Hacker News Reader

Comment by mjburgess | Hacker News Reader

If I ask, "how did you like new york?" and you answer the question, sure, you can have you `I`.

That requires having gone to new york, having tastes, having judgements, having memory -- having a subjectivity which is unique and composed of the history of your experiences which can be related to novel environemnts. It requres existing in a shared world we can both co-refer to with our words. It requires being embedded in a space of social and physical significance.

It isnt google. "Language models" just zip the internet and search for patterns in text already written about the world, and reports it.

It's one step above a tape recorder on playback.

View on HN

If the model responded, "I have not visited New York because I am a computer", would you consider that to be a good step?

If you require physical embodiment, that is fair - but I don't think there's any principled reason to expect a multi-modal neural network could not be embodied with appropriate sensors or actuators.

Transformer models are universal function approximators. They do not "search" - in the same way that a calculator does not "search" when you ask it to perform an operation.

Being a universal fn approximator is a pretty trivial property. The question is (1) does a fn exist to approx? (in most cases: No); and (2) how is the aprox constructed?

With NNs, fns are approximated by compressing a sample of points along an empirical dataset which stands in for the fn. In most cases, however, the fn doesnt exist.

There is no fn from "Image->Animal". The world is ambiguous, functions do not describe it. The different `x`s genuinely correspond to the same `y`, here: the same image (occluded animal) could either be a dog or cat. There is no function to approximate.

Even when there is, compressing empirical samples along a trivial interval is a terribly fragile method: consider sampling the addition function (x,y)->x+y over anything less than the infinite domains of its inputs.

Starting with an empirical distribution of "the solved problem" will never build an intelligent system. Intelligence is what animals have because the world is ambiguous and not already ready-to-hand.

Intelligence is what you do when there isnt "functions lying around" to approximate. I doubt almost anything in CSci is even relevant to solving this problem. It's largely a bioengineering one. Being the agent in the world which manufactures functions by disambiguating it.

"Animal" is an arbitrary definition that exists in one's head. Unless you assume an idealist position, the only way it can be defined is by a function that runs in one's head. It might be a very complicated function - sufficiently complicated that we can't really articulate it and resort to "I know it when I see it", and possibly self-modifying - but the same person will still produce the same output consistently when given the same input.

> Transformer models are universal function approximators.

You have to be very careful with the wording here. The class of Transformer models is a universal (continuous) function approximator. A single Transformer model can only approximate 1 particular function.

Also, backpropagation/gradient descent are not universal function approximator constructors - they can't be used to construct an ANN that approximates an arbitrary (continuous) function even with arbitrarily many random input/output pairs from that function's domain.

For example, given a function that is f(x) = floor(x) if floor(x) is even; x if floor(x) is odd, defined for any `double x`, there is no guarantee that doing back-propagation over a random sample of (x, f(x)) pairs will yield the correct function approximator for some arbitrary precision (especially if you don't start with the right architecture).

People confuse the "representational capacity" of `recurse(AW^X)` with the algorithm which sets `W`.

> That requires having gone to new york, having tastes, having judgements, having memory -- having a subjectivity which is unique and composed of the history of your experiences which can be related to novel environemnts. It requres existing in a shared world we can both co-refer to with our words. It requires being embedded in a space of social and physical significance.

are you saying that I would not be considered an intelligent being if I had a jar in a brain left with my thoughts since my birth ? I don't think that this is what humans mean by intelligence...

also I don't really see the fundamental difference between "tastes", "judgements" and "memory" - making a judgment, having an opinion on something, etc... is just stopping the brain from reasoning past a certain step and internally saying "this is good enough" most likely because this is tiring to our frail physical bodies. Like, I don't like eating eggs for instance, but that's just because I don't make the effort of convincing myself that I like them. There's no intelligence in this, just laziness.

Yes, you wouldnt be intelligent if you were literally a brain in a jar.. you'd have nothing to say. You dont have any "thoughts at birth". A brain absent stimulation from the environment, doesnt do anything.

The brain-in-the-jar example presumes that the whole nervous system receives identical stimulation as-if it were embedded in the world.

