Artificial intelligence is impossible | Hacker News Reader

Artificial intelligence is impossible

mindmatters.today

10 pointsyters19 comments

The human brain, like everything else, ultimately runs on the laws of physics. If the laws of physics can be simulated on a computer, then the brain is theoretically simulatable on a computer. So any argument that a computer cannot even in theory do some thing that a human mind can do must be flawed, unless the assumption that of the simulatability of physics is wrong.

(This particular instance is thinly veiled question-begging; it takes as its assumption that a human mind is special ("A defining aspect of the human mind is its ability to create mutual information"), and proceeds to conclude exactly what it just assumed. See also: sloppy dualism — http://goodmath.scientopia.org/2013/01/17/sloppy-dualism-den... )

Or, the mind is not the brain. No one has ever proven it is.

Sure, of course. That's a philosophical position known as Dualism[0].

The trouble is that what the author has done is base their argument on an assumption of dualism, but not made that assumption explicit. If made more explicit, the question-begging becomes more obvious: "If we assume there is something special and meta-physical about the human mind which is not replicable anywhere else, then artificial intelligence is impossible". Well, yeah. But putting it that way doesn't make for a very dramatic article...

[0] https://en.wikipedia.org/wiki/Mind%E2%80%93body_dualism

Author here. I would not say I am making the assumption that dualism is true, merely that it is a possible explanation. Given the human mind generates a whole lot of mutual information, seems dualism might be a good explanation in this case. There are also other possibilities. At any rate, the possibility the mind is a Turing machine is excluded by the premise that it creates mutual information.

I don't know, it looks like the conservation of mutual information is like the fact that entropy always remains constant or increases. This is true for the overall universe, but not true for individuals.

A human isn't some magical entropy reducer -- we reduce entropy in ourselves by increasing it elsewhere. I suspect mutual information works the same way.

That is a good insight. The law does seem similar to the 2nd law of thermodynamics. The form of the free energy formula looks a lot like mutual information.

So, this could go two ways. Either humans are not net entropy reducers, and also not information creators. Or, they are information creators, and thus are net entropy reducers. Not sure there is evidence the latter is false.

A broader point is that regardless of whether the human mind is an entropy reducer, something in existence is an information creator/entropy reducer, otherwise we wouldn't have information/low entropy to begin with.

Mmmh, that would imply that the brain is not a Turing machine (or compatible). I'm not convinced as the author does not mention what the brain is instead.

If the mind creates mutual information, then it has to be a limited halting oracle.

#relevant

https://en.wikipedia.org/wiki/G%C3%B6del%27s_incompleteness_...

Utter junk. Does anybody find articles like this convincing? This has about the same intellectual content as YouTube “proofs” that the earth is flat.

Physics (randomness + determinism) cannot create mutual information. This has been proven. So, there must be some other causal entity that creates the mutual information we see around us, and this causal entity cannot be reduced to randomness + determinism.

I saw a similar (junk) article a while ago about how our minds must be special and can't be replicated on silicon because we can implement pure functions in our minds, but that is impossible in any stateful machine. The author's point hinged on the idea that when I add `2+2`, I am doing the platonic ideal pure operation of addition which is demonstrably false so the entire argument falls away.

Similarly, it hasn't been established that our minds possess mutual information in the ideal sense. One should first note that there are well known Bayesian methods for approximating mutual information which sound much closer to what our brains do operationally.

Now consider the fact that if you actually were creating mutual information when you looked at a sign, it would be impossible for you to be mistaken about the implication of the sign! You would infallibly understand every streetsign that you read because your undertanding of that sign is mutual information. This is demonstably false. People are mistaken about trains of reasoning they make all the time, in fact this article is a fantastic example of that.

Given that, it sounds much more like our minds are Bayesian machines approximating mutual information but not creating it.

Not sure why you make the inference that because we don't have the most elegant encoding of the information that absolute mutual information has not been created. As it is, any approximation of absolute mutual information is predicated on the existence of said information.

> As it is, any approximation of absolute mutual information is predicated on the existence of said information.

That's not true. I can approximate a circle on a piece of paper or with a computer or whatever without there ever actually being a true, platonically ideal circle in the entire universe. Similarly, I can construct a heuristic which approximates shared information without that shared information existing in the first place.

That is true, but algorithmic information theory would say the Platonic circle is the best explanation for those approximations, so the approximations are instances of the absolute.

So, uhh, how do humans get mutual information? Big gap in the argument there, especially if you have a determinism bent in your philosophy of physics (in which case the article proves that humans can't create mutual information either)

The point is randomness + determinism cannot create mutual information. Yet humans create loads of it. So, by definition, the human mind must be something other than randomness + determinism. One good candidate is a halting oracle.

I get the point they're trying to make, I just don't think they've done anywhere near a good enough job showing that humans create mutual information (rather than just transforming it)

Author here. Yes, I did not formally show humans create mutual information. That part of my argument is just an intuition pump. The article was meant to illustrate why artificial intelligence could well be impossible based on one formal premise and one intuitive premise. The goal was to demonstrate artificial intelligence is not a foregone conclusion.

A broader point is that the existence of mutual information (i.e. all the order in the world) is evidence that something not Turing reducible exists, regardless of whether the human mind is Turing reducible or not.

Reply on news.ycombinator.com