This is How Wrong Kurzweil Is | Hacker News Reader

This is How Wrong Kurzweil Is

asserttrue.blogspot.com

20 pointstechdog58 comments

Honestly fully-sentient machines in 16 years seems totally ridiculous on its face to me regardless of what kind of graphs and charts about technological progress can be shown off. While I'm not really up to speed on AI research, in the consumer space at least we've just barely achieved a semi-automated house-vacuuming robot and fumbling speech-recognition in call centers. Unless I'm totally off-base with what's capable today, emotions and intelligence in a decade and a half would seem to require breakthrough after breakthrough in a vast number of nearly unrelated disciplines; plus there'd have to be the commercial demand needed to fund it. (Space stations and flying cars may be technically possible today, but nobody wants to pay for them.)

Kurzweil has invented a few cool things but today the guy is essentially a talented sci-fi writer. While those kinds of people are important, I have no idea why so many people take him so seriously.

Breakthrough after breakthrough already happens today, and happens more and more often due to growing economies around the world, along with more people doing more reasearch with faster computers.

Also, academic research and writing software is significantly less resource-intensive than building a space station or flying car.

> Breakthrough after breakthrough already happens today, and happens more and more often due to growing economies around the world, along with more people doing more reasearch with faster computers.

This statement brought to you by the 1950s, 1960s, 1970s, ...

People always seem to assume that the typical "human" behavior (speech, reasoning, emotions, making plans) are the hard parts while walking around without bumping into stuff is the trivial thing. But aren't those just the things added by evolution at the last 3 million years or so on top of stuff that evolved for around 500 million years? So "just" being close to robots moving around nicely might already be 90% of the way. To get that far with AI we took around 60 years - so another 16 years is maybe not so little. If you take the evolutionary time-spans as basis we seem to be even closer already.

AI hasn't progressed because it hasn't been commercially viable in a lot of applications. PID, a bit of hardcoded logic, and a bit of common sense has been far more useful than a machine which will try to outsmart the operator, and screw up badly.

Now that it's actually becoming useful, things could progress very quickly. Think about how long it too the internet to be something more than a toy for geeks, and how quickly things started moving once it started getting a real a market.

People take him seriously because he has a good track record at making predictions.

Plenty of people also don't take him seriously because lots of his predictions turn out to be hand-wavey or quite obvious.

http://spectrum.ieee.org/computing/software/ray-kurzweils-sl...

Basing your idea on where the cutting edge will be in a decade on the current state of consumer products will be misleading. A roomba is a $300 piece of hardware with an embedded computer, IBM's Watson was a $3,000,000 system. Both Apple's Siri & Google's voice recognition blow the average call-center system away.

I'm not sure we'll have sentience in 16 years, if ever, but we're a lot further ahead than you give us credit for.

I'm curious how many people have actually read his books before attacking his positions. A lot of the criticisms about his predictions seem to be based around the charts and graphs he's used in his Ted talks and articles.

In his books, he goes into great detail describing each aspect of the GNR (Genetic, Nano, Robotics) revolutions that will contribute to the singularity. He's actually addressed a lot of the criticisms that seem to crop up in HN, etc. over and over.

Forgive me if your position was based on having read his books (especially The Singularity is Near), but it didn't seem like it. Especially considering that cracking general AI is specifically mentioned as not being a requirement to obtain greater than human intelligence (for example, augmentation of biological human intelligence, specialized AI coming together to solve general problems, etc.).

Also, his prediction for the Singularity is 2045 (32 years from now), so your figure of 16 years seems to be incorrect.

Edit: D'oh, you were talking about the 2029 figure mentioned in the article, not the date of the technological singularity. I'm an idiot.

Kurzweil's whole thing is exponential growth. He is fond of recounting that the Human Genome project was 1% done half-way through the project, seven years in. This in fact was right on schedule, they finished 99% of the project in the final seven years, most in the last year.

"seems totally ridiculous" is a ridiculous argument. Human brain being very bad at estimating exponential processes. Do you have anything more scientific?

Any way there are gonna be big changes in the near feature. Either the current progress in data, energy, etc. will stop/drastically slow down or in a few years will have what it takes to build AIs and move planets around.

Less inflammatory title: These are some cherrypicked reasons why Kurzweil might be wrong, ignoring reasons why Kurzweil might be right.

But that wouldn't make the front page.

Also, I'd be a little less critical if this article wasn't citing all this quantum consciousness pseudoscience as one of the main reasons why Kurzweil is wrong.

