3 ms·

Humans who have heard of Monty Hall might also say you should always switch without noticing that the situation is different. That's not evidence that they can'

Humans who have heard of Monty Hall might also say you should always switch without noticing that the situation is different. That's not evidence that they can't think, just that they're fallible. People on here always assert LLMs don't "really" think or don't "really" know without defining what all that even means, and to me it's getting pretty old. It feels like an escape hatch so we don't feel like our human special sauce is threatened, a bit like how people felt threatened by heliocentrism or evolution.

no_wizard 1y ago
On the other hand, computers are suppose to be both accurate and able to reproduce said accuracy. The failure of an LLM to reason this out is indicative that really, it isn’t reasoning at all. It’s a subtle but welcome reminder that it’s pattern matching
- chpatrick 1y ago
  Computers might be accurate but statistical models never were 100% accurate. That doesn't imply that no reasoning is happening. Humans get stuff wrong too but they certainly think and reason. "Pattern matching" to me is another one of those vague terms like "thinking" and "knowing" that people decide LLMs do or don't do based on vibes.
  - no_wizard 1y ago
    Pattern matching has a definition in this field, it does mean specific things. We know machine learning has excelled at this in greater and greater capacities over the last decade The other part of this is weighted filtering given a set of rules, which is a simple analogy to how AlphaGo did its thing. Dismissing all this as vague is effectively doing the same thing as you are saying others do. This technology has limits and despite what Altman says, we do know this, and we are exploring them, but it’s within its own confines. They’re fundamentally wholly understandable systems that work on a consistent level in terms of the how they do what they do (that is separate from the actual produced output) I think reasoning, as any layman would use the term, is not accurate to what these systems do.
    - chpatrick 1y ago
      > Pattern matching has a definition in this field, it does mean specific things. Such as? > They’re fundamentally wholly understandable systems that work on a consistent level in terms of the how they do what they do (that is separate from the actual produced output) Multi billion parameter models are definitely not wholly understandable and I don't think any AI researcher would claim otherwise. We can train them but we don't know how they work any more than we understand how the training data was made. > I think reasoning, as any layman would use the term, is not accurate to what these systems do. Based on what?
      - no_wizard 1y ago
        You’re welcoming to provide counters. I think these are all sufficiently common things that they stand on their own as to what I posit
        chpatrick 1y ago
        Look, you're claiming something, it's up to you to back it up. Handwaving what any of these things mean isn't an argument.
    - Timwi 1y ago
      You're derailing the conversation. The discussion was about thinking, and now you're arguing about something entirely different and didn't even mention the word “think” a single time. If you genuinely believe that anyone knows how LLMs work, how brains work, and/or how or why the latter does “thinking” while the former does not, you're just simply wrong. AI researchers fully acknowledge ignorance in this matter.
- papichulo2023 1y ago
  I guess computer vision didnt get this memo and it is useless.
bscphil 1y ago
> Humans who have heard of Monty Hall might also say you should always switch without noticing that the situation is different. That's not evidence that they can't think, just that they're fallible. At some point we start playing a semantics game over the meaning of "thinking", right? Because if a human makes this mistake because they jumped to an already-known answer without noticing a changed detail, it's because (in the usage of the person you're replying to), the human is pattern matching, instead of thinking. I don't think is surprising. In fact I think much of what passes for thinking in casual conversation is really just applying heuristics we've trained in our own brains to give us the correct answer without having to think rigorously. We remember mental shortcuts. On the other hand, I don't think it's controversial that (some) people are capable of performing the rigorous analysis of the problem needed to give a correct answer in cases like this fake Monty Hall problem. And that's key... if you provide slightly more information and call out the changed nature of the problem to the LLM, it may give you the correct response, but it can't do the sort of reasoning that would reliably give you the correct answer the way a human can. I think that's why the GP doesn't want to call it "thinking" - they want to reserve that for a particular type of reflective process that can rigorously perform logical reasoning in a consistently valid way.
- chpatrick 1y ago
  I'm not sure what your argument is. The common claim that annoys me about LLMs on here is that they're not "really" coming up with ideas but that they're cheating and just repeating something they read on the internet somewhere that was written by a human who can "really" think. To me this is obviously false if you've talked to a SOTA LLM or know a little about how they work.
johnnyanmac 1y ago
>People on here always assert LLMs don't "really" think or don't "really" know without defining what all that even means, Sure. To Think: able to process information in a given context and arrive at an answer or analysis. an LLM only simulates this with pattern matching. It didn't really consider the problem, it did the equivalent of googling a lot of terms and then spat something that sounded like an answer To Know: To reproduce information based on past thinking, as well as to properly verify and reason about with the information. I know 1+1 = 2 because (I'm not a math major, feel free to inject number theory instead) I was taught that arithmatic is a form of counting, and I was taught the mechanics of counting to prove how to add. Most LLM models don't really "know" this to begin with for the reasons above. Maybe we'll see if this study mode is different. Somehow I am skeptical if this will really change minds, though. People making swipes at the community like this often are not really engaging in a conversation with ideas they oppose.
- chpatrick 1y ago
  You're just deferring to another vague term "pattern matching". If I think back to something I was taught in primary school and conclude that 1+1=2 is that pattern matching? Therefore I don't really "know" or "think"? People pretend like LLMs are like some 80s markov chain model or nearest neighbor search, which is just uninformed.
  - johnnyanmac 1y ago
    Do you want to shift the discussion to the definition of a "pattern" or are we going to continue to move the goalpost? I'm trying to respond to your inquiry and instead we're just stuck in minutia. Yes, to make an apple pie from scratch, we need to first invent the universe. Is that productive conversation to fall into or can we just admit that your dismissing any opinion that goes against your purview? >If I think back to something I was taught in primary school and conclude that 1+1=2 is that pattern matching? Yes. That is an example of pattern matching. Let me know when you want to go back to talking about LLMs.
    - chpatrick 1y ago

