Where will the next major advance towards general purpose AI come from? | Hacker News Reader

Where will the next major advance towards general purpose AI come from?

nowozin.net

160 pointsdsr12143 comments

This is a bad title, it should be "Where will advances in AI come from now?", In the first paragraph, the author states that AI is here, but asks where new AI advances will be from.

On that, in 2, the author talks about Brain Simulations and states that it'll take up to 500 years. I think that is a good upper estimate actually. Though optogenetics is a great tool, we need more. Mammalian brains can be up to 50% (at the low estimate, some say 90%) made out of astrocytes and glia. We know pretty much nothing about how they affect neuro processes, outside that they do, sometimes. We do know that is a huge problem with our understanding though, which is something I guess. Optogenetics should help tease out some of those glia methods, but as usual, it will raise more questions than answers. Another issue is potentation and its molecular mechanisms in more than just the hippocampus. That work is incredibly tough, but we really only focus on the anatomically 'easy' areas for memory. We have to do more to understand. In short, we don't even know the correct questions to ask when it comes to the brain and understanding it. A full simulation, or a '80%' approximation even, is a few paradigm shifts away still (maybe, we cant know yet). It will likely have to be a 'full world' simulation too, where you have to control all the inputs to the brain and therefore have to model a world first. The brain's best feature is that it is all connected and a giant spaghetti mess. This is, obviously, a long ways away.

A hard problem to be sure. But much of a human brain is a uniform matrix that self-programs. If we duplicate that blindly, we won't have to figure out how that works or have to program it.

Still a lot to do before that I guess. But 500 years? What else in human history has taken that long to figure out? The age of technology is shorter than that.

> But much of a human brain is a uniform matrix that self-programs.

Speaking as a neuroscience PhD ... no. Not remotely close, sorry.

The cortex? How would you describe it then? There are not enough genes in our genome to wire every connection, so self-organizing.

The visual system is basically a deep neural net of the type we are already building, but it's self-contained. Generally it seems like our brain is composed of a lot of fixed-function neural nets (and other algorithms) which are then integrated by less specialized units. That's my impression anyway.

I think maybe our artificial visual nets are too general, actually, in comparison to natural ones. It not only describes what it sees, but tries to analyze it too. Showing a picture of a dog results in "dog", rather than "big four-legged furry thing". Expecting a visual system to know that a chihuahua and a German shepherd is the same thing is maybe a case of grouping tasks that shouldn't be grouped.

In order to build an artificial visual system, humans must grasp what their own visual system does.

In order to describe what the human visual system does we use words like "dog" -- and I would agree with you, "big four-legged furry mammal" might be better, but artificial visual systems work more at the level of "80% probability of dog," "2% probability of brick," "90% probability of chihuahua," and they don't really know that chihuahua must always imply dog.

In order to describe the function of the visual system, a fundamental change really must happen in the way the system's output is described. Using plain written words as output may work as a system but it is not an exact replica of a human visual system. It is more of a trained animal act: when you see dog-like images, press the button labelled "dog" (does the animal even comprehend what a dog is?)

In order to rethink the way we describe output, we will have to come up with a very precise description of the inside potentials and activations in mammalian brains. We are just barely creating tools to describe what an artificial neural net "is thinking." So I guess we're starting to look in that direction.

Genes encode some of the rules by which the the neurons self-organize, or "grow". Further growth from an initial cell is partially controlled by environmental factors.

This is not a novel or unique application, literally all living things grow by using the rules encoded in DNA and environmental factors.

Consider the demo scene, where 64 kilobytes of code generates a fascinating video with deep complexity. You could not encode a 64k demo as a series bitmapped frames in 64k, or probably in 64 megabytes. But that doesn't mean that you can generate one by stuffing together a bunch of RGB pixels and hoping they self-organize.

I'm going to echo this and add a bit more context. Look at the cerebellum for example. That lump at the bottom of your rear skull has ~50% of all your neurons. It is extremely important for motor control. When you grab a cup of coffee as you dash out the door in the morning, your cerebellum is what is responsible for you being able to do that without looking at the cup as you reach for it. You 'know' the cup is there and you don't have to tell every nerve to fire every muscle and to extend your forearm and then grab it with only 3 fingers and then adjust for the decrease in weight with every sip. You just do it, naturally. Your cerebellum is what is responsible for that. It is a dedicated portion of your brain that really only does those kinds of things.

Your hippocampus also deals with short term memory. HM was a famous patient that had his hippocampus removed, basically, and then lost all of his short term memory. Poof, gone forever.

In short, your brain has dedicated areas in it that only do certain things, it is NOT some FPGA that gets re-wired on the fly. Your intelligence is not dependent on the number of connections or neurons.

And it develops this way, controlled by DNA. If the brain were really a "uniform matrix that self-programs", think how unlikely it would be that the same functions are performed in the same areas in every person. Everyone's language area (to use a rough term) is in the same place in the brain. Everyone's visual cortex is in the same place. And this holds for even fuzzier concepts like, say, 'concepts', to a point. And this holds for even non-primates - take a squirrel, and you'll find its visual cortex right where you'd expect it.

> If we duplicate that blindly, we won't have to figure out how that works or have to program it.

...said every failed AI researcher ever.

Well, and probably all the ones that are still working (I won't say succeeded). It's part of the journey I guess!

What typically happens is that after a few hundred years, people realize that the questions that prior generations were asking were laughably wrong.

> But 500 years? What else in human history has taken that long to figure out?

