We can not unequivocally explain what thinking is (in a human sense), and what makes it vastly different from Cephalopod thinking or a ridiculously large NN 'thinking'.
Until we reach a body of knowledge in Neuroscience that can give such an explanation on is it that makes human thinking somehow special (other than scale) then it is equally religious to assume that human consciousness or human creativity are somehow unique (in nature) and irreproducible (with technology).
This is an argument from ignorance: you dont know what caused the universe, so maybe god did.
Well, we do know much more than enough to regard any use of NNs today as having nothing to do with intelligence; and likewise, that any model of correlations in symptoms of human social activity (text, books, etc.) as not relevant to modelling/simulating intelligence.
A neural network model is just a set of averages over training data, which in this case, are simply recording correlations in word frequency. There's no neural "Neural network", it's just some mean()s.
We could claim the human brain is just 100 million neurons arranged in a mesh were each neuron has their own 'activation potential'.
Neither your or my statement are wrong. But again until we can pinpoint what (human) intelligence is, then claiming squid or neural networks are unable to be intelligent agents is just as emotional as the people you criticize.
It doesnt matter what algorithm in this family we choose, they all just produce correlative models of the data given. So if we give a NN a library of books, it produces a function whose structure is just correlations of words in those books.
To believe that the structure of this function has anything to do with language, the brain, or anything else is pseudoscience. It isnt close to science, it isnt an open or a philosophical question, it's homoeopathy.
Language use in humans is not a function of correlations of word frequencies. We know this, as much as we know where mars is.
No human is shown a trillion books and arrives at an internal model of language corresponding to word frequencies in those books. We know that it would be impossible to do this, since words refer to the world -- and correlations between word frequencies dont.
Your comment here defending this is equivalent to a homoepath saying "we dont know everything about chemistry". We dont know everything, but we do know that diluting with water doesnt make substances more efficaious.
And we do know that the correlational structure of text in books captured with any algorithm whatsoever, has nothing to do with modelling language. We know any such system cannot use language, as at least, it has no model of how words refer to the environment.
However:
> Language use in humans is not a function of correlations of word frequencies.
I would argue that this is untrue. For example, people pick up on common expressions (and other behaviors) of those around them. If I hang out with people that say "catch my drift" a lot, I will probably also start using "catch my drift" more. I find it very plausible that "my internal model of language" is in many (but not all) aspects driven by the word correlations I observe around me.
I do wonder about the absoluteness of your statements though. Would a bed-ridden person that has no sense of touch/sight/smell etc. except hearing (i.e. only able to interact with the world via language) be "not human" to you? If not, how is "the environment" different for it and GPT-3 (other than context window)?
When I say "not a function of" i mean, explicitly, that the model does not have a term which depends on these correlations. Just as F=GMm/r^2 does not have a term for "image pixel weights".
Real Language is what wrote all those books, as much as gravity is what made the night sky. Pixels and word frequencies are symptoms, a few of an infinity of symptoms -- distant effects, whose correlations are not models of their causes.
You actually can't because, unlike neural networks (we know exactly what they do), we don't have anything even close to an exhaustive view of the functioning of the human brain.
How can you not be impressed by something that would have been considered inconceivable just a few years ago?
At no point since digital computers were invented was it inconceivable that we could continue prompt phrases by searching through similar phrases, near such prompts, in digitized text.
It was thought computationally infeasible, in just the same way that a harddrive would never be TBs, and so wouldnt be big enough.
But there's no magic here, and nothing suprising. It's presented by charaltans who have money to make as an achivement that it isnt. This isnt intelligence, and has nothing to do with it.
It can complete phrases by having seen TBs of similar ones no more suprisingly than google can. It is as limited as that process implies, namely, radically limited.
Likewise, it was never inconceivable that one could 'play' chess by searching through billions of chess permutations and selecting a simple high-value path. It was thought infeasible given, say, an abbacus to do it with.
We have never thought totally dumb solutions to these problems were unimaginable. They were just so dumb, that the only route to using them would be to somehow run a Ghz processor. Well, since we now have those, dumb partial solutions to dumb problems are possible.
Hard disagree here. Google can't do what GPT-3 does and it has access to a lot more than 350gb (size of GPT-3 model). I doubt you literally believe that.
> We have never thought totally dumb solutions to these problems were unimaginable. They were just so dumb, that the only route to using them would be to somehow run a Ghz processor. Well, since we now have those, dumb partial solutions to dumb problems are possible.
Did the GPT-3 architecture exist 10 years ago? Could an ANN from 10 years ago have produced comparable results, given access to the same hardware? I doubt so.
It basically seems that in your worldview, anything below "human level intelligence" is just unimpressive? Seems quite a high bar.
The mathematics of NNs are Victorian, it's just regression. The NN architecture is largely irrelevant: that only determines how easily the regression converges. The strategy is the same: just compress the data into a correlative statistical model, and estimate using statistical associations.
That's the wrong category of solution, and as a category, it's anchient. The hardware of modern digital computers is impressive, NNs are an absurd joke -- as is any correlative model of data -- a magic trick carried entirely by the performance of CPUs and HDDs.
Newton did not predict the existence of an undiscovered planet by finding correlations in the stars of the night sky. He first produced an actual theory of gravity.