More importantly

The implementation details are not relevant

We don't need to know how something works in the lower level to reproduce it and especially copying is not necessary.

It may be helpful. Very helpful. But not essential

ARM does not copy x86 and yet it achieves the same goal.

How a gecko attaches itself to the wall was not known until recently, however it's possible to study how to do something similar without studying the Gecko. Sure, if you study it it's a shortcut.

And this is a very specific and low level example.

For higher level things, the levels of abstraction that even nature creates make the lower level details irrelevant.

The functioning of a wing is independent of it's a bird or an insect wing (it is dependent of size, shape and speed, not origin)

A Shark and a Whale are completely different animals, and yet, they have lots of similarities.

So I'm thinking Ray's success or failure has almost nothing to do with our knowledge of the brain.

Cherrypicked? He's describing how consciousness currently manifests itself. Which cherries did he forget to pick?

Also considering much of Kurweil's "evidence" is best-fit graphs of cherry-picked data points, it's kind of ironic to lob that the other way.

> if this article wasn't citing all this quantum consciousness pseudoscience

Yeah, I guess if by "all this pseudoscience" you mean a two mentions of two very minimalistic claims. It's not like he's fabricating Deepak-Chopra-esque word salad.

Cherrypicked? He's describing how consciousness currently manifests itself. Which cherries did he forget to pick?

Any cherry that doesn't fit his preconceived conclusion. Kurzweil has eloquently made his arguments in his books and elsewhere, which are summarily ignored here. But it's not like Kurzweil is some random crackpot that nobody supports, there are tons of scientists who agree with him. For instance, Henry Markham and his collaborators have made similar predictions about how their work will lead to human-level AI in the relatively near future.

Yeah, I guess if by "all this pseudoscience" you mean a two mentions of two very minimalistic claims. It's not like he's fabricating Deepak-Chopra-esque word salad.

Two brief bullet points out of a list of six brief bullet points is a fairly substantial chunk of the argument.

Why not supply some actual reasoning to make your points? Such as 'bullet point x' should be valued lightly because of 'insert evidence/arguments here'. Just calling things pseudoscience or stating that he has 'lot's of support from other people' only implies that you're being defensive.

> But to say that we will see, by 2029, the development of computers with true consciousness, plus emotions and all the other things that make the human brain human, is nonsense. We'll be lucky to see such a thing in less than several hundred years—if ever.

2029 may be a bit early, but I think that actually to say it will take at least several hundred years is nonsense.

Look at the timescales of scientific and technological progress. Pretty much 99% of all knowledge and technologies we use are less than 200 years old. Most of it is less than 100 years old. We went from zero to space in a single life time. And the progress is not steady, nor is it slowing down, it's accelerating. One thing Kurzweil is definitely right about is that people don't understand exponential growth. Or any superlinear growth for that matter.

It's a bit presumptuous to think a sentient AI/machine has to replicate the human brain. While there are complexities we still do not understand there are also many shortcomings/defects that we do not necessarily need to replicate to a machine. One of the cognitive biases happen to be us thinking very highly of ourselves. Our brains are far from "perfect" (natural selection and all).

While 2029 may be too early (note that it actually might not be. Nature doesn't exactly follow Moore's law, but so far electronics and computing do), I doubt it will be several hundred years. The other thing to keep in mind are geo-political event which may have adverse effects on scientific development/progress. To expect there to be relative peace in the next 15 years is a far better prediction than hundreds of years. We might even loose some advances by then.

Expecting an AI to replicate/model every detail of a brain is like expecting a house to replicate/model every detail of a cave.

So after reading "On Intelligence" by Jeff Hawkins [1], it would seem that the algorithm of the cortex is mostly figured out and it missing only the hardware capable of simulating enough neurons. One comment I read a while back was that the description in the book was good as far as it goes but there's more to do with regards to making it a two way process. I.e. the part where the nervous system takes action to confirm a prediction (I probably didn't explain that very well).

My question is why would you want to build a human like intelligence? Wouldn't it be enough to build tools that simulate parts (pattern recognition, inference, reasoning etc) of our own intelligence so that me may augment what we're capable of? What purpose would it serve to include emotions feelings and what makes us human?

I'd love to hear from people who actually know more about this stuff than reading a couple of popular science books!

[1] http://en.wikipedia.org/wiki/On_Intelligence

To be honest, it's likely that the why is simply “because we can”. I'm sure there are other advantages—having a fully-functioning brain in a digital realm where you can modify it at will has a myriad of probably psychological and biological research implications. But even if most people asked themselves ethical questions about whether doing that is right etc etc, someone would produce it simply because it is possible.

