4 ms·

Human teachers aren't omniscient, but often have absolute authority over their pupils. This flaw isn't unique to AI.

godelski 3y ago
I have more confidence in a teacher, especially in niche subjects. One of the big reasons why is that I can ask a teacher a question, they will infer what I am trying to ask, and not easily flip flop. Because a teacher has confidence and GPT doesn't. A teacher will even find their own mistakes without prompting (though of course, everyone needs it sometimes. Just a teacher won't say that a kilogram of feathers and a pound of bricks weigh the same and then explain how 1 kg is 2.54 lbs). Sure, no system is perfect, but not all imperfections are equal, and I'm tired of pretending like they are.
- ben_w 3y ago
  > Sure, no system is perfect, but not all imperfections are equal, and I'm tired of pretending like they are. I think this is the actual issue, rather than the other specifics you raise — my anecdote of human mistakes includes having a 44kHz/16-bit audio recording of a whole university lecture where the lecturer insisted that nobody can record mocap joint positions for a mere dozen joints at a mere 24Hz/8-bit for more than a few minutes because that is too much data to store, and it was their confidence that kept them from updating. But, back to the point. Every instance of any specific AI release (GPT or LLM or whatever else) are identical; errors are therefore necessarily correlated. Humans get an advantage from our imperfections being un-equal, we benefit from the wisdom of the crowd (with conditions) and second opinions are meaningfully independent (again with conditions), but not so for any specific AI. (Possible twist: Mixture of Experts).
  - godelski 3y ago
    That's not true even before mixture of experts because the systems are stochastic. They are only identical if we were to always initialize with the same starting seed and rng_state.
    - ben_w 3y ago
      While they do indeed have a random number generator, I think the main strengths and weaknesses of these models (and humans) are in the structures rather than at the token temperature stage — I expect a human to answer any given question slightly differently at different times, but asking the same doctor for a second opinion as the first is considered the setup for a joke rather than a sensible idea.
      - godelski 3y ago
        This is the assumption of a flat and smooth latent manifold. If either of those conditions are not met, then seeds can have a wild effect on the model output. There's only bounds on the latent structure being locally smooth or locally flat btw. This is the underpinning of such works as the Lottery Hypothesis but there's a whole rabbit hole to dig down here and it is why Alpha Go did so poorly in round 2. This is an important part of the discussion of generalizability btw. I do not think your analogy works out and it is inconsistent with your prior comment, because you're right that anyone asking GPT is akin to anyone asking the same doctor. The second doctor is a second model (or potentially a second GPT checkpoint depending on how we want to consider divergence).

I have more confidence in a teacher, especially in niche subjects. One of the big reasons why is that I can ask a teacher a question, they will infer what I am trying to ask, and not easily flip flop. Because a teacher has confidence and GPT doesn't. A teacher will even find their own mistakes without prompting (though of course, everyone needs it sometimes. Just a teacher won't say that a kilogram of feathers and a pound of bricks weigh the same and then explain how 1 kg is 2.54 lbs). Sure, no system is perfect, but not all imperfections are equal, and I'm tired of pretending like they are.

ben_w 3y ago
> Sure, no system is perfect, but not all imperfections are equal, and I'm tired of pretending like they are. I think this is the actual issue, rather than the other specifics you raise — my anecdote of human mistakes includes having a 44kHz/16-bit audio recording of a whole university lecture where the lecturer insisted that nobody can record mocap joint positions for a mere dozen joints at a mere 24Hz/8-bit for more than a few minutes because that is too much data to store, and it was their confidence that kept them from updating. But, back to the point. Every instance of any specific AI release (GPT or LLM or whatever else) are identical; errors are therefore necessarily correlated. Humans get an advantage from our imperfections being un-equal, we benefit from the wisdom of the crowd (with conditions) and second opinions are meaningfully independent (again with conditions), but not so for any specific AI. (Possible twist: Mixture of Experts).
- godelski 3y ago
  That's not true even before mixture of experts because the systems are stochastic. They are only identical if we were to always initialize with the same starting seed and rng_state.
  - ben_w 3y ago
    While they do indeed have a random number generator, I think the main strengths and weaknesses of these models (and humans) are in the structures rather than at the token temperature stage — I expect a human to answer any given question slightly differently at different times, but asking the same doctor for a second opinion as the first is considered the setup for a joke rather than a sensible idea.
    - godelski 3y ago
      This is the assumption of a flat and smooth latent manifold. If either of those conditions are not met, then seeds can have a wild effect on the model output. There's only bounds on the latent structure being locally smooth or locally flat btw. This is the underpinning of such works as the Lottery Hypothesis but there's a whole rabbit hole to dig down here and it is why Alpha Go did so poorly in round 2. This is an important part of the discussion of generalizability btw. I do not think your analogy works out and it is inconsistent with your prior comment, because you're right that anyone asking GPT is akin to anyone asking the same doctor. The second doctor is a second model (or potentially a second GPT checkpoint depending on how we want to consider divergence).