There's a famous paper by Dijkstra where he claims that using anthropomorphic terms for computers is a sign of immaturity of the discipline. I used to think that was a bit extreme, but the more people keep talking about fucking AI, the more I'm convinced he was actually right.
Machine learning is linear algebra, nothing more nothing less. Making a model, using it wrong, and then complaining that "the model failed" is a unique kind of stupidity that's becoming more and more popular with "hoi polloi" due to garbage articles like this.
I hate to be that guy, but most ML is definitely not linear algebra. The most popular and useful ML algorithms are boosted trees and random forests; they're more or less geometric approximators (which are all nonlinear). And as ANN nerds like to say, there's plenty of nonlinearity in neural approaches.
But mostly I agree with you: modeling is horse shit, worshipping models (linear and otherwise) is childish idolatry and Dijkstra was completely correct about just about everything (speaking as an APL bum).
> I hate to be that guy, but most ML is definitely not linear algebra...
Yes, I'm not disagreeing with any of that (see also my other reply), it was a misuse of terminology on my part.
> But mostly I agree with you: modeling is horse shit, worshipping models...
Absolutely. The way I see it is that models are like theorems: they are useful and worthy of being studied, but if you forget to validate that your hypotheses are correct, you get wrong results: garbage in, garbage out.
An alarming amount of stuff these days is the equivalent of people trying to apply the pythagorean theorem to a non-rectangle triangle and then complaining the results don't match.
Good ponits you guys. I think it would be fair to say that "AI" is nothing but algebra ran by an automata. The rest is a mix of hype and nerds taking advantage to manipulate people that has this concept out of their intellectual reach. Depending on the case, this can enable evil at global scale. As always with humanity, ideology is always a problem.
This kind of argument is common in this sort of discussion, and more than a bit reductionistic. Extreme complexity often comes from very simple processes at scale.
Your computer's abilities are the result of little switches flipping between two states.
> Your computer's abilities are the result of little switches flipping between two states.
I don't think this is a reductionist view of a computer. That's exactly what it is, and with a very little imagination describes what it has the capacity to do. What computers do now wouldn't be a surprise to people in the 1940s.
It is reductionistic, because so much comes from how you arrange those switches, the current state of the switches, and the sequence in which you tell those switches to switch. And I think most in the 1940s would be very surprised at what computers can do now. It only takes little imagination because we’ve already seen it in action.
Similarly, you can say that deep learning is just a series of matrix multiplications with nonlinearities applied, which is true, but it certainly wasn’t obvious to most that when you scale that way up, that that would lead to computers being able to interpret images, sound, and text. For example, my AI course professor described neural nets as something of interest mostly for historical reasons, something that was tried, but was an evolutionary dead end. And he was no slouch.
The hype has gotten ahead of its current abilities, but I simultaneously think that it will be one of the most impactful inventions we’ve ever created.
> The hype has gotten ahead of its current abilities, but I simultaneously think that it will be one of the most impactful inventions we’ve ever created.
This statement is one of the things I find most fascinating about the current state of AI. Do I think deep learning is incredibly useful? Absolutely and I look forward to seeing more awesome stuff done with it in the future.
Do I think the singularity or "true" AI will be based on deep learning? Not really. I strongly suspect anything capable of reaching scifi levels of intelligence (which a lot of hype says or strongly implies is imminent) will have to be a difference of kind, not of degree.
I could be wrong, of course, but it will be interesting to wait and see.
Yeah, I tend to be more in the difference in degree camp, with the caveat that it's pretty clear we're at least a few discoveries of structural mechanics away from mimicking human-style learning. We need to make advances in semi-supervised learning, as well as increasing sample efficiency, and reducing catastrophic forgetting. But my intuition says that the solutions will be pretty simple, and seem obvious in hindsight.
But I'm also guessing we'll come up with an endless number of complex tweaks to get the current stuff closer before someone figures out the simple mechanics and sweeps most of that away.
It manipulates people who can't grasp it. It sets up implementors for failure when the hype settles. To an extent, it's responsible for the peaks and valleys you inevitably see in funding.
I wouldn't say modeling is horse shit. It's really marketing surrounding modeling that's horse shit.