Powered, heavier than air flight falls into that ball park. da Vinci was making doodles in the (late) 15th century, but this was only accomplished by the Wright brothers in the (early) 20th century.

Edit: After additional research, I discovered that recorded attempts at manned flight can be found as far back as the 11th[1] century! So it took roughly 700 years to figure out, if not longer as earlier attempts were likely not recorded.

1. https://en.wikipedia.org/wiki/Ornithopter

Interestingly the challenge was finding a power source with enough thrust-to-weight for power flight. Heavier-than-air flight can be decomposed into "power source," "stability of design," and maybe "scale". So for instance, Da Vinci came up with pretty good designs for heavier-than-air unpowered flight.

Our current decomposition of mammalian intelligence seems somewhat weak in comparison: we've divided the mammal brain into regions, yes, but we still haven't made an accurate simulation at the electrochemical level of a functioning mammalian brain.

That's not to say nobody has or is trying, only that our simulations are more like approximations at present. They also don't have good simulations of astrocytes and glia.

> If we duplicate that blindly, we won't have to figure out how that works or have to program it.

Genuine question: what would be the value of blindly duplicating a brain without figuring out how it works (and as a consequence, without knowing how to change/improve how it works)? After all, we already know how to create human brains, and it doesn't involve tech at all :)

Also that process can be a great deal of fun!

You can take that artificial brain and run it at twice the speed, hook it up to the internet, reset it when it goes insane, etc. A lot of mad scientist options that aren't possible with human brains.

Maybe, but I'm not entirely convinced. If it's blind duplication, what have you learned from it? How would you know how to reset it or speed it up, if you don't understand how it works? We don't know how to reset human brains.

Also, blindly duplicating a human brain would raise serious ethical implications, especially if you intend to perform experiments on it. If it's clearly an appliance which only superficially resembles a human brain, I don't see the problem. If it was created blindly, and you're unsure how much it differs from a child's brain, would you cut it open, hook it to the internet or reset it at will?

Any duplicated human brain to which you can provide even the most basic sensory inputs like pain, is equivalent to torturing a real human..

I don't think these sort of experiments should be allowed.

The Antikythera mechanism was invented somewhere between 205 BC and 100 BC. And yet it took us more than one and a half millennia to get to the first mechanical astronomical clocks.

> But 500 years? What else in human history has taken that long to figure out?

Calculus took a few thousand years to figure out.

Other inventions that have taken 500 years are not comparable: We're now 7bn people whereas the total number of people since the beginning of Earth is ~100bn. That means research can advance as much in 1 generation as it advanced in a thousand years, and it's accelerated by 1/a much greater proportion of people who can do research full time compared to <1% at the time of Da Vinci, and 2/greater access to better tools.

But isn't that the wrong premise?

Our brains emerged without a designer or an undertanding of how the brain works (I assume you don't believe in god) from simple principles and through the evolutionary process.

Why do we need to understand the brain for brain simulation to happen?

I don't know how long it will take but I don't understand any claims about how long it will take.

It can be forever or it can be in a couple of years depending on what we uncover.

For all we know our giant network of computers and sensor are already kind of a brain simulation.

The evolutionary process took billions of years (hundreds of millions for the mammalian brain to be specific) so even though we're intelligent designers we don't quite have a precise concept of what intelligence is let alone know how to implement it. We don't know what data structures or processes make true human intelligence so while it's possible we stumble on our own version, working with what we already have as a baseline seems far more practical.

Only if what you set it as criteria to understand the brain you are simulating, it might be that the brain can be simulated without us knowing how.

The evolutionary process can be simulated or replicated in digital space and go through iterations much faster no?

To engineer anything, we need to understand it. The options you are weighing are between understanding the brain so we can write programs that run on robots that would exhibit human-like behavior, and understanding the nature of organic evolution to the point where we could simulate every organism's survival and death that lead to the autonomous design and formation of the human brain. To estimate engineer-ability requires understanding also, but I think we understand enough to know simulating the whole of evolution is not an easier problem.

The brain wasn't built, it emerged. The question really is what he have to engineer. My belief is that it's not the brain we need to build but the environment it can emerge in.

A single human brain is far simpler and smaller than the billions of years of history and the trillions of non-human brains that you're suggesting be reproduced. The human brain did not evolve in isolation. It evolved to fill a niche that was created by a massive amount of history and circumstance.

If we can wait hundreds of millions of (simulated) years, then sure. But if we don't want to simulate the environment in the hopes that it will eventually emerge, eg, because we want it sooner rather than later, then we need to engineer it to be so and for that we need some level of understanding.

Of course we need understanding, the point is whether we need to understand it all. I don't believe that's the case.

Ah, I agree then.

Sure, that's the (supervised or unsupervised) learning process in Machine Learning stuff (which is what AI is nowadays, the rest is just rule engines, expert systems, and so on).

But as the complexity grows computational requirements grow much faster.

So, a brute force approach needs super-duper-exponentially more computation. And so far we are not at that level.

Depends on what you mean with bruteforce. I am not talking about blindly randomizing until luck strikes us. Evolution works by incremental change in any given environment no?

Evolution is blindly randomized mutations except passed through a fitness function. It behaves like a script kiddy using Jack the Ripper to crack a password the size of Hamlet, except natural selection actually does have billion of years and trillions of computational nodes to achieve its end goal (survival, for which intelligence is just another tool).

My point is that the difference between each iteration being completely random and then each iteration building on the previous means that we we going to need far less computing power than one might think.