I hate to say “it's in our nature”, but I think there is something (perhaps each other?) that is constantly pushing us to prove that the boundaries others thought were there aren't. And even if we know intellectually that they aren't, we want to manifest the breakdown of those boundaries.

So why would someone want to build a human-like intelligence? Because they can.

Upvoted for the OI reference, which I found to be a remarkable book. I highly recommend it.

But I wouldn't go so far as to say "it would seem that the algorithm of the cortex is mostly figured out": my takeaway from the book was that a lot of it was solid educated guesses and connecting of dots. E.g., the concept that the brain stores and manipulates patterns and this is the key intelligence behind all the senses, and the multiple "levels" of intelligence and how they connect to each other.

I totally agree with your second para (this is another point that is present throughout in the book). Jeff Hawkins did a good job (for me) to distinguish between intelligence and "human-ness". Recreating the human cortex will still leave us far far away from creating an artificial human brain.

Whenever I see these Kurzweil naysayer articles, I always feel like they are overly pessimistic, or short-sighted. Just because problems seem insurmountable now, doesn't mean that they will be in a few years. People always seem to forget about the exponential growth of information technology. If you look at an insurmountable problem in the context of today's technology and speed of discover, then yeah, it will be tough to believe that we'll have sentient machines any time soon. But please do not forget that tomorrows technology will be exponentially better, faster, stronger than what we have today. In five years, the problems outlined in the article may seem like child's play.

The hardware may be exponentially better, faster, stronger, but you've still got the same people writing the software. I don't buy the whole "Moore's Law will give us strong AI." - it just feels a bit like Asimov-era sci-fi with "nuclear power will fix everything".

To provide a more concrete objection, I'll say that a human-level AI would have to have a general and expandable framework for learning, and I don't think we really have any idea how to build one. Watson is cool and all (and probably the best current counterexample to my point), but it needed a lot of domain experts with doctorates to set it up to "understand" problems in a relatively limited field.

Kurzweil's Discover article says if we build a sufficiently complex machine it will be conscious as fully as humans are conscious. And he says we might do this by 2029.

This blog post says there are all sorts of low-level physical processes in the brain like detailed interactions of neurotransmitters and calcium ion channel dynamics which are super complicated and scary. And it will take hundreds of years to figure it all out and replicate it.

The answer is we don't have to replicate the human brain to create a system complex enough to exhibit consciousness. Modeling is not simulation, we can model a system after the human brain without simulating every last detail.

>> Modeling is not simulation, we can model a system after the human brain without simulating every last detail.

Maybe so, but I think the author is arguing that those details are part of the emergent whole; in other words, without them, you might get an intelligence, but a very alien one.

Thomas waves his hands and says "the brain is really complicated". He doesn't address how much complexity needs to be modeled, but he implies a lot of it does. Kurzweil says a "sufficiently complex" program will exhibit consciousness. So the rub is exactly what are the essential properties of the brain, and what are the accidental ones or the ones that can be implemented some other way.

Is there really a strong argument that human cognition and/or consciousness relies on quantum effects? (That is, stronger than the vacuous argument that everything relies on some properties of a quantum mechanical system, including our computers, by virtue of being made of matter.)

I seem to recall Roger Penrose's argument to that effect a couple of decades ago being soundly criticized.

No, there is not. The linked article is vacuous, and the quantum consciousness has not come up with a testable hypothesis yet that explains any observed data better than simpler models.

This explains it better than I could: http://rationalwiki.org/wiki/Quantum_consciousness

We've always groped for ways to explain things we don't understand, and have always preferred irrational explanations to the uncomfortable feeling of ignorance. I guess it's passe to attribute things we don't understand to God, so now we choose to attribute them to quantum effects.

We're talking about a guy who believes he will be able to literally resurrect dead people by feeding primitive personality data into an AI which in turn would emulate the deceased. So, yes, he's been wrong for a long time. He has always been simplifying issues to the point where they become absurd parodies. This is not new.

Having said that, I'm glad he's around though. AI and transhumanism are need of a skilled public advocate, which he is.

While I instinctively agree with the author's conclusion, I have some questions about his argument.

1) The foundational assumption is that, by "sentient", Kurzweil means "Homo-complete":

    > Because of some of the criteria Kurzweil has set for sentient machines (e.g. that they have emotional systems indistinguishable from those of humans), I like to go ahead and assume that the kind of machine Kurzweil is talking about would have fears, inhibitions, hopes, dreams, beliefs, a sense of aesthetics, understanding (and opinions about) spiritual concepts, a subconscious "mind," and so on.