Modeling is actually very very useful but is too often oversold, which drives me bananas. It's especially bad in the world of data driven or data intensive models because that's trendy and sells software, hardware, and services. The problem is that various forms of modeling are complex and therefor expensive and convincing people to invest in furthering the domain results in these poor public misrepresentations and overselling.
Modeling gives us ways to abstract reality and attempt to predict the future or find patterns we might otherwise miss. Sometimes it works, a lot of times it doesn't, but when it does it's great. Everytime--its expensive and people need to feel they didnt waste their money.
And here's a famous old paper by McDermott that makes the same point but specifically for AI. Using terms like "understanding" to mean the effect of a certain algorithm is fooling yourself into thinking you've accomplished something you haven't.
> There's a famous paper by Dijkstra where he claims that using anthropomorphic terms for computers is a sign of immaturity of the discipline.
If you take the word "computer" out of the sentence this statement is likely backwards. Putting the word computer back in doesn't change that.
Look, for example, at Dennett's seminal book, "The Intentional Stance" for a philosophical discussion of the expressive and reasoning power behind statements like "The thermostat tries to keep the temperature at 60 degrees".
Exactly. Anthropomorphic language is such a useful and normal way of thinking and expressing ourselves.
If you say "the classifier thinks this roadsign is a truck" everyone knows you mean "the classifier has labeled this roadsign as a probably a truck with a high confidence". No-one seriously would take you to mean the classifier thinks in the same way a person thinks any more than if you say "The toaster wants to pop up but it's jammed," they would take you to mean the toaster has a yearning like unrequited love.
If we remove our ability to use language in this way we give up on a mental model that is incredibly useful for simplifying lots of explanations and reduce ourselves to a very clumsy literalism instead. As long as we understand the limitations of this model and are clear that it's not literally what is going on it's very useful.
My problem is not with natural language in and of itself, it's obvious that natural language is extremely useful. I'm arguing one should be extremely careful, however. Natural language doesn't have deduction. It doesn't have definitions. It doesnt't have ways to restrict us from making vague statements. It is subject to non-unique interpretations.
This is particularly relevant when technicians and scientists speak to the general public. Let me quote one of your examples:
> "the classifier thinks this roadsign is a truck" everyone knows you mean "the classifier has labeled this roadsign as a probably a truck with a high confidence".
The fact that the classification is probabilistic is obvious to me and to you, but nothing in the natural wording suggests that, and it's a pretty important caveat.
It's not enough for us to understand the limitations of natural language, everyone in the conversation has to, for it to be a useful tool.
"Thinks" in that context is inherently probabilistic. No one uses the phrase "I think..." to literally state that they have a thought process going through their head, they do it to clarify that the following statement is based on their own conclusions rather than objective reality.
No one except Descartes: "I think therefore I am."
But all this does is underline that natural language is a vague mesh of concept categories anchored to words with multiple overloaded meanings.
One of the reasons philosophy exists is because sometimes it attempts to untangle the overloading.
The argument about AI is no different, but with the complication that the marketing concept mesh has different requirements to the research concept mesh, and both are different to "Make this work by next Tuesday" mesh.
Google used to have an ancestor of word2vec which tried - very crudely - to map some of these meshes. Predictably it was much better at mapping word meshes than concept meshes, but it was still an interesting approach.
The underlying problem is that we literally don't have a clear representation system for these concept meshes. Words are overloaded, code and math are too specific in their different ways. The various shades of ML are still too concrete.
So we're going to keep having these debates until someone invents a representation system for natural language concept maps that isn't based on verbal approximations.
Not arguing, just suggesting an alternative phrasing that might better convey the intent. When people say thinks in this sense they mean determines to a certain level of confidence, rather than modelling consciousness.
I think ML is a bit too complex (different approaches trees, DNNs,...) to oversimplify it like that. One could say, it's just math - but even experts can't really tell some of the why it works behind back propagation in DNNs. A lot of research is empirical, like switching to ReLU...
ML is a commodity now, even popular with artists to create art. Being a commodity means you dont have to understand the inner workings to benefit from it. I don't see anything wrong with that, but a lack of proof for biases/backdoors/bugs in models is an issue that I'm worried about when its applied too enthusiastically in critical areas.