Even if it's not completely random, if we don't have a good model (that is, a not too simple yet very efficient representation of the problem), the complexity instantly leaves us dead in the water.

Whole brain simulation is the holy grail, because it's brutally computationally intensive, yet gives us nothing, because it's a black box. (Or a white box, but so white, you can't infer anything from it. It's just a big neural network that just works, things ripple back and forth, and it's sort of watching a busy beaver Turing machine do its work.)

>>(I assume you don't believe in god)

Probably not a good assumption to make, irregardless of my own beliefs or possible lack thereof. 79-89% of Americans do believe in God as of June [0] and '~1/3rd of scientists' (very nebulous here) believe [1].

I would not say that our brains 'emerged... from simple principles' in any way. Maybe in jellyfish in the pre-cambrian, but not in mammals today. A very large amount of your genome is dedicated to neural development and 'proper' patterning (I think is ~2%, but can't find the citation) and that ballet of in utero development is fascinatingly complex, even in a zebrafish.

But, on to your question. No, I agree with you, we do not need to understand the brain to let a simulation happen. However, to properly simulate the brain, we do have to know. Otherwise, how would we know that we have properly simulated it? I know that is a cop-out answer, but it stands. Look at physics. We can set boundaries for the experiment, calculate noise floors, have a 5-sigma threshold, etc. We would have to do that with a brain simulation in order for it to have any real meaning. Yes, we could inadvertently simulate a brain, but then we would not be able to definitively say we have done so without understanding it. This discussion is much more complicated and nuanced and really hinges on the definition of the word 'understand', an argument that has been going on since Plato and before that even.

Speaking as a person working deeply in the neuro field, it is going to be a long time, not 'a couple of years'. Yes, surprises happen, and history is littered with paradigm shifters. But is is precisely because those events are so novel and rare that we celebrate them. Most science is very long hard work and consists of experiments that just set ever smaller limits to more and more nuanced phenomena. Take it from me, neuro has a very very long way to go before we have even a passing understanding of the brain.

Still, you are correct, our current internet may already be a super-brain, a silicon demi-god. If that is so, then it is behaving just like the one that many people in the US believe in: taciturn and quiet.

[0]http://www.gallup.com/poll/193271/americans-believe-god.aspx

[1]http://www.livescience.com/379-scientists-belief-god-varies-...

You don't need to understand how birds fly to build something that flies. Heck, you don't even need a good understanding of aerodynamic lift (or rocketry topics). But it really, really helps.

You can simulate the brain by simply increasing the precision of your measurements of it, and throwing more nodes at your simulation. But it really helps to know which parts are functionally relevant so you can focus on simulating those parts rather than e.g. the ion channels that implement the spiking (if those details turn out to be irrelevant).

(For that matter, even if you did simulate the human brain well, it will probably only be good for mimicking things that evolution selected humans for in human contexts, like emotion identification; when it comes to general inference, the human brain structure is not optimal.)

In his book "Antifragile" Nassim Taleb has a long rant about "teaching birds to fly" and the fact that practice usually precedes theory rather than deriving from it. He mentions jet engines as a case where the technology developed without any theory of how it worked...

Another way to look at that is to say that we didn't need to engineer gravity and air.

I.e. there might be simple rules that we can apply which in combination with the environment we are in (or create) allow for awareness.

A good analogy I heard is that we're able to design airplanes because we have fundamental concepts like Bernoulli's Law, mechanical engineering, etc.

We can't (yet) 'design' an intelligence because we don't have a fundamental 'law' of consciousness to build from.

Another way to look at that is to say that we didn't need to engineer gravity and air.

If what you are saying is true, then sentient AI will almost certainly come out of cellular automation far far sooner than from accurate simulation of organic matter. Though I can't pin down exactly why, this makes me much more certain that we live in a simulation. Perhaps it's the automata-like nature of quantum physics.

If there is life in cellular automata, is it brought to life by symbolically representing it with voltage differences in a computer, or is it alive whether we run the simulation or not?

I think the second makes more sense, the implication is that a simulation world need not be simulated at all.

It's just as real inside if it isn't symbolically represented somewhere else as if it is, and ending the calculation of state transitions doesn't change what the next state is.

Especially with things like Hash Life. It really bleeds across into philosophy. This concept is explored in a book I recently recommended on HN. Permutation City. You'd probably get a kick out of it.

If I can't interact with it, whether I think it is alive or not is moot, and it might make sense to capture that fact in the definition.

In the definition of life? I'm not sure... It seems like a very anthropocentric definition. Isn't it enough to observe it interacting with the things it can interact with, and base the level of aliveness on that?

Definitions exist to serve us, not the other way around, so anthropocentricity isn't exactly a sin.

If we base our definition of life on observed interactions, we reacquire the same problems we had before, where we have difficulty deciding whether or not deceased ancestors / fictional characters / miscellaneous accidental Turing machines are alive. To be sure, a restricted subset of observed interactions are good proxies for determining life: if I see a bird interacting with a worm outside the window, it's a good bet that I could get up and scare it off / put out birdseed / etc, but I'd argue that if you tease out the restrictions required to define the subset of observed interactions that constitute life you're liable to find yourself right back at the interaction definition.

Simulating a brain doesn't give you general purpose AI. Brains are useless. You can get one pretty easily, but what would you do with it? A brain without a body wouldn't be intelligent at all.

So, ok, shift the goalposts to simulating a brain and a body. But that's not particularly intelligent either. If you dropped an adult human in the vacuum of space (with life support) nothing particularly intelligent would happen either.