I'm not sure how Kurzweil feels about that but to me the whole point of creating sentient machines is to achieve human-like intelligence without all hindrances and idiosyncrasies that come with being a human. I'm thinking of a thing like Data (or Spock) from Star Trek. Are these guys not "sentient"? I know they're not real but they don't seem outwardly implausible or self-contradictory.

Furthermore, wouldn't some humans fail to qualify as "sentient" (or, at least, "Homo-complete") based on this definition? What about early homo sapiens? Do we know for sure they had the capacity for schizophrenia or autism? Is it at least conceivable that very early humans lacked the capacity to love in the same way as modern humans? How about future generations? If they find a cure for depression, won't tomorrow's pill-popping, ultra-content humans also be "sentient"?

2) Even if you grant him this assumption, it seems like the author is essentially saying that the only way to be "Homo-complete" is to be a "Homo" (pardon me). That may very well be true and I think the author makes a compelling argument that it is. But I still don't see why being a human is the only way to be "sentient".

So in addition to serifs being worthless now Kurzweil is completely wrong! This guy is going to debunk all the truths we hold dear.

It is cool to learn about dendrites: https://docs.google.com/viewer?a=v&q=cache:FIejhf3q_y4J:...

I see now, more intuitively, how the brain's nonlinear triggering might come about, and how plasticity can be achieved.

Isn't it worth noting, however: that the singularity may not require 100% replication of the human brain, but rather replication of aspects of it? In other words, mental illness should not be viewed as a requirement... it should be avoided, if possible, in such modeling, assuming we're developing these models to help run our infrastructure.

Creating an exact replica of a human brain on a machine seems an unnecessary goal. What we want is a machine that can think like us but orders of magnitude faster/broader. That doesn't mean we want it to think like us in every sense, but rather to accomplish specific goals: like the end of poverty.

"imagine that you are the subject of some mental event--perhaps you are experiencing an intense pain. Now imagine that one of your neurons is replaced by a silicon chip prosthesis that has the exact same input/output profile as the neuron it replaces. At the core of this thought experiment is the presumption that such a replacement would be unnoticeable to you or to anyone observing your behavior. Presumably, you would continue to experience pain even though the physical realization of those mental events includes a silicon chip where an organic neuron used to be. Now imagine that, one by one, the rest of your neurons are swapped for silicon prostheses. Presumably there would be no change in your mental life even though your brain, which was once made of lipid and protein neurons, is now entirely composed of silicon neuronoids." This thought experiment, first presented by Pylyshyn (1980) does not seem too far fetched today.

TL;DR: we are very, very far from understanding the human mind, much less simulating it.

I have always thought that SF accounts of intelligent robots were far too hand-wavey.

I predict we'll never have machines with minds like humans. They will always exceed our abilities in certain tasks for which they've been programmed, but will never pass a long-term Turing test; for example, you'll never mistake a bot for a human remote programming colleague, with whom you chat and collaborate on features every day.

I take it as an axiom that the human mind cannot understand the human mind fully, for the same reason that a box cannot contain another equal sized box. The "singularity" requires a series of boxes, each of which contains a bigger one.

Out of curiosity, why would you take that as an axiom? Do human minds possess some qualities above physical reality? Are they unknowable in principle? If you answer yes, how did you arrive to that answer? Isn't there a physical experiment or demonstration that will change your mind?

I think the box analogy is flawed. We have certainly come a long way in understanding our own brains and bodies, even in the past couple of decades, and you could've said the same thing about boxes when people were first venturing into e.g. neuroscience.

The model doesn't have to be entirely contained in a human mind either, that's why we have a series of tools to aid us in modeling new tools, layer after layer. A human mind cannot contain a complete specification / working model of a modern computer either (if you think of all the layers involved), but modern computers do exist.

That would be philosophical questions.

In addition to that, human minds may possess some qualities above our current understanding of physical reality.

And wouldn't the experiment "proving" that we understand it all, precisely, be the creation of an AI equaling or surpassing the human mind?

"human minds may possess some qualities above our current understanding of physical reality."

While you are free to believe that, you should be aware that you've chosen to leave science and logic behind, and expect to be treated as such. I do not mean that in the sense that you are therefore wrong, or that that is necessarily a bad thing; we must all make choices that leap beyond the light that science can cast, as the light is quite limited in some ways. But nevertheless, you have explicitly tossed anything that can be discussed in concrete terms overboard.