Article is weak though, and hard to follow conclusion.
I knew an older programmer who used the phrase "parasitic interpretation" to describe the human tendency to assume a things' name related to its function. This is the idea that something called `loop_index` is the index for a loop, when it could actually be anything. He strongly felt that semantic content in variable names should be minimized.
I haven't run into the phrase anywhere else, and Google doesn't turn up any programming-related results that I can tell, but I've found the idea endlessly useful. In talking about programs and the goal of programs, we often use language that describes things we have not yet been able to construct. The language has a parasitic relationship to the thing it names - misdirecting our understanding of the machine it sits on.
I try to practice a lot of skepticism around language and computers. It is very easy to name a thing and much harder to make the machine behave in the way the language suggests.
You may like the article "Troubling Trends in Machine Learning Scholarship", Lipton & Steinhardt, 2018 https://arxiv.org/abs/1807.03341
Especially the trend of "misuse of language, e.g., by choosing terms of art with colloquial connotations or by overloading established technical terms"
Yeah, this is an interesting take. Broadly speaking I agree, but you have to be pragmatic about that. There exist some accepted conventions it's just senseless to break just to make a point, for example, if you are defining a ring, making + distributive over \* rather than the opposite as is usual is not inherently wrong, but I would classify that as bad notation.
Similarly in OOP, you could technically name getters in any way you want, but why would you not conform to the accepted "getValue()" standard?
Of course fooling yourself into thinking you have achieved things you actually haven't by clever usage of natural language is the opposite end of the spectrum, and it's exactly what I was ranting against.
I'm not claiming it's literally a linear model, obviously. The point is that they are relatively simple mathematical models (that actually use linear algebra inside, https://en.wikipedia.org/wiki/Artificial_neuron does the output formula remind you of anything?) different in scope/use but not conceptually from linear regressions.
Attaching (meta)physical meanings ("neurons", "intelligence") to any of that is dubious at best.
How is this different from saying that a brain is "just" a bunch of atoms, which individually are quite simple? It's when you put the individually simple pieces together and run an optimiser that you get something clever. The result of that optimiser is a complex and very large array of weights.
The difference is that we have neither the scientific knowledge nor the computational power to make an exact model of the human brain (nor we are anywhere close to either requirement, AFAIK).
In other words, we cannot pinpoint the exact sequence of operations and events that produce a certain behavior, while we absolutely can do that with ML models (you could in principle run machine learning models with pen and paper, obviously).
If you could show me a program that emulates my brain to a certain degree of precision (it's unlikely that such a program would be a deterministic one, by the way), I would have no problem accepting that my brain is not meaningfully different from that program.
"Intelligence" is a word that, etymologically and semantically, is related to human or human-like capabilities. You wouldn't say that a leaf floating on a lake is swimming, and likewise, claiming that computers are "learning" or "intelligent" is at best a thin analogy and at worst a mischaracterization of the process.
What's happening in my brain is something we don't have full scientific knowledge of, but we know it's not x86 machine code. While the two processes may be in many ways similar, conflating the two into this ill-defined concept of "intelligence" is a discussion about semantics more than anything else.
> "Intelligence" is a word that, etymologically and semantically, is related to human or human-like capabilities. You wouldn't say that a leaf floating on a lake is swimming.
The definition of words changes in response to increasing knowledge - just take 'energy' for example. One cannot establish truths about the world by arguments from usage. (On the other hand, to be clear, I do not think that the current state of AI merits being called "intelligence". What happens in the future is speculation.)
> What's happening in my brain is something we don't have full scientific knowledge of, but we know it's not x86 machine code.
The introduction of x86 machine code at this point seems to be moving away from your original claims about "AI" being "just" relatively simple (though not simply linear) mathematical models, which are not "just" machine code either. The interesting (and very much open) question is how much of intelligence can be modeled in this way, and what else, if anything, is necessary.
The more you stress the simplicity of these models, the more intriguing their achievements seem.