Incidentally, this is what people mean by intelligence... by since no one has studied the relevant science, people do not know what's required to provide it.

> The brain-in-the-jar example presumes that the whole nervous system receives identical stimulation as-if it were embedded in the world.

I should have chosen a similar sounding thing, I really meant what I said, not the original brain-in-a-jar thought experiment.

> Incidentally, this is what people mean by intelligence... by since no one has studied the relevant science, people do not know what's required to provide it.

I don't understand how these two sentences work together: the word's meaning is defined by its usage. If tomorrow most people in the world says that being intelligent is defined by waking up at 6AM and doing 10 jumping jacks every day, then that is what the word will mean whether you want it or not.

In my "human frame of reference", "Yes, you wouldnt be intelligent if you were literally a brain in a jar.." really does not square with how intelligence is talked about for instance so... ¯\_(ツ)_/¯

You name things intelligent; then you have a theory as to why you name them that way.

Your theory is wrong.

Those thing which count as intelligent aren't intelligent because they're finding statistical associations in historical datasets that have been pre-made and pre-disambiguated by people.

You don't know this, because any time you theorise to explain, you require empirical investigation and decades/centuries of science. You cannot be "right by definition", you're not omniscient.

The reason that animals are intelligent isnt their cognition, narrowly. It's their bodies, broadly, which produce the input cognition requires.

In "doing all the hard work" by manufacturing data for machiens to process, we can produce computational models of cognition -- but these immitate intelligence, precisley because, the thing which we call intelligent is the very production of that data.

> You name things intelligent; then you have a theory as to why you name them that way.

but here a human being (and I suspect not the only one) names a computer program as "intelligent". The theory has to take this case into account to make it fit in its definition of intelligence: the will of human beings naming things overrides everything else. Maybe there is something else that humans have and this program doesn't, but another name has to be found for it if enough humans agree that this fit as what "intelligent" means.

> You cannot be "right by definition", you're not omniscient.

This doesn't make sense. Definitions are outside rightness / wrongness.

Ah, well if you start this project by calling present AI intelligent, then you're right.

I take the A to apply to the I, ie., AI is artificially Intelligent. Its not actually intelligent.

We didnt arrive at a notion of intelligent via digital computers. We have been talking about it for thousands of years, at least.

I am talking about the pretheoretical, intuitive notion of intelligence that doesn't prejudge the question of AI. What is it we call intelligent?

Intelligence is a kind of rapid adaption and coping to environmental change. It's a way of being responsive to an environment. Its a kind of radical underminism of response: a thermometer's response is determined. An animal, in having reasons to do things, is radically under-determined.

And thus, in acquriing specific hisotries, animals bring their own reasons to environemnts, and thus are able to cope with a vast amount of sudden change that would, eg., destroy a thermometer.

And what we are concerned with, both with real AI, and with animals is systems which are in the world with us, in this way: that are likewise radically able to cope.

> What is it we call intelligent?

> Intelligence is a kind of rapid adaption and coping to environmental change.

Who's the "we", there ? I have absolutely never heard anything remotely close to the definition of intelligence given there

Exactly in some sense we (collectively) are doing most of the hard work, by providing these models with a vast number of samples of "intelligent" behaviour and expression. An embodied system wouldn't be able to learn anything as it stands, because none of the "bootstrapping" steps humans do are adequately captured by current training methods.

Human languages exist to compress the world around us in a way that will survive being sent over sound waves and uncompressed in another brain that experiences a similar enough world around them.

Meaning that with enough words and their relations, sufficient complexity forms for its processing to result in intelligence.

Remember that an AI can process far more textual data than we can. It is still orders of magnitude smaller the amounts of information we process through our senses, but it is still a start.

> I really meant what I said, not the original brain-in-a-jar thought experiment.

In that case please explain your thought experiment in more detail. In reality if we scoop your brain into a jar your brain tissue will be dead before it hits the jar due to the lack of bloodflow. Then after a while it becomes mouldy and stinky and even more goopy than it was before. Not sure anyone would call any of that inteligent.

Reply on news.ycombinator.com