What we call AI today is nothing more than a statistical machine: a digital parrot regurgitating patterns mined from oceans of human data

Prove to me that this isn’t exactly how the human mind – i.e., “real intelligence” – works.

The challenge with asserting how “real” the intelligence-mimicking behavior of LLMs is, is not to convince us that it “just” is the result of cold deterministic statistical algoritms running on silicon. This we know, because we created them that way.

The real challenge is to convince ourselves that the wetware electrochemical neural unit embedded in our skulls, which evolved through a fairly straightforward process of natural selection to improve our odds at surviving, isn’t relying on statistical models whose inner principles of working are, essentially, the same.

All these claims that human creativity is so outstanding that it “obviously” will never be recreated by deterministic statistical models that “only” interpolates into new contexts knowledge picked up from observation of human knowledge: I just don’t see it.

What human invention, art, idé, was so truly, undeniably, completely new that it cannot have sprung out of something coming before it? Even the bloody theory of general relativity–held as one of the pinnacles of human intelligence–has clear connections to what came before. If you read Einstein’s works he is actually very good at explaining how he worked it out in increments from models and ideas - “what happens with a meter stick in space”, etc.: i.e., he was very good at using the tools we have to systematically bring our understanding from one domain into another.

To me, the argument in the linked article reads a bit as “LLM AI cannot be ‘intelligence’ because when I introspect I don’t feel like a statistical machine”. This seems about as sophisticated as the “I ain’t no monkey!” counter- argument against evolution.

All this is NOT to say that we know that LLM AI = human intelligence. It is a genuinely fascinating scientific question. I just don’t think we have anything to gain from the “I ain’t no statistical machine” line of argument.

Obligatory SMBC: http://www.smbc-comics.com/comic/2009-12-11

That’s perfect. You already know your lines!

It is fully possible, quite likely even, for models to both be “more accurate than humans” on average while at the same time suffer occasional “accuracy collapses”.

I very much understand wanting to have a say against our data being freely harvested for AI training. But this article’s call for a general opt-out of interacting with AI seems a bit regressive. Many aspects of this and other discussions about the “AI revolution” remind me about the Mitchell and Web skit on the start of the bronze age: https://youtu.be/nyu4u3VZYaQ