> The definition of words changes in response to increasing knowledge
The usual process in mathematics and science is that you have a phenomenon that everyone agree exists but nobody can quite put their finger on it, so someone proposes a formal definition and if that definition turns out to be adequate, people work on the formal definition, and that's much easier because you now can use math, statistics, formal methods, etc; a prime example of this is the notion of "computability".
I don't believe that we are seeing the same thing with the concept of "intelligence", this is probably in part because it's much harder to capture the concept in a formal definition. Computers do computable stuff. Overlapping that notion with "intelligence" serves no purpose in my opinion: it explains nothing, it doesn't clarify anything, and it's certainly not obvious that the two are related.
> which are not "just" machine code either
I'm using "machine code" as proxy for "instructions/lambdas/whatever for a computational model of your choice", which they certainly are.
> The more you stress the simplicity of these models, the more intriguing their achievements seem.
It's not my intention to downplay any of the achievements of "AI". They are certainly not less intriguing when viewed from my perspective, the same way a compiler is not less intriguing if you think it's "just code".
My point is that any association of a formal concept (math, models, etc.) with philosophical concepts (intelligence, "truths about the world", consciousness, etc.) is always on thin ice, because natural language and formal concepts are hard to mix. Especially so when the concepts at play are so ephemeral.
In the past, when a construct like 'intelligence' has been hard to pin down, science moves on--leaves it to 'philosophy' and works with formal definitions.
Would you say that's one part of your point? Just to clarify, I am only responding to that part of your point.
By way of example, behaviorists declared a strict subset of the human experience to be in the purview of scientific study--and that may even have been just fine for that era.
You seem to be sweeping something important under the rug because it's hard to pin down, and saying that this is what science has done in the past--and if so, you're right.
But there's an assumption--an assumption that it is safe to sweep things under the rug like that. That assumption may prove false, and if it does, we're screwed.
Convinced that AGI is not something to worry about? Fine--but surely you agree there's such a thing as an information hazard? That is, information that can be deadly in the wrong hands: like how to create the next COVID, or how to make a nuclear bomb. In past eras of human history, knowledge was not as powerful. Today and ever more so in the future, whether humanity can Get Things Right will matter.
So from my perspective, it doesn't matter that whatever 'intelligence' is, is hard to pin down: it's still got to be figured out, whether or not it's difficult.
> In the past, when a construct like 'intelligence' has been hard to pin down, science moves on--leaves it to 'philosophy' and works with formal definitions.
Yes, that's part of the point, your wording is better than mine. If we're sticking to a purely historical perspective, this is definitely what happened. Most disciplines that today we (rightfully) regard as fully independent, originally splintered off philosophy (the most obvious examples are mathematics and physics, but even something like economics, in spite of having become more formalized recently, undoubtedly originates from moral philosophy).
> You seem to be sweeping something important under the rug because it's hard to pin down, and saying that this is what science has done in the past--and if so, you're right.
I won't deny that strictly speaking there is a bit of inductivism at play here. Historically, the scientific approach of limiting the domain of discourse to a tractable subset has been so much more productive and successful than any alternative that my, as it were, "bayesian prior", is that we should replicate the same approach if possible at all.
> So from my perspective, it doesn't matter that whatever 'intelligence' is, is hard to pin down: it's still got to be figured out, whether or not it's difficult.
This is a reasonable position, but wouldn't you agree that it's more of a "moral intuition" (not that there's anything wrong with that!) than a position regarding how ML results ought to be interpreted? As such I have no real counterpoints to offer, except perhaps an utilitarian point of view: are you really sure that banging your head against this very specific wall is the most productive thing to do?
Most problems I see with AI arise from either flat out using the models incorrectly (i.e. mathematically wrong, not ethically wrong, which is what I was pointing out in my previous comments) or from already familiar "political" problems, i.e. incentives, transparency, privacy, openness of the decision-making process.
The good news is that none of that is new. The bad news is that our track record as a species on problems of that kind is abysmal. I doubt that, in any case, trying to halt scientific progress makes sense.
People learn to hit a target by changing the structure of their brain to fit the task. Computers become better at hitting a target by changing a data structure. That seems directly analogous to me. Critically, learning doesn’t imply the ability to perfectly execute the task.
