Comment by mjburgess | Hacker News Reader

Comment by mjburgess | Hacker News Reader

Parent Full thread

Approximately,

You "actually understand" if in saying words, w, about a situation, s, the reason, r, you say those words counts as as an understanding of s.

So if you're asked, "what is the weather today?" and a system replies, "its sunny" -- it only understands if the reason it replied was because its saying "sunny" had come about because it understands sunny sitations; and therefore has a justfiied reason on this occasion to say that this occasion is sunny.

If I ask a NN trained on a trillion documents, "do you like what I'm wearing?" it cannot answer. It can say the words, "yes I do!" but it cannot have a reason to say those words. It's just a "weighted average retrievla across a compressed dataset of a trillion documetns".

The question isnt "on average, what -- historically -- would a generic person reply to the question: do you like what i'm wearing?"

The question is if the systems likes what I'm wearing. Its replies are only "actual replies" if it can see what i'm wearing, has experiences/preferences for taste, has some aesthetic judgement, has a disposition to like/dislike -- etc.

No ML system on the planet, in this sense, has any understanding whatsoever. Interpolation over historical data is just a means of compressing history into parameters, ie., its a compressed lookup table. That isn't ever a /reason/, exactly, if a person simply used such a table -- they wouldnt mean what they said

Given that your squishy brainmeats also consist of neural networks that have been trained directly by years of experience and indirectly by billions of years of evolution, I can write off your response as just a weighted average retrieval across a compressed dataset of trillions of your experiences.

If, on the occasion I say, "I like you" my saying it is caused by my liking you -- then you can describe this process however you wish.

Since my liking you is caused by my immediate environment, it isn't reducible to a weighted average of my history.

Another way of putting it: the historical positions of all the molecules in some water aren't sufficient to determine its present state. It's state depends on its container (ie., the pressure & temp of its environment). And there are a very very large number of states of water, many still being discovered.

In this sense my state in any moment is a point in an infinite space of states -- not determined by my history. But also extremely complexly by my container -- my social, etc. environment. The world hitting my senses is doing more to me than the air on the water. It induces in me a state which cannot be "averaged" from my history.

Thus, no, we are not weighed averages of our histories. We are profoundly chaotic and organic organisms whose growth in our environments enables us to respond to our enviroments by entering a near infinite number of states. These states arent in our history, they are how our biophysical structure -- via history -- responses to the near infinite depth of the here-and-now.

We are more like water than a computer. A computer is a deterministic machine which is a deterministic function of its deterministic inputs. Water is a chatoic system whose state "isnt up to it". Water's state is /in/ its container, and water itself is a non-determinsitic chatoic soup.

The chaos of water is the least of what one nanometer of a cell has; a cell is a trillion times that adaptive and responsive. And we are a trillion of those.

We are a cascade of chaotic state changes provoked by an infinitely rich environment action on our bodies shaped by a long history of organic growth.

We don't have "neural networks", we have cells. That some form "networks" has nothing to do with what we are. A complete misdirection.

> Since my liking you is caused by my immediate environment, it isn't reducible to a weighted average of my history.

It is not clear to me that this cannot be the case of a weighted average very heavily weighted to the immediate past.

> The historical positions of all the molecules in some water aren't sufficient to determine its present state. It's state depends on its container (ie., the pressure & temp of its environment).

AFAIK, given that you could determine the momenta of the molecules from a history of their positions, this would be sufficient to determine its state (maybe you need their angular momenta independently?) The relevant information about the container has been impressed on the motion of the molecules.

Similarly, we can suppose that the history of your environment becomes manifest in your mental states (and a predisposition towards certain state transitions) - though, on account of the complexity of that environment, in a compressed form.

> We are more like water than a computer. A computer is a deterministic machine which is a deterministic function of its deterministic inputs. Water is a chatoic system whose state "isnt up to it". Water's state is /in/ its container, and water itself is a non-determinsitic chatoic soup.