Intelligence is theorising, ie., it is an explanatory process. Correlations in the frequencies of characters in books (we wrote!) is the basis for a neat trick.
There's a tremendous amount of pseudoscience in the claim that these are models of anything at all. Inasmuch as correlations in patterns of pixels in the night sky is not a model of gravity, and provides no means to "predict" anything other than pixel patterns of the night sky.
Back to my point, your bar is pretty high. Artificial intelligence to the level you describe will basically be a "world as you know it is now obsolete" type discovery. It will likely go down as the most impactful discovery of all time. Most of us have a lower bar for what counts as impressive.
I am, yes, anti-wonder; I do think people get hopped-up on wonder and that lowers their threshold for reasonable belief and they end up in (religious) absurdities.
My issue here, in these terms, is that the "wonder" associated with the magic trick of modern ML is preducing a gross cognitive bias in which people believe this is an example of intelligence; or otherwise, an example like us in some significant sense. This leads to pseudoscience, charlatanism and a religious techno-utopian faith.
At the heart of it is a wonder-scam in the vein of all miracle-based shysterism. The fool looks at the miracle, is impressed out of all reason, and concludes miracles exist.
For sure, I dont really trade in wonder and I have never found it to do anything other than profoundly distort people's beliefs.
It's like someone in crypto talking about a future crypto-based financial system whilst at the same time being unable to processes basically any transactions, stabalise any currency, ensure any regulation, deal with any fraud, etc.
Or likewise, an MLM health company talking about cures for cancer all the while hocking vitamin supplements.
The question of what properties animals have which enable them to implement intelligence isn't relevant to whether correlative statistical techniques which capture word frequencies are "intelligence" -- inasmuch as vitamin 'science' isnt relevant to curing cancer.
This gross level of chalatanism, wishful-thinking, scamming, delusion and out-right fraud is overwhelming.
Try telling that to Sutskever or Karpathy (or indeed a bunch of "rationalists" on twitter, at least a subset of which works in AI) because both of them have made the claim you assert you haven't seen.
For me the real question is how well can GPT-3 identify pseudoscience.
I think I answered it myself... I have been seeing people's fervour to attribute intelligence to AI as either charlatanism or misunderstanding... but I think, now, it's actually wishful thinking (and hence essentially religious).
Trading in the usual superstitious foundations of religions: "AI works in mysterious ways", "correlations in symbols are meaningful", etc.
I still don't know why people wish that AI were intelligent; but at least now I have a clearer sense of it. There's some utopian faith behind it which creates this odd urgency to "require" machines to really be intelligent. I'd be interested in knowing why this wish exists.
I can never quite put my finger on how irreligious tech people end up in these tech religions. I guess there's some attachment to there being "a heaven somewhere", which with tech, will be heaven-on-earth.
As human beings, we slowly evolve from one generation to the next but our tools can evolve much more quickly and it can give us in a sense, superpowers. Cars were invented and allowed us to travel great distances while Self-Driving cars (which will require some sort of AI) will reduce car accidents and provide mobility for everyone.
AI has the potential to greatly improve the lives of a vast number of people and so it might be looked-upon as some sort of "saviour" and that has some religious connotations but most people still would treat it as just a tool but possibly a life-changing one.
What makes you sure that your thinking process is different?
What's the difference between a model and complicated correlations?
The most basic, most broken down, most simplified answer is what is called "poverty of stimulus".
A NN has an abundance of stimuli, orders of magnitude greater than any biological being could ever possibly take in. Finding enough correlations (enough for productivity) in such abundance is not actually that surprising. What is surprising is kids being competent at speech after less than 2000 days of stimuli, often really low-quality stimuli.
With that said, my guess is that we'll probably have to take a few more hints from the biology of the brain before we are able to achieve "human level intelligence".
A contemporary CPU and/or GPU is 'a bunch of connections that are pre-encoded' - just not in DNA, but in silicon.
No, CPUs/GPUs are not relevant if we are interested in the speed of learning in relation to the quantity of stimuli processed. You could even compute an ANN's learning algorithm with a pen and paper and it wouldn't change its "learning speed" within that definition. Pre-trained weights would.
Regardless, this head start is probably not sufficient to explain the disparity between the brain's capacity to learn and modern ANNs. ANNs are probably just not a very good approximation of how the brain works, for now at least.
In any case, my question here is: how do you not go mad in this techno-religious environment where people (seemingly desperately and earnestly) equivocate correlations in written text with thinking?
I feel at once overwhelmed, despairing, and a sense of need to point out that it isn't true. It's the most obscene form of wishful thinking and superstition.
HNers seem capable of identifying pseudoscience when its an article in nutrition (namely, eg., that correlations in the effects of human behaviour are not explanations of it)... but loose all sense when "AI" becomes the buzzword.
Correlations in *TEXT* are not models of language, intelligence, thinking. This is pseudoscience. And the most extreme and superstitious form: thinking that correlations are meaningful because we find them meaningful.
A person using language uses it in response to an environment, so "Pass me the salt" is said on the occasion that one needs salt. The "language model", `language = person(...)` includes the environment as input (and many other things).
Correlations in pixel patterns of the night sky are not a model of gravity. Correlations in text are not a model of language. Correlations in the detritus of human behaviour are not models of intelligence.
But to my question: how does one cope in an environment of this kind of wishful thinking?