Your response on definitions actually supports my point on the matter: definitions follow from knowledge ("a phenomenon that everyone agree exists") and are modified in response to new knowledge ("if that definition turns out to be adequate..." - and if not?) As before, "energy" stands as an example of how it works, and "computability" did not enter the lexicon until there was a use for it.
Nevertheless, I agree that in the specific case of current AI, using the word "intelligence" is misleading. I do not, however, think this misuse has any serious consequences, as, to reverse how I put it before, usage does not establish truths about the world.
>> which are not "just" machine code either
> I'm using "machine code" as proxy for "instructions/lambdas/whatever for a computational model of your choice", which they certainly are.
Then that is an unfortunate choice of proxy, unless, perhaps, you intended to imply that it is a priori impossible for intelligence to be created by running x86 machine code. It was not clear to me whether, by introducing machine code into the discussion, you were not making some sort of argument from incredulity against the possibility of AI.
> My point is that any association of a formal concept (math, models, etc.) with philosophical concepts (intelligence, "truths about the world", consciousness, etc.) is always on thin ice, because natural language and formal concepts are hard to mix. Especially so when the concepts at play are so ephemeral.
At least since Newton, mathematical models have proved very useful in discerning truths about the world. Are we to just assume they will not work for the biological phenomena of intelligence and consciousness?
> Your response on definitions actually supports my point on the matter
I'm afraid I failed to understand your point, then. I don't have a problem with what you said there.
> using the word "intelligence" is misleading. I do not, however, think this misuse has any serious consequences
This is where I fundamentally differ. Its misuse implies a connection between a formal model (algorithm expressed in a computational model) and a philosophical concept (intelligence) that's dubious at best. On a conceptual level, this makes it harder to reason clearly about those fundamentally mathematical and abstract concepts, and on a concrete level, it misleads the public at large, implying that certain goals have been reached when that's plainly untrue. That's pretty "serious" in my book.
> Then that is an unfortunate choice of proxy, unless, perhaps, you intended to imply that it is a priori impossible for intelligence to be created by running x86 machine code.
Again, I'm afraid I don't understand your objection. It's widely accepted that all reasonable computational models are equivalent. Citing x86 was colorful language, it has clearly no bearing on the point at large. Machine Learning algorithms are clearly computable, which means they are expressible as Turing machines, terms of a classical untyped lambda calculus, Python scripts, C++ template metaprograms, or anything else. They are literally just programs.
> At least since Newton, mathematical models have proved very useful in discerning "truths about the world." Are we to just assume they will not work for the biological phenomena of intelligence and consciousness?
I certainly believe mathematical models to be useful, you would be hard pressed to say otherwise.
The ontological status of scientific theories is however at the very least a debatable topic. One needs not believe Newtonian mechanics is ontologically true, it's a tenable position to claim it's just a model, and we accept that model because it's useful.
Specifically, one could easily argue that Newtonian mechanics is false, because, for example, it fails to accurately predict Mercury's orbit.
Similarly, one needs not believe ML is anything more than relatively simple math to find it useful.
> I'm afraid I failed to understand your point, then. I don't have a problem with what you said there.
You have to go back a couple of posts to see the point. There, you wrote "'Intelligence' is a word that, etymologically and semantically, is related to human or human-like capabilities. You wouldn't say that a leaf floating on a lake is swimming." As we are now agreed that definitions follow from knowledge and are modified in response to new knowledge, it would not be somehow wrong to extend the concept of intelligence to a certain class of machines, if it turns out to be useful and informative to do so.
> They are literally just programs.
I take it, then, that you don't agree with the sort-of Platonist view that algorithms have an existence independently of any implementation? I'm on the fence, myself, but lean towards the Platonist side.
Regardless, it follows from your position here that your original statement "What's happening in my brain is something we don't have full scientific knowledge of, but we know it's not x86 machine code" can be rewritten as "What's happening in my brain is something we don't have full scientific knowledge of, but we know it's not computable." - but while the former is true, the status of the latter is not yet decided, so they are not identical propositions.
> One needs not believe Newtonian mechanics is ontologically true, it's a tenable position to claim it's just a model, and we accept that model because it's useful.