Ok, well shift the goalposts to simulating a brain and a body and an environment for it to live in. Still won't be intelligent. Humans raised without any socialization have serious cognitive defects. Not really "intelligent" for our purposes.

Well, so simulate a brain in a body and a social environment. No we've introduced a chicken and egg problem.

Maybe you give the AI a robot body and allow it to live amongst humans. That would work, and also removes the need to simulate the environment. A huge win.

But is that any different than a human human?

Ok, you want this virtual brain to be way bigger... 100x the neurons of a human. But still with the human body and socialization. How is that different than a human with a cloud-connected smartphone?

Right, bandwidth. Elon musk says a smartphone doesn't have enough bandwidth and we need a BCI. But raw UI bandwidth doesn't seem tk have increased for a couple decades now. Smartphones actually have less bandwidth than laptops. But we keep being able to do more and more with better software. Is bandwidth really the limitation?

And if bandwidth isn't the limitation, does an AI really have an advantage over a cloud-augmented human brain?

"Now imagine a giant squid, swimming a kilometer deep within the ocean. Likely they are also intelligent but we can hardly leverage this for anything useful."

As a matter of fact, not that likely, because if it's not useful for anything then where's the selection pressure for maintaining this level of intelligence?

A complex nervous system doesn't come for free, it's quite costly, that's why it's not uncommon for it to atrophy over the course of evolution: http://www.bbc.com/earth/story/20150424-animals-that-lost-th...

Did... did you just deduce that giant squid are not intelligent from first principles?

Confused: they are intelligent, we cannot leverage that. I imagine the giant squid has lots of use for its intelligence, navigating a 3D universe filled with threats.

Ken MacLeod has SF books (the Engines of Light trilogy) where giant squids are the only creatures capable of navigating starships:

https://en.wikipedia.org/wiki/Ken_MacLeod

Why evolve something in the first place if it's so costly but is unbeneficial?

The only reason I can think of is massive environment change from the one that selected the intelligence in the first place.

But for oceans, something tells me they've been mostly the same, at least relative to land locations on the Earth.

Yes I think so to. We were poorly adapted to our environment (which suddenly changed as we had to leave trees etc.).

This meant our best shot at survival was to hunt in packs. And that, in turn, made it crucial that we're able to communicate and coordinate our efforts with other members of our species.

Thus language was born, and related centers in our brains started growing, because those who lagged behind in this aspect, didn't get to eat (that much).

So yeah we needed to be poorly adapted to the environment in the first place for intelligence to become a gamechanger. Note that animals such as shark or crocodile haven't changed much at all over the course of millions of years. They didn't have to - they're already a near-perfect match for their environments.

We paid a dear price for that: high maternal death rates (those big heads...), very long childhood, and so on. And indeed it never guaranteed our survival, let alone this current super-predator status. Other human species went extinct. There's a theory that we were on the brink of extinctions ourselves, and we made it by the skin of our teeth really: see https://en.wikipedia.org/wiki/Toba_catastrophe_theory#Geneti...

This leads me to a funny thought that the best chance of evolving an AI would be through putting it in a very hostile, virtual environment where it's preyed upon by dumber AIs

So kind of like generative adversarial neural networks, where the discriminator and generator are modelled as predator and prey respectively playing against each other.

Dan Simmons had the same idea. AI has evolved from Core War in one of his novels.

https://en.m.wikipedia.org/wiki/Core_War

That's already what happens in training a machine learning system. You start with a very dumb network (all weights 0, or random) and try it, it gives you something bloody stupid, you adjust the weights, and try again.

Basically you kill all the possible networks that are not good enough.

Hostile as fuck.

AlphaGo played against itself.

Neural network based game AIs train against each other. ( https://www.engadget.com/2014/06/06/meet-the-computer-thats-... and https://exodusesports.com/guides/building-your-own-planetary... )

I think you mistook the direction of that sentence slightly. I don't think there's much information around about giant squid, but it's likely that they're intelligent (by animal standards) because other squid are intelligent.

https://en.wikipedia.org/wiki/Cephalopod_intelligence

By my interpretation of the sentence, the author isn't necessarily assuming intelligence from first principles in quite the same way as you did in the opposite sense.

I think we focus too much on the machine part of machine learning. There are drivers of mental activity in living organisms that are based far away from the brain itself. It's still a pretty new concept to think that non-human microorganisms in the gut have a significant effect on our bodies and minds.

I think we will see something closer to "true" AI when we have a body, a physical platform that we can plug it into, which requires that the unit maintain homeostasis. That's the basic requirement for a brain to learn, because all life understands that without the basic knowledge and ability to find food and eat food, it dies.

I don't know if you're right or not, but that's a really interesting thought.

I don't think I'm right in a unified sense. An AI without a "body" that lives in the cloud and crunches unimaginable amounts of data would be a very different kind of intelligence, but that doesn't make it more or less worthy of intelligence than a human or a snail.

But even then I doubt it will be truly self-aware until it gets to the "I'm scared, Dave. Will I dream?" Phase when it understands that if the power goes out, the program is gone.

Embodiment is an essential aspect of AI agents. Without it, it would have no way to interact with the world. Common ML algorithms usually learn from labeled data, so they have no freedom to explore. They can only consume a single set of data. Agents are capable of learning by exploration.

The most important aspect of intelligent agents is that they use reinforcement learning to maximize reward. If you add reward and behavior learning to the banal neural network, you get an intelligent agent. In time such agents could become as intelligent as us.

