1
$\begingroup$

A teacher of mine said something of about AI being self sufficient and operating on experience. They backed this statement by this example-a car programmed to avoid ramming into apple, mango and banna tree also stopped at a walnut tree which was not what it was programmed to do, thus ai proves itself to be capable of faulty reasoning or associative learning, which technically is not possible for ai as it can only generalize to some extent if it is trained on diverse data and designed to recognize patterns.

They also mentioned something regarding emotions, attributing it to the video released last year where Sophia (the robot) said that it wanted to have a baby (https://youtu.be/q6xWrhD7p5o?si=rEFI7ReRzEQ4e9DQ)

Which is bull and I know.

Problem is that they insist that I provide valid proof that's not too outdated. I would like your help in this regard. Please help me find articles or any books that are new and up-to-date so that I can shut his non-existent brain off before he corrupts others with his faulty teachings.

On the side note I would like to mention this whole incident started when they asked the question (what is the computer?) and my friend replied "it's a dumb machine" and he scoffed and laughed at the response as if hearing some funny joke. He also mentioned that while it might have been so in the past, the present is vastly different and that computer is no longer dumb and instead can think and do it's work on its own.

$\endgroup$

1 Answer 1

1
$\begingroup$

AI are based on the idea of distribution matching, no matter what you are doing.

Is it a discriminative problem? then match $p_\theta(y|x)$ with the true one, using examples
Is it a generative problem? then match $p_\theta(x)$ with the true one, using examples

The only one that is not under this umbrella, is Reinforcement Learning, which is indeed sold as one of the most promising ways to achieve AGI (big word, take it with a grane of salt)

Indeed, evolution can be formulated as a RL process, trying to maximize the probability of procreation (only the strongest ones, that are able to arrive to such moment, will keep their genes "alive")

However, this does not mean that we are at that point, not even close.

What your teacher is observing is just generalization, which is far from extrapolation

If you show to an NN a light that is turned on every time it rains, that NN cannot understand if it's the light that turns on when it rains, or if it rains when the light turns on (the only way is, for example for LLM, give enough common knowledge to the AI to reasonably deduce it)

However, on the other hand, if there is no evidence that any AI at the moment can extrapolate (which is not equivalent to "generalize"), there are some Researcher that thinks that LLM training is a sort of "compression", and the best compression possible will indeed lead to abilities of deductive reasoning (however, without interaction with the world, I still see some counterexamples)... check any podcast with Ilya Sutskever, cofounder of OpenAI, on ChatGPT abilities or LLM in general

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .