4 ms·

The way GPT-4 works is by having built a world model of the generative processes that produced the data it was trained on. The more data you train it on (and th

The way GPT-4 works is by having built a world model of the generative processes that produced the data it was trained on. The more data you train it on (and the larger and therefore more capable he model is), the better it performs - i.e the more complete and consistent this world model has evidentially become. I'm not sure where you are seeing daylight between this and your own definition of intelligence. FWIW GPT-4, being a neural net, is more analog than not. It's driven by floating point values not 1's and 0's. The values are imperfectly calculated (limited accuracy) as computer math always is. There is also a large element of pure randomness to the output of any of these LLMs. They don't get to control exactly what words they generate ... the model generates probabilities over 10's of thousands of possible output words, and a random number generator is used to select one of the higher rated words to output. This semi-random word is then fed back into the model, for it to "generate" the next word ... it is continuously having to adapt to this randomness forced upon it.

mike741 4y ago
> The more data you train it on (and the larger and therefore more capable he model is), the better it performs - i.e the more complete and consistent this world model has evidentially become. Increasing training data doesn't increase consistency. Each data point acts as a potential new axiom, and each axiom decreases consistency. GPT-4 is trained to satisfy humans, and humans are wildly inconsistent. Even if humans were perfectly consistent, attempting to satisfy multiple different humans simultaneously results in inconsistency. Additionally, even if GPT-4 were perfectly complete and consistent it still wouldn't have reached this state autonomously. So the difference between GPT-4 and intelligence, by my definition, is night and day. > FWIW GPT-4, being a neural net, is more analog than not. It's driven by floating point values not 1's and 0's. Floating points are digital 1's and 0's. Adding more digits is never going to make something analog. > The values are imperfectly calculated (limited accuracy) as computer math always is. Agreed. >There is also a large element of pure randomness to the output of any of these LLMs. Strongly disagree. There isn't a single element of randomness during the training stage. We know the exact architecture of the neural net, we know the exact data it was trained on, and we know the exact beam selection algorithms used to synthesize outputs. Every single step can be simulated, traced, and recreated to achieve the exact same results. The number of steps involved might overwhelm us, but that doesn't make it random. > They don't get to control exactly what words they generate We do get to control it, we just lose track of the inputs and then pretend it was all out of our control. But of course every single step was willed and controlled by us. We call it "random" for personal convenience, not because its actually true.
- HarHarVeryFunny 4y ago
  These models don't output sequences - which is where you'd use beam search - they output a single word at a time. The output is a set of probabilities (from a SoftMax) which is then sampled at a given sampling "temperature" (degree of randomness). There's no point discussing it when you obviously don't have clue how these models work, won't listen when you're told, and just prefer to make stuff up.
  - mike741 4y ago
    Beamsearch is just one example, the same applies to top sampling and greedy search. Focusing on one approach suggests you've missed the actual point: if you know how a given output is synthesized, then its not autonomous. You're trying to nitpick as an excuse to avoid a substantive response. If you want me to agree with you, you'll need to offer a counterpoint. Saying things like "It's driven by floating point values not 1's and 0's." and conflating pseudo randomness with actual randomness does not inspire confidence.
    - HarHarVeryFunny 4y ago
      I was curious what your definition of intelligence was such that you thought GPT-4 doesn't exhibit it. It's your opinion - doesn't have to agree with mine. You seem to place an importance on whether the models are entirely predictable or not, which is why I pointed out that the output is randomly sampled.
      - mike741 4y ago
        It's not random, its pseudo-random. Those are not the same thing.
        HarHarVeryFunny 4y ago
        You could choose to use a truly random hardware random number generator, and it would make zero difference to the system, empirically or philosophically.

mike741 4y ago

> The more data you train it on (and the larger and therefore more capable he model is), the better it performs - i.e the more complete and consistent this world model has evidentially become. Increasing training data doesn't increase consistency. Each data point acts as a potential new axiom, and each axiom decreases consistency. GPT-4 is trained to satisfy humans, and humans are wildly inconsistent. Even if humans were perfectly consistent, attempting to satisfy multiple different humans simultaneously results in inconsistency. Additionally, even if GPT-4 were perfectly complete and consistent it still wouldn't have reached this state autonomously. So the difference between GPT-4 and intelligence, by my definition, is night and day. > FWIW GPT-4, being a neural net, is more analog than not. It's driven by floating point values not 1's and 0's. Floating points are digital 1's and 0's. Adding more digits is never going to make something analog. > The values are imperfectly calculated (limited accuracy) as computer math always is. Agreed. >There is also a large element of pure randomness to the output of any of these LLMs. Strongly disagree. There isn't a single element of randomness during the training stage. We know the exact architecture of the neural net, we know the exact data it was trained on, and we know the exact beam selection algorithms used to synthesize outputs. Every single step can be simulated, traced, and recreated to achieve the exact same results. The number of steps involved might overwhelm us, but that doesn't make it random. > They don't get to control exactly what words they generate We do get to control it, we just lose track of the inputs and then pretend it was all out of our control. But of course every single step was willed and controlled by us. We call it "random" for personal convenience, not because its actually true.

HarHarVeryFunny 4y ago
These models don't output sequences - which is where you'd use beam search - they output a single word at a time. The output is a set of probabilities (from a SoftMax) which is then sampled at a given sampling "temperature" (degree of randomness). There's no point discussing it when you obviously don't have clue how these models work, won't listen when you're told, and just prefer to make stuff up.
- mike741 4y ago
  Beamsearch is just one example, the same applies to top sampling and greedy search. Focusing on one approach suggests you've missed the actual point: if you know how a given output is synthesized, then its not autonomous. You're trying to nitpick as an excuse to avoid a substantive response. If you want me to agree with you, you'll need to offer a counterpoint. Saying things like "It's driven by floating point values not 1's and 0's." and conflating pseudo randomness with actual randomness does not inspire confidence.
  - HarHarVeryFunny 4y ago
    I was curious what your definition of intelligence was such that you thought GPT-4 doesn't exhibit it. It's your opinion - doesn't have to agree with mine. You seem to place an importance on whether the models are entirely predictable or not, which is why I pointed out that the output is randomly sampled.
    - mike741 4y ago
      It's not random, its pseudo-random. Those are not the same thing.
      - HarHarVeryFunny 4y ago
        You could choose to use a truly random hardware random number generator, and it would make zero difference to the system, empirically or philosophically.