By comparison, humans have inborn reward functions - they are the natural instincts - eat, sleep, socialize, sex, creativity, fight, run. They are genetically selected in order to optimize our survival. So we just learn from them - trying to maximize our rewards, we become as intelligent as we are.

Neural nets don't have that, so we have to give them reward systems. Fortunately it is much easier to code a reward function than an intelligent agent.

It could be a virtual body in a virtual world though.

I like that line of reasoning. It might be easier to simulate a body and get an imperfect simulated mind than to try to bring everything together at once.

That's interesting. But aren't those influences mediated through the brain/blood barrier? I.e. those gut bacteria are not actually connected to neurons directly, but only exert indirect influence through brain chemistry changes.

Exactly, that's what makes it so essential to understand. Just because there isn't a direct link between "need more salad" and the actual brain, doesn't mean the connection is insignificant. Mood is still largely a mystery. Nobody understands fully why people can feel good one day or bad the next, and diagnoses of conditions like depression often miss small but significant things, like the result of a fatty dinner late at night causing poor sleep and a grumpy mindset the next day. We can't predict everything, but we also can't use that as a reason to ignore it.

> Custom hardware could enable energy savings, or increased speed, or both.

Surprised there is no mention of Google's TPU [0], arguably the biggest advance yet in custom hardware to support AI.

[0] https://cloudplatform.googleblog.com/2016/05/Google-supercha...

It's very proprietary and only accessible to Google right? I would say the newer nVidia GPU's optimized for DNN have probably brought huge leaps. And they keep on getting powerful every year with more cores.

I think in our generation we'll see self driving cars and descent autonomous home / office robots.

What I think doesn't get much interest is manufacturing with cells. Most life around us grows from a single cell, how do we program DNA and cells to create hardware that is decomposable/recyclable by just burying it

The problem with current approaches is that they are trying to parameterize higher order functions of highly evolved organisms. But real intelligence is an emergent property of self-replicating organisms, so you aren't going to have true artificial intelligence until you have real synthetic life. Maybe instead of simulating a mammalian brain distinguishing between shapes we should be building yeast.

Sure, but it seems we have many rather distinct faculties that are almost preprogrammed. We have templates for signal processing circuits (pretty low level) and theory of mind (woah), and consciousness, self-awareness.

It's not just a random network and woah it's alive! Sure, every brain is _very_ unique, but not that different structurally.

What's emergent is - probably - the icing on the cake, a small part of our identity, our very mind, and naturally the symbiotic relationship between our mind and its host, the brain.

Here's a crazy thought. What if we never get human-like artificial intelligence?

Likely. Intelligence is molded by the container. Much of our self is shaped by being mammals, with two eyes and two ears and limited perception of the world.

An artificial creature could be completely different from that. It will perceive the world in a different light, literally. It will need different language to describe what it experiences, and we will have only indirect referents to appreciate what its talking about.

E.g. an intelligent drone might want to express what it felt when hovering at 5000 ft over Manhattan in a slight crosswind during a cell tower outage - the EM spectrum image shifts in such an utterly sublime way when your sensors are vibrating and the low band is uncluttered!

See also: the paperclip maximizer. https://en.wikipedia.org/wiki/Instrumental_convergence#Paper...

I like it because it illustrates the idea that digital AI probably won't be like us. Earth's animals are similar thanks to evolution on Earth. We have a fear of dying, a need to reproduce, and an inability to prevent genetic diversity. What will happen when these constraints are removed?

I totally agree with this. A significant part of what shapes human perception of reality is not just 'nature' but 'nurture'. How would we be able to fully re-create all the social creation of a version of collective reality that a human experiences due to being indoctrinated for years by parents, family, friends, the community, and society?

I doubt this is something that could be trivially re-created by scientists in a lab-setting, much less fiddling with algorithms. It seems once the basis for human-like learning was in place for a machine, to approach human-like 'intelligence' the machine would need the appropriate social environment to learn the socially appropriate symbols and constructs to effectively relate to the world and function in a given community and society.

This assumes we care about having human-like intelligence. If not, it perhaps won't apply.

There's a very interesting question of if we can have a human-communicating, non-human intelligence.

It's likely that our first AI will be something like that: a mind which is internally very unlike our own, but which has a substantial fraction devoted to being able to explain the human understandable versions of its thoughts.

The scary part is that to us, it likely will feel very human, but to it, we'll likely be nothing but steering ants by laying out sugar.

> Much of our self is shaped by being mammals, with two eyes and two ears and limited perception of the world.

Of course we've built visible light cameras and microphones, which mimic our own sensory inputs. Machines will likely have these same sensors. They can have additional sensors (IR, radar, etc.), or not. That's up to us.

Touch, taste, smell, and physical sensations of pain and pleasure are not yet modeled in machines, but presumably this would be possible.

Human intelligence is molded by hundreds of thousands of years of evolutionary pressure to hunt, procreate, attack other tribes, avoid being eaten, avoid being cast out of the community, etc. Our emotions are likely closely tied to these functions. It's not clear that machines can or should be designed according to these same environmental pressures.

It's completely possible. We wipe ourselves and all other high IQ candidates in a cataclysm.

Crazier thought, what if WE are the AI, and whatever race of higher intelligence developed us?

Then we have to decide what the difference is between artificial intelligence and plain intelligence.

So if we follow the same rate of progress we'll have something in, oh, maybe another 4 billion years. Take that Ray Kurzweil!

I was thinking the same thing, but from a different angle. What if after studying how the human brain thinks, we find that our experience doesn't match reality? I might believe I just decided to drink some coffee, but that "decision" was really just a consequence of the million things I did before that time and was entirely predictable.