And yet we can usefully model fluid dynamics on a computer, even though the mathematical representation of the problem is analytically intractable. This line of argument does not appear to be leading in the direction you think it does.

> We don't have "neural networks", we have cells. That some form "networks" has nothing to do with what we are. A complete misdirection.

I am generally distrustful of these ontological arguments - quite often, it seems, things that were once thought of as being completely different turned out to be similar in some relevant way.

Ok, but this is all irrelevant. The trained history of evolution is the Chinese book. The environmental interaction is the symbol being passed into the room.

Without your trained history, you would not be able to say "it is sunny" when the sun hits your skin. We know this for sure, as an elephant or a baby are unable to say it, while a Frenchman would say different words.

Me saying "I like you" is not caused by me liking you. It is caused by me deciding, based on my experience and my perception of the current situation, to utter these words. I may be lying, I may be exaggerating, I may be misdirecting. Human speech is extremely rarey caused directly by interaction with the (external) environment (exceptions would be onomatopoeia exclaimed when in extreme pain, or when surprised, which tend to bypass most computation in the brain).

Okay, but how do you test whether any speaker (human or otherwise) understands something without asking them to explain it? If you have some other test (like perhaps some analysis of what the human brain is doing when it is understanding something), great! But even then, why can't the computer do the same thing that the human brain is doing which constitutes understanding? Or if the only test is to ask the speaker to explain something to you, well then, the Chinese room can do that too! Of course you end up with an infinite regress where you can ask the speaker to explain the previous explanation forever, but that's true for human speakers as well.

The question is whether what causes us to understand things is a computational process.

See my comment to the other reply, an others in this thread about chaos/non-determinism to see why I doubt.

Ie., I dont think our organic growth and adaption to our environment, in being profoundly chaotic (and via QM, thef. non-deterministic) is likely to be describable as a computable function.

Non-determinism is a red herring. It's trivial to model and produce a non-deterministic computer given a deterministic one and since source of non-determinism (it's in fact impossible to produce a truly deterministic computer in real life).

Conversely, you can interpret QM to mean that the universe evolves perfectly deterministically, except that no particular observer can predict in which branch they will find themselves.

Also, non-determinism has little to do with computability. All known rules of physics are computable, even if you include the Born rule for non-determinism. A quantum computer is computable by a Turing machine (in exponential time, as far as we know).

All laws of physics aren't computable. None of them are, they are all defined over the reals.

You've a bigger issue with talking about the computability of those laws.

I am assuming that somehow the reals can be ditched (which, incidentally, seems unlikely). And in that reality, we still have a problem: the role the reals play is to make reality infinitely precisely measurable. Without this chaotic systems become non-deterministic over realtively short time horizons.

Nearly all actual systems are chaotic in many of their properties. It isnt a matter of just adding an RNG to a CPU. A trivial object has 10^30 atoms, in organic stuff, much of that is behaving chaotically on environmental boundaries. The engery to simulate this, assuming 10^30 RNGs (etc.) is vastly greater than the target system and "cannot outrun" that system.

Not only this but I think simulation of such things quickly surpasses physical bounds on physically possible computers, not least as the RNGs have an energy cost.

The real numbers used in physics are all computable (pi [0], e, sqrt(2) etc are all computable numbers). There are non-computable real numbers, but so far those have never been necessary for describing physical systems (e.g. Chaitin's constant).

When talking about computability in the absolute, theoretical sense, the size of the universe is irrelevant - all we ask is if there exists some arbitrarily large, but finite, Turing machine that performs that computation and then halts after an arbitrarily large but finite amount of time.

If we're taking about practical computability, than my base assertion that the physical world is itself a computer also implies that this computer is exactly fast enough to simulate physics. Since we know that the laws of physics are computable in the theoretical sense above, and since we know that the universe is accurately simulating itself, we have proven my assertion.

Your other points about chaos and non-determinism I have addressed in another thread.

[0] https://stackoverflow.com/questions/4134282/is-pi-a-turing-c...

Reply on news.ycombinator.com