Comment on We need to stop pretending AI is intelligent

TheObviousSolution@lemmy.ca ⁨2⁩ ⁨weeks⁩ ago

• Mistic@lemmy.world ⁨1⁩ ⁨week⁩ ago

  They’re not. It’s a math formula that predicts an output based on its parameters that it deduced from training data.

  Say you have 2 sets of data.

  Y0 = 3, X0 = 1 Y1 = 4, X1 = 2 Y2 = 5, X2 = 3

  We can calculate a regression model using those numbers to predict what Y3 would equal to if X3 was 4.

  I won’t go into detail, but

  Y = 2 + 1x + e

  e in an ideal world = 0, that’s our model’s error, which is typically set to be within 5% or 1%

  If x = 4, then

  Y = 2 + 1×4 + 0 = 6

  Our model just predicted that if X is 4, then Y is 6.

  That’s what AI does, but instead of numbers, it’s tokens (think symbols, words, pixels), and the formula is much much more complex.

  This isn’t intelligence and not deduction. It’s only prediction. This is the reason why AI often fails at common sense. The error builds up, and you end up with nonsense, and since it’s not thinking, it will be confidently incorrect.

  source

  • TheObviousSolution@lemmy.ca ⁨1⁩ ⁨week⁩ ago

    Wikipedia is literally just a very long number, if you want to oversimplify things into absurdity. Modern LLMs are literally running on neural networks, just like you. Just less of them and without the same structure. It is also on average more intelligent than you on far more subjects, and can deduce better reasoning than flimsy numerology. Another thing entirely is that it is cognizant or even dependable while doing so. Modern LLMs waste a lot more energy for a lot less simulated neurons. We had what you are describing decades ago.

    source

    • Mistic@lemmy.world ⁨1⁩ ⁨week⁩ ago

      People don’t predict values to determine their answers to questions…

      Also, it’s called neural network, not because it works exactly like neurons but because it’s somewhat similar. They don’t “run on neural networks”, they’re called like that because it’s more than one regression model where information is being passed on from one to another, sort of like a chain of neurons, but not exactly. It’s literally just a name for a transformer model.

      I don’t know enough to properly compare it to actual neurons, but at the very least, they seem to be significantly more deterministic and way way more complex.

      Literally, go to chatgpt and try to test its common reasoning. Then try to argue with it. Open a new chat and do the exact same questions and points. You’ll see exactly what I’m talking about.

      source

      • TheObviousSolution@lemmy.ca ⁨1⁩ ⁨week⁩ ago

        Literally, go to a house fly and try to test its common reasoning. Then try to argue with it. Find a new house fly and do the exact same questions and points. You’ll see what I’m talking about.

        There’s no way to argue in such nebulous terms when every minute difference is made into an unsurpassable obstacle. You are not going to convince me, and you are not open to being convinced. We’ll just end up with absurd discussions, like talking about how and whether stochastic applies to Alzherimer’s.

        source