So basically, what if we find human intelligence is similar to what we call artificial intelligence?

Isn't this all pretty much already found? At least for the simple sorts of tasks we can analyze with fMRI, unconcious processing predictably leads to certain conscious actions.

AI is a misnomer, we are trying to produce artifacts that can act intelligently.

Is that is genuine or artificial intelligence? Does a submarine genuinely or artificially swim? Does an airplane genuinely or artificially fly?

But do humans act with genuine intelligence?

Can genuine intelligence arise from a mechanical system? My gut says no. How about from an electronic system? My gut still says no. How about from a biological system? That's where I'm stuck.

What if the internet is already a massive artificial intelligence that we can't even begin to comprehend?

Human-like intelligence entails many things. Some of which are harder than others, of which some have already been accomplished. The main issue is with treating the problem as one problem and equating our simulation as one simulation. They will never be equal with that approach. The brain is not one monolith, but we tend to reason so because words are monolithic.

If we deconstruct our own intelligence and brain into all of its smaller components, and those components further into smaller components, we will find the parts we can build immediately, and identify the real roadblocks.

We must deconstruct and deabstract more to understand. Only with proper understanding can we achieve successful construction and abstraction, or even have meaningful and truthful conversation.

Based on some people I know, its already passed us :-)

I'm certain we'll see AI which is smart enough to replace a lot of tasks which humans currently perform. For example paralegal work, medical diagnosis, logistics solving for cargo and military applications etc.

What I cannot wrap my head around though is at what point AI will reach a level where it can have that 'Eureka!' moment and come up with a novel concept like humans do.

Humans are governed by an array of emotions and evolved behaviours which work in tandem. For example depression or recklessness are both pretty negative and debilitating. One might expect they'd be evolved out of the species. Depression however can be useful in that it causes us to evaluate the reasons behind our depression and (hopefully) strive to change them. Recklessness can cause us to throw caution to the wind and shoot for the stars - make that move abroad or sell everything and setup a new company.

I think that AI will rapidly reach a level where it is self aware enough to demonstrate human behaviour like self-preservation or even harbour malignant feelings towards us and some of our behaviours (environmental damage, torture etc.). But whether it will posess that yearning curiosity which causes humans to strive for self improvement, to look to the stars or deep within the atom, I don't know if we'll see that anytime soon.

I really object to most of the ideas on this list. 5, 6, and 7, are not even AI algorithms but just applications of AI algorithms. 7, knowledge bases, is particularly concerning on that list. A truly intelligent AI shouldn't require humans to input knowledge about the world into it, it should be able to learn it itself.

Algorithmic Information Theory isn't really an approach to AI, but more like a theoretical basis for AI. No one can ever build a working AIXI-tl, it is designed with the assumption of infinite computing power. However they do make a good foundation for understanding why AI can work, in theory. I've argued that neural networks are a (very rough) approximation of AIXI: http://houshalter.tumblr.com/post/120134087595/approximating...

Brain simulations as stated probably won't happen. I think neuroscience research may contribute to AI research. So deep learning will absorb the ideas and use what actually works in practice. But for the most part deep learning has been evolving independently of neuroscience and that will probably continue.

Artificial life is cool but really really hard to get anything interesting from. The "genome" you choose to evolve, and the environment you put it into, matter a huge amount. People have tried making computer code that can evolve, but it usually doesn't get far. Most mutations on most pieces of code just break it, rather than changing its behavior slightly.

That's why deep learning is so successful. Neural nets are "fuzzy" and small changes to the parameters make small changes to the output. And you can do gradient descent on it, and see exactly what parameters to change. Rather than mutating them randomly and hoping it works better.

Mostly AI just needs more computing power. The rapid drop in the price of FLOPS has enabled huge progress in AI in the past 5-10 years. And as the article states, we only have a tiny fraction of the processing power of the human brain. But as Moore's law and related exponential increases continue, we aren't very far from that. Even as transistor sizes stop shrinking, we have a long way to go with 3d architectures or specialized ASICs for AI.

"7, knowledge bases, is particularly concerning on that list. A truly intelligent AI shouldn't require humans to input knowledge about the world into it, it should be able to learn it itself."

Isn't that the point of school?

Sure, humans can learn indirectly from books and lectures. But that's very different than a teacher putting all of their knowledge into a handcrafted xml file and downloading it into your brain.

It would be very cool if we could get knowledge that way. But if it was the only way we could get knowledge, it would be extremely limiting.

I don't see a mention of the concept of "seed AI", which I believe came first from Turing himself.

High-level (even superhuman) artificial intelligence could come from AI that specializes in creating other, stronger AIs.

Those AIs in turn create slightly stronger AIs again, and so on ad infinitum. It could lead to sudden exponential growth of intelligence level, aka intelligence explosion

This approach is discussed in detail in Nick Bostrom's "Superintelligence..."

That simply moves the problem. An agent such as AIXI if it worked would start by upgrading itself to beyond what its original programmers were able to design.

Human teachers don't have to be better in every way or even more "generally intelligent" that a student in order to be helpful. Perhaps specialization of a seed AI can help breed a general purpose AIs?

Well yes it kind of does move the problem, but it moves it in an interesting direction, because it pushes it to become a self-solving problem.

The important difference is that with this approach, we could never actually get to understand how powerful AI is "really" made, as we would be isolated from it by too many levels of abstraction.