One could say the same about specific ontologies - they are as subject to revision in the face of increasing knowledge as are both mathematical models and individual words - and if it turns out that a mathematical model of biological intelligence or consciousness is effective and useful, it would be tendentious to imagine an ontological line between that model and intelligence.
> it would not be somehow wrong to extend the concept of intelligence to a certain class of machines, if it turns out to be useful and informative to do so.
No objections, but it's a pretty big "if". You could restate my point as "there is no evidence it is in fact useful".
> I take it, then, that you don't agree with the sort-of Platonist view that algorithms have an existence independently of any implementation? I'm on the fence, myself, but lean towards the Platonist side.
I don't, my position is essentially formalist. While I believe most research mathematicians would side with me here, your position is absolutely valid; famously, Kurt Godel was a Platonist, as are many others. My only observation here is that even from a Platonist point of view you are not really rejecting formalism, at least in the sense that while you don't agree it's the only view, I find it impossible to argue that one can't view mathematical objects as formal constructs.
> "What's happening in my brain is something we don't have full scientific knowledge of, but we know it's not computable." - but while the former is true, the status of the latter is not yet decided, so they are not identical propositions.
No, I don't really hold that view. I probably worded that badly. My position is that the latter is unknown, and I would be content to accept that my brain is not fundamentally different than an algorithm if you showed me an algorithm that can effectively emulate my brain within an acceptable margin of error. Ironically, if it were possible to do that, it would be proof there is no "intelligence", only "computability", making the first entirely redundant.
> One could say the same about specific ontologies -
Again, this is really thin ice. I don't really know what to think about this because it's getting too abstract for my monkey brain, but it's certainly not obvious that a better model implies an ontological line between itself and reality. To put it bluntly, is a better model really "more true", i.e. qualitatively different from a worse one?
Can you specify what you mean by "emulate my brain within an acceptable margin of error"? What would this mean as an actual experiment? Depending on your answer, I think we can actually test your implicit proposition that no such algorithm exists.
I'm not actually claiming no such algorithm exists, as I have already stated, but simply that such a thing is not known to exist.
I'm not an expert in neurosciences so I can only give an informal description. Let's also remove "me" from the equation, let's talk about a randomly chosen human H. We know for a fact that there is nothing physically special about human brains, they are just ordinary organic matter. This matter forms a system subject to the laws of physics. With enough computational power and scientific knowledge (we have neither as of now, AFAIK), we could write a program for a quantum Turing machine that runs a 1:1 simulation of H's brain in software. Any quantum program can be emulated by a Turing machine equipped with sufficient random numbers with at most an exponential slowdown, making this program computable in exactly the classical sense.
My questions are (1) is it possible, even in principle, to make a program of this kind? (2) Would such a program be sufficiently predictive (with any statistical notion of that concept you prefer) of H's behavior?
If there exists a program that satisfies both (1) and (2), then I'm content with the notion that I am, myself, not significantly different from such a program.
Hmm, this still doesn't quite answer my question at the level of concreteness I was looking for. But thank you for clarifying.
The thing is, you can only define predictive accuracy relative to some experimental design. Otherwise you can always claim that there is some unknown, unperformed experiment where the predictions of the model and your actual behaviour would diverge to a greater degree than is permissible by your accuracy threshold, no matter how many successful experiments have already been done in constrained conditions.
Imagine a task where you have to classify images as being of dogs or non-dogs. We can already train a model that can almost perfectly predict the choices you would make during the runs of such an experiment. But we obviously wouldn't call such a model a "model of your brain"!
My question is this: what would be a sufficient experimental design or empirical criterion to decide that some program is a model of you? The loosest criterion I could imagine would be something like "can successfully deceive your loved ones into believing they are you in a single text chat of unbounded duration with some extremely high success rate." Recent advances in NLP lead me to believe that we'll be able to reach at least this level of fidelity quite soon.
To be fair, qsort is not insisting on seeing an algorithm that is metaphysically identical with a human mind, nor claiming that such an algorithm would be a p-zombie, devoid of subjective experiences, which are both positions that you might find from dualist philosophers.
> You could restate my point as "there is no evidence it is in fact useful"
And if you had originally stated your point that way, I would probably pointed out that there is equally no evidence that it will not be useful, if it turns out to be the case.
