of utilty, effort, interest, preference, etc.).
  - mjburgess 3y ago
    > Reasoning is a cognitive process You could have a non-cognitive view of reasoning, or an embodied one. NNs are a-cognitive systems, they do not engage in reasoning of any form. Reasoning concerns inference across truth-apt propositions (eg., A->B, A thef. B). NNs have no propositions, nor are any parts truth-apt, true or false. NNs are statistical systems which select answers by weights found from optimisation. No process either in optimisation or prediction is an inferential one in the sense of cognition. I also deny that the "world" as used by frankly just philosophically incompetent ML researchers, whose gross lack of familiarity with basically anything outside pytorch, is even relevant to the sense of "world" that propositions bare a truth relation to. A world in the relevant sense isn't the state space of the training data -- this is an insane supposition which makes the claim "AI has world models" actually circular. The relevant sense of world is the cause of the training data. If the training data is about an abstract game then you collapse the distinction since the rules are the data. The famous "NNs learn WMs" paper is just this: choose a system whose data is just a restatement of the system; rather than a measure of a world. NNs do not form representational models of the cause of their measurement data because all they do is induce (ie., compress by function-fitting) across the measurement space. They model the measurements not their causes. This is only "predictive" of features of the causal origin of measurement data in rigged scenarios, and in general, fails catastrophically to be predictive. Consider running a NN across photos of the sky: it is impossible for this process to produce newton's law of gravity. The weights are just models of the pixels, and these are not distributed according to this law. Worse, in general, there is no function from the measurement space to properties of its causal origin -- so it is impossible to build representations by induction. (eg., there is no function Photo->Cat|Dog, the distributions of pixels in photos is ambiguous, and changes over time). Reasoning, as in cognition, is an (logically) inferential process which considers propositions that bare a truth relation to the world which is the causal origin of concepts which the proposition comprises (created by a biogenerative process). It is the activity of an agent with an interior subjectivity and ecological rationality. Reasoning is done by an agent about something of interest to that agent, with motivation towards a goal the agent has, in the service of the agent's preferences, etc. If reasoning is an abstract pattern, then rice falling to the ground is likewise "reasoning".

mjburgess 3y ago

It never has an "internal subjective experience" -- AlphaGO wasnt reasoning. Reasoning is a cognitive process in which propositions, which model the world, are considered in turn and subject to ecological rationality (concerns of utilty, effort, interest, preference, etc.). At no point in the flight of an aeroplane does it ever lay eggs. You are using smoke to establish fire, these ways of measuring internal mental states of animals only work on animals. If you can produce a robot with no prior conceptual scheme of, say, a novel apartment it is thrown into; a robot which can then determine what is in that apartment, how roughly it works (eg., light switch -> lights turn on), of an account of that apartment; explain why it has explored it; show that its behaviour is moderated and caused by these stated goals; ask it for opinions about the apartment etc. -- then we are actually playing the intelligence game, at least. Rather than stupid magic laterns. Now, does this robot have a subjective experience? Well I think we need to keep going with our tests: does it have an aversion to toxic stimulous? Is this aversion moderating its goals and behaviour? Are its memories contextualised by these aversions (eg., does its process of remembering display a variety when remembering negative vs. positive experiences)? And so on. If I can ask, "Did you find my apartment fun?" and it can answer because it did, or did not -- then we're very close. That is if we can show the reason it says, "yes" or "no" had to do with a history of taste, judgement, preference, curiosity, etc. all built up by itself -- not under "supervision with the right answers" but with no problem-specific answers ever given... and so on. Questions of these kind arent even revenant to anything in AI. Any sincere AI engineer will say that they have nothing to do with the goals of the system theyre building. All AI that we can actually access, 100% has no interest, methods or ambitions to deliver any of the above. AI isnt even in the category of intelligence; it isnt even trying to produce it.

calf 3y ago
So I'm trying to understand your argument here, but why isn't "Reasoning is a cognitive process" circular logic? AlphaGo wasn't reasoning, how? I think reasonably AlphaGo has modeled a world, and it is by design subject to the bounded rationality of a game-theoretic optimization problem. So two of your criteria are satisfied. So - I'm just reading your definition here - AlphaGo wasn't reasoning because it is not a cognitive process. Is that your specific argument? And if so, then what is a cognitive process? I'm just focusing on this part here and I don't see the argument clearly: > It never has an "internal subjective experience" -- AlphaGO wasnt reasoning. > Reasoning is a cognitive process in which propositions, which model the world, are considered in turn and subject to ecological rationality (concerns of utilty, effort, interest, preference, etc.).
- mjburgess 3y ago
  > Reasoning is a cognitive process You could have a non-cognitive view of reasoning, or an embodied one. NNs are a-cognitive systems, they do not engage in reasoning of any form. Reasoning concerns inference across truth-apt propositions (eg., A->B, A thef. B). NNs have no propositions, nor are any parts truth-apt, true or false. NNs are statistical systems which select answers by weights found from optimisation. No process either in optimisation or prediction is an inferential one in the sense of cognition. I also deny that the "world" as used by frankly just philosophically incompetent ML researchers, whose gross lack of familiarity with basically anything outside pytorch, is even relevant to the sense of "world" that propositions bare a truth relation to. A world in the relevant sense isn't the state space of the training data -- this is an insane supposition which makes the claim "AI has world models" actually circular. The relevant sense of world is the cause of the training data. If the training data is about an abstract game then you collapse the distinction since the rules are the data. The famous "NNs learn WMs" paper is just this: choose a system whose data is just a restatement of the system; rather than a measure of a world. NNs do not form representational models of the cause of their measurement data because all they do is induce (ie., compress by function-fitting) across the measurement space. They model the measurements not their causes. This is only "predictive" of features of the causal origin of measurement data in rigged scenarios, and in general, fails catastrophically to be predictive. Consider running a NN across photos of the sky: it is impossible for this process to produce newton's law of gravity. The weights are just models of the pixels, and these are not distributed according to this law. Worse, in general, there is no function from the measurement space to properties of its causal origin -- so it is impossible to build representations by induction. (eg., there is no function Photo->Cat|Dog, the distributions of pixels in photos is ambiguous, and changes over time). Reasoning, as in cognition, is an (logically) inferential process which considers propositions that bare a truth relation to the world which is the causal origin of concepts which the proposition comprises (created by a biogenerative process). It is the activity of an agent with an interior subjectivity and ecological rationality. Reasoning is done by an agent about something of interest to that agent, with motivation towards a goal the agent has, in the service of the agent's preferences, etc. If reasoning is an abstract pattern, then rice falling to the ground is likewise "reasoning".