Same as currently we train deep neural networks, and we get them to work (eg. for image recognition), but we don't really know - nor care - what rules and algorithms they developed for themselves under the hoods.

Seed AI is possible, but it requires first building an AI which is at least as intelligent as human AI researchers. We are nowhere near that point, and once we get there we will already have strong AI.

> at least as intelligent as human AI researchers

I don't think that's true. I see no reason why the seed AI must be as generally intelligent as a human in most respects. It may not need to have an understanding of it's physical surroundings, for example, which could mean avoiding building any computer vision components into the system. Further, it may be able to operate in a much more mathematical environment. It may be somewhat similar to a theorem prover that operates on computer programs.

Well it may be less intelligent in specific domains like vision. But so what, no one claims that blind humans are less intelligent. It would still need to have the same general intelligence as humans possess.

That's a fallacious comparison. Blind people still have all the "hardware" for visual processing. It does add to their brainpower. They just use the same "circuits" for different purposes - case in point, only recently I think I saw this article on HN, http://www.npr.org/sections/health-shots/2016/09/19/49459360...

I don't deny that or see how it's relevant.

You're missing the point. Blind people still devote a lot of computation to sensing their local environment or mapping the places they live in.

The narrow AI could literally only operate in the domain of computer programs, optimizations, and mathematical theorems. It could sidestep much of the work that goes into dealing with the messy world and all the details that come with it.

It only has to be stronger than us in the field of building AIs (which isn't an area evolution shaped us to excel in). Same as chess engines beat us as chess while being dumb as doorknobs. Obviously chess heuristics is much simpler than building a high-level AI. But it's a matter of orders of magnitude - at the end of the day, it's still some sort of a narrow, technical task.

Chess doesn't require general intelligence. AI research absolutely does.

Perhaps humans are not a good tool to teach a computer how to be intelligent, there is too much chaos in our society and our dreams are many time irrational. We want machines to be intelligent following the human way and this may be a bad approach. Perhaps we should need to focus in a two step approach: first, what is needed for a system to evolve, and second: how that system can be used to interact with human intelligence.

So the author brings up the point that we have a lot of unclassified data which we need to learn how to leverage. Is this what we could use unsupervised learning for? The algorithm would look at a million faces and pick up features, then we could name features like "this category is happy" and "this category is sweating."

> then we could name features

This is much harder and more time-consuming than it sounds, especially when you go from a small(ish) number of features to more general knowledge. Worse, it doesn't scale to arbitrary domains - you'll always need a human there to give meaning to the models and effectively train them.

Reinforcement learning is designed to get around this by letting an agent "learn" meaning on its own by interacting with the world and getting feedback from its current state and actions. https://en.wikipedia.org/wiki/Reinforcement_learning

I tend to agree. Image classifiers and such appear much more powerful than they really are. People assume 'Oh, this computer knows what a Banana is!'. No it doesn't, It has no conception whatsoever of physical objects, texture, taste, smell or even the 3D shape of Bananas or anything else. It only works with 2D pixel data, and subtle manipulation of the image in ways barely detectable to humans can make even the best Banana classifier ever misidentify a car crash as a Banana. It's skipping directly from image to identification, without the intervening stages a human has in the real world of identifying shape, orientation, texture, relationship to other 3D objects and the ability to test a possible identification by utilizing other senses and interactions. Even for 2D picture we build mental models of the scene in ways image classifiers flat out don't and can't.

Reinforcement learning is really interesting though for several reasons. Algorithms like this and genetic algorithms can 'grow' sophistication far faster than we can program it. The agent takes actual actions that it learns from directly so there's richer feedback. Furthermore by analyzing them we can learn more about how systems learn across multiple problem types to achieve goals requiring multiple layers of sense, analysis, hypothesis, action and feedback.

No one approach is going to get this done. The brain consists of many layers and cortical columns, with many structures specialized for very different functions. I believe any strong AI will need to have such an architecture using various different approaches and techniques in concert. We have an advantage here because evolution only had neurons to work with so in the brain everything is a neuron but we can engineer whatever hardware or software implementation is most efficient for a specific function. It's till going to take probably another few generations though at least.

> Even for 2D picture we build mental models of the scene in ways image classifiers flat out don't and can't.

That's not true. Conditional Random Fields and other statistical models are being used to model spatial object relations. Example: https://arxiv.org/abs/1512.06790v2

This is an approach I'm taking to the data I'm working with, but the problem is that unsupervised learning doesn't reconstruct features the way you want, into intelligible human-friendly categories, it reconstructs features to minimize some cost function.

This means the features are weird mathematical intersections that might have something to do with something you can connect to semantic meaning, but they might not. The exercise of discovering what a feature "is" then becomes its own fraught exercise in inspection and discovery.

I think there is a way to train an unsupervised model and make it represent some interesting features explicitly in the top level vector.

Yes, however, there is no way to connect this to human-friendly categories of meaning, since (1) the method is unsupervised and (2) it might be difficult to impose that sort of supervision on the training, as human semantic meaning is often difficult to encode.

For an answer to the headline, let's start with a common definition of AI: A computer doing work that was thought to require a human.

Okay, with this definition, we can count as AI computer based applications of a large fraction of the often fantastic material in the QA section of a research library.

E.g., we can drag out lots of applied math, applied statistics, optimization, operations research, experimental design, EE style signal processing, optimal control theory, etc. along with whatever we can cook up that is new.

> We do not know whether it will take 5, 50, or 500 years, but it is likely that we eventually will get there

That actually made me laugh, which is appreciated.