Usually I'm saying this in the context of whether we'll ever have FTL. Yes, you are free to believe that the next supercollider will produce some bizarre result that turns out to be something we can harness for FTL... but meanwhile, in the real world, the evidence gets ever tighter that it is impossible. If you wish to take a flight of fancy, fine, but please be aware that is what you are doing. And do not mistake it choosing to toss rationality overboard for rational argument; you can't have it both ways.

>> A human mind cannot contain a complete specification / working model of a modern computer either (if you think of all the layers involved), but modern computers do exist

This is an excellent counterargument. I still think minds are many of orders of magnitude more complex, but your example shows that my box analogy is flawed.

You don't need to understand something to build it. After all natural selection built our minds with just trial and error and no design at all.

Yes, but it took a couple billion years for natural selection to do it, and it's a Texas sharpshooter fallacy to draw a target around a currently observed humanity that could have turned out much differently. Us hitting that same target in decades rather than millennia is a staggeringly different exercise.

Except that we're rapidly becoming better shots, and we're aiming at a target that's already been drawn.

Exactly. I could build a program to play checkers, and have it beat me at checkers every time I play it.

Great, now build a program to play Go, and have it win every time.

> ... not only computational capabilities but all the things that make the human mind human. .. this requires a developmental growth process starting in "infancy." A Homo-complete machine would not be recognizably Homo sapiens-like if it lacked a childhood, in other words. .. have the potential of becoming depressed, .. compulsivities, .. panic, .. addictions, ...

I don't see all that. I don't see why a strong AI would necessarily be equal in emotions to humans and even develop similar behavior.

I haven't read any books by Kurzweil though, so I cannot tell if he makes that requirement on a strong AI.

Other great impossibilities: sequencing a human genome in 15 years.

Cost of first human genome: $2.7b

Cost 20 years later: $250.

I think 2029 might be pushing it, but neither would I be surprised. Strong AI is always "20 years away". One day it wont be.

That presumption, that strong AI is inevitable, is precisely the point in contention. The fact that it has been a decade away for nearly a half century argues against its inevitability, not for it.

> Other great impossibilities: sequencing a human genome in 15 years.

Citation needed. Who said it would be impossible? And: were they actually taken seriously?

This is spot on, and something I have thought for a long time without the time or ability to elucidate it so clearly.

The complexity of the human brain is not just difficult, not just complicated or hard to understand, it is several orders of magnitude beyond our ability to comprehend. IMHO we will need a fundamental leap in our own intelligence before we have the means to comprehend even how the brain works, and one more leap to understand how we might reproduce its behavior.

It's a long way off. People who believe otherwise are doing just that—believing.

The bits about the brain using quantum computing are only theories. TMK there isn't any evidence of these theories being true or false.

I would hesitate to use the term 'spot on' when discussing something we know so little about.

What we do know is that by 2030, a standard CPU will be doing somewhere in the neighborhood of 10^19 computations per kw/hr. That capability, presumably on a single chip, will rival or exceed the brain's computational ability, perhaps by an order or two of magnitude.

What we do with that horsepower is up to us coders. Maybe we'll code up a brain. Maybe we'll figure out how to sell you soap faster.

Uff, 1st theory =/= hypothesis, 2nd if the brain quantum computation in microtubules is true, 10^19 computations per kw/hr is not going to be nearly enough.

What do you mean by "comprehend"? What do you say to the idea that we might -- rather than build an AI piece by piece and thought by thought, a project I agree is several orders of magnitude beyond our ability to comprehend -- instead "grow" an AI by putting together evolutionary computation and enough processing power to eventually converge on something we would understand as "intelligence"? The extent of such an intelligence wouldn't be fully understood by us, or itself for that matter, but what of it? Our intelligence evolved without anyone understanding it.

Ellen Ullman's "Programming the Post-Human" from the October 2002 issue of Harper's sheds some really interesting light on AI research, how the brain functions & what it means to be a human. Definitely worth a read if you're interested in the subject & available for free on Harper's site:

http://harpers.org/archive/2002/10/programming-the-post-huma...

People really ought to read "On Intelligence" if they want to understand how wrong the entire academic AI field is.

It's not just the approach that is wrong. It is the entire mindset that is wrong.

At the same time, people in this field are so wrong that "machines as intelligent as human" may not be that far off because we really cannot be that smart if we have so many experts so wrong in such a field ; )

But "On Intelligence" is a subset of "the entire academic AI field". And, for any of its variances from the norm with regards to opinion, it's fairly typical of the "academic AI mindset".

Reply on news.ycombinator.com