no_wizard 1y ago

On the other hand, computers are suppose to be both accurate and able to reproduce said accuracy. The failure of an LLM to reason this out is indicative that really, it isn’t reasoning at all. It’s a subtle but welcome reminder that it’s pattern matching

chpatrick 1y ago
Computers might be accurate but statistical models never were 100% accurate. That doesn't imply that no reasoning is happening. Humans get stuff wrong too but they certainly think and reason. "Pattern matching" to me is another one of those vague terms like "thinking" and "knowing" that people decide LLMs do or don't do based on vibes.
- no_wizard 1y ago
  Pattern matching has a definition in this field, it does mean specific things. We know machine learning has excelled at this in greater and greater capacities over the last decade The other part of this is weighted filtering given a set of rules, which is a simple analogy to how AlphaGo did its thing. Dismissing all this as vague is effectively doing the same thing as you are saying others do. This technology has limits and despite what Altman says, we do know this, and we are exploring them, but it’s within its own confines. They’re fundamentally wholly understandable systems that work on a consistent level in terms of the how they do what they do (that is separate from the actual produced output) I think reasoning, as any layman would use the term, is not accurate to what these systems do.
  - chpatrick 1y ago
    > Pattern matching has a definition in this field, it does mean specific things. Such as? > They’re fundamentally wholly understandable systems that work on a consistent level in terms of the how they do what they do (that is separate from the actual produced output) Multi billion parameter models are definitely not wholly understandable and I don't think any AI researcher would claim otherwise. We can train them but we don't know how they work any more than we understand how the training data was made. > I think reasoning, as any layman would use the term, is not accurate to what these systems do. Based on what?
    - no_wizard 1y ago
      You’re welcoming to provide counters. I think these are all sufficiently common things that they stand on their own as to what I posit
      - chpatrick 1y ago
        Look, you're claiming something, it's up to you to back it up. Handwaving what any of these things mean isn't an argument.
  - Timwi 1y ago
    You're derailing the conversation. The discussion was about thinking, and now you're arguing about something entirely different and didn't even mention the word “think” a single time. If you genuinely believe that anyone knows how LLMs work, how brains work, and/or how or why the latter does “thinking” while the former does not, you're just simply wrong. AI researchers fully acknowledge ignorance in this matter.
papichulo2023 1y ago
I guess computer vision didnt get this memo and it is useless.

bscphil 1y ago

> Humans who have heard of Monty Hall might also say you should always switch without noticing that the situation is different. That's not evidence that they can't think, just that they're fallible. At some point we start playing a semantics game over the meaning of "thinking", right? Because if a human makes this mistake because they jumped to an already-known answer without noticing a changed detail, it's because (in the usage of the person you're replying to), the human is pattern matching, instead of thinking. I don't think is surprising. In fact I think much of what passes for thinking in casual conversation is really just applying heuristics we've trained in our own brains to give us the correct answer without having to think rigorously. We remember mental shortcuts. On the other hand, I don't think it's controversial that (some) people are capable of performing the rigorous analysis of the problem needed to give a correct answer in cases like this fake Monty Hall problem. And that's key... if you provide slightly more information and call out the changed nature of the problem to the LLM, it may give you the correct response, but it can't do the sort of reasoning that would reliably give you the correct answer the way a human can. I think that's why the GP doesn't want to call it "thinking" - they want to reserve that for a particular type of reflective process that can rigorously perform logical reasoning in a consistently valid way.

chpatrick 1y ago
I'm not sure what your argument is. The common claim that annoys me about LLMs on here is that they're not "really" coming up with ideas but that they're cheating and just repeating something they read on the internet somewhere that was written by a human who can "really" think. To me this is obviously false if you've talked to a SOTA LLM or know a little about how they work.

johnnyanmac 1y ago

>People on here always assert LLMs don't "really" think or don't "really" know without defining what all that even means, Sure. To Think: able to process information in a given context and arrive at an answer or analysis. an LLM only simulates this with pattern matching. It didn't really consider the problem, it did the equivalent of googling a lot of terms and then spat something that sounded like an answer To Know: To reproduce information based on past thinking, as well as to properly verify and reason about with the information. I know 1+1 = 2 because (I'm not a math major, feel free to inject number theory instead) I was taught that arithmatic is a form of counting, and I was taught the mechanics of counting to prove how to add. Most LLM models don't really "know" this to begin with for the reasons above. Maybe we'll see if this study mode is different. Somehow I am skeptical if this will really change minds, though. People making swipes at the community like this often are not really engaging in a conversation with ideas they oppose.

chpatrick 1y ago
You're just deferring to another vague term "pattern matching". If I think back to something I was taught in primary school and conclude that 1+1=2 is that pattern matching? Therefore I don't really "know" or "think"? People pretend like LLMs are like some 80s markov chain model or nearest neighbor search, which is just uninformed.
- johnnyanmac 1y ago
  Do you want to shift the discussion to the definition of a "pattern" or are we going to continue to move the goalpost? I'm trying to respond to your inquiry and instead we're just stuck in minutia. Yes, to make an apple pie from scratch, we need to first invent the universe. Is that productive conversation to fall into or can we just admit that your dismissing any opinion that goes against your purview? >If I think back to something I was taught in primary school and conclude that 1+1=2 is that pattern matching? Yes. That is an example of pattern matching. Let me know when you want to go back to talking about LLMs.
  - chpatrick 1y ago