> ...but it's certainly not obvious that a better model implies an ontological line between itself and reality.
Clearly, I failed to get my point across, to the point where I cannot guess where this question is coming from. Let's see if I can be clearer...
My position on the definitions of words is that they are contingent on our knowledge and that new knowledge can change our definitions (I gave 'energy' as an example, and you appear to have accepted this point a couple of posts back.)
I take the same view of ontologies; the categories we see are contingent on what we know and may change as our knowledge increases. This should not be surprising, given that ontologies are specific cases of words with meanings that nominally pertain to how the world is. There is no implication here that a better model implies an ontological line between itself and reality; rather, the point here is effectively a "so what" reply to your statement, "it's a tenable position to claim [Newtonian mechanics] is just a model, and we accept that model because it's useful." Mutatis mutandis, as they say, and consequently, there is no justification for holding on to old ontologies if new facts suggest a better alternative, any more than there is for models or theories.
> To put it bluntly, is a better model really "more true", i.e. qualitatively different from a worse one?
Did you mean to write that, especially given that, in your previous post, you offered an argument for the proposition that "Newtonian mechanics is false, because, for example, it fails to accurately predict Mercury's orbit." If there is a relevant point here, I think it is that "more true" models are qualitatively better (and quantitatively better, also.)
> Ironically, if it were possible to [show an algorithm that can effectively emulate your brain within an acceptable margin of error], it would be proof there is no "intelligence", only "computability", making the first entirely redundant.
How so? If "intelligence" is a useful concept now (and your objection to "artificial intelligence" seems to be predicated on it being so), when we do not know if the mind is computationally modelable, why would this usefulness necessarily vanish if this turns out to be the case?
> And if you had originally stated your point that way, I would probably pointed out that there is equally no evidence that it will not be useful, if it turns out to be the case.
No objections, but isn't it a bit weird to argue that we should do that just in case it might be useful someday? We'll deal with it when it comes up.
> Clearly, I failed to get my point across
I'm sorry, I'm pretty sure there's an argument but I just don't get it. I'm not really following the train of thought anymore.
>> I'll take the seventh.
>? - I'm not familiar with this expression.
> No objections, but isn't it a bit weird to argue that we should do that just in case it might be useful someday?
That is not an argument that we should do that just in case it might be useful someday, it is just a response to the quoted statement. As you know, I am not in favor of the current usage of "intelligence" in AI, and the only thing we differ on in that regard is whether it matters much.
> I'm sorry, I'm pretty sure there's an argument but I just don't get it.
It is an argument that ontologies are not privileged, canonical or fixed ways of representing the world; they have to conform to current knowledge as it evolves, or be replaced, and they are only interesting if they can "earn their keep" by being useful. Consequently, I do not think your argument from ontology, that this abusage of "intelligence" is a big deal, is definitive.
It’s weird that you compare AI to the human brain. I think the goal is here in the long term is to surpass the mere human brain. For that reason it’s stupid to try and model it at all. Maybe some future neural architecture will make the human brain look like a computer built using vacuum tubes.
I don't disagree with this. It was a misuse of terminology on my part, my broader point has absolutely nothing to do with the mathematical details of neural networks.
Reason for the nitpicking is, that for me - as a trained mathematician - and seemingly some other commenters, the gap between linearity and non-linearity is yet inside mathematics quite substantial. This is true also in general. Non-linearity might be not sufficient, but it is certainly necessary for any definition and/or serious discussion of AGI.
Just insert the words "these days" and "mostly", and you're pretty much covered. Also a honorable mention for calculus perhaps. Even decision trees use gradient boosting very commonly.
I searched for the "famous paper" in question and could not find it - can you point us to a link?
So far I find two of Dijkstra's talks/writing on anthropomorphism
I thought gradient descent was mostly calculus, not linear algebra. I was under the impression linear algebra was used to frame calculations so that GPUs could be utilized (since GPUs are very good at LA operations)
Machine learning is linear algebra, nothing more nothing less. Making a model, using it wrong, and then complaining that "the model failed" is a unique kind of stupidity that's becoming more and more popular with "hoi polloi" due to garbage articles like this.