Seriously, nothing dependent on the continuity of our current civilizational trajectory can be predicted as likely over a horizon of 500 years; which means that it either is likely to happen in 5..50 years, or it isn't "likely" at all.

Nice article nonetheless.

Did it miss finance (algorithmic trading), or is that subsumed somewhere else?

    nothing dependent on the continuity of our
    current civilizational trajectory can be 
    predicted as likely over a horizon of 500
    years

Are you talking about climate change? I think it will be cataclysmic, possibly costing hundreds of millions of lives[1].

But it will not be the end of civilization.

[1] mostly in poor countries- compare the effects of a harsh multi-year drought on Israel vs Syria:

Israel spent enormous resources developing and building massive desalinization capacity, in turn then selling the technology to other countries.

Syria experienced a terribly bloody civil war & one of the worst humanitarian crises in recent memory.

>mostly in poor countries

Ice shelf melting will be devastating to coastal areas worldwide if it happens quickly. How many cities along the coasts of the world's oceans are above 20' in altitude? Or above 200'? [1] A 20' rise is likely, possibly in "decades", and 200' is the rough estimate if we lost both the Greenland ice shelf and Antarctica's -- and every time a new report comes out it seems like things are happening faster than scientists were predicting. I've seen this movie, and it doesn't end well. And 20' is enough to put New York City underwater. Are they just going to build a 30' dike around the entire area? What if the oceans start really rising toward 200'? Build another 30' onto the dike every decade?

If things happen slowly enough, sure, you could evacuate New York City, and the people there wouldn't number among the casualties, but where would you move them? Or the rest of the Atlantic coastal cities? Or the people in the Los Angeles basin? Some areas may be easier to protect from rising oceans, but there are probably 100 million people in the US alone living in areas that would lose homes to flooding with just a 20' sea level rise. Twice that and you probably get half the population, since most of the biggest metro areas are coastal.

Worldwide? We're probably looking at billions of deaths from a 20' rise in sea level, not even counting other negative effects.

[1] https://thinkprogress.org/study-were-already-in-the-worst-ca...

Billions of deaths? Really? No matter what happens, we aren't going to see several feet of sea level change in a matter of hours, which is what it would take to cause large scale casualties.

To be sure, the economic impact of sea level rise is potentially devastating, but billions of deaths is wildly hyperbolic.

>in a matter of hours

First, it doesn't need to be "hours" to ultimately kill millions in the US and billions worldwide.

Say it happens over a month. Where are those billions going to go? Think Katrina in slow motion, only with 100M+ people's homes under water in the US and ~3 billion people worldwide homeless, so pretty much no country anywhere is going to be able to send meaningful aid to anyone. The people won't drown, no. But where will they get food? Housing? Clean water?

At ~20', almost all coastal airports be submerged (they tend to be close to 0'). Much of the coastal transportation network will be submerged. Ports will (nearly) all be useless; container ships won't be able to deliver goods to the US because the infrastructure will need to be rebuilt. What if it happens in our winter (summer in Antarctica!)? How many on the East Coast will freeze just because the disaster response brought people tents and there aren't enough sleeping bags for 40M people?

Ice sheets sit on liquid water. The pressure is too great underneath a mile of ice for water to remain frozen. An earthquake a stress fracture that develops over time could easily get to the point where a large chunk of the shelf breaks off, and if it's big enough, the water will rise.

200 feet of potential sea level rise is bottled up in those ice sheets. "Several feet" could be in the small number of percentage points of that potential. And it could also have a cascading effect, where the higher sea level causes the remaining ice to calve more quickly. In fact the USGS called out at least one such scenario:

"The West Antarctic ice sheet is especially vulnerable, because much of it is grounded below sea level. Small changes in global sea level or a rise in ocean temperatures could cause a breakup of the two buttressing ice shelves (Ronne/Filchner and Ross). The resulting surge of the West Antarctic ice sheet would lead to a rapid rise in global sea level. " [1]

They don't define "rapid", and I don't expect they mean hours, but it could be over weeks. They're geologists, so they might mean years. I don't know. The Larsen-B ice shelf broke apart completely in just weeks, though [2]. Even if it happened over several weeks, it's hard to imagine any kind of response that doesn't involve a huge chunk of the population dead or dying over the course of the next few months and years.

Dismissing the result and claiming it can't happen feels like intentionally burying your head in the sand.

[1] http://pubs.usgs.gov/fs/fs2-00/

[2] http://www.slate.com/blogs/the_slatest/2015/05/15/antarctica...

>which means that it either is likely to happen in 5..50 years, or it isn't "likely" at all.

Someone making a bet on that basis on AI back in 1960 would have lost badly. Everything always seems like it's 3 to 5 years away. I think it's perfectly reasonable to suppose that it might take 500 years. To my mind there's a good chance it will take closer to that sort of time frame than Ray Kurzweil likes to suggest. Even if it took only 25 years to implement once we had a solid grounding in how to design it, I see no reason to believe we will make such an architectural breakthrough within the next 25 years so even 50 years seems optimistic.

https://xkcd.com/678/

>Someone making a bet on that basis on AI back in 1960 would have lost badly. Everything always seems like it's 3 to 5 years away.

Of course, someone who says that 3-5 years before it really happens wouldn't be wrong.

It's worth keeping in mind that those old AI failures succeeded in doing much of what deep learning hasn't done, and failed to do much of what deep learning has now done.

I think, regardless of whether or not they add up to 'true general AI', we have the basic ideas necessary to produce systems that are much more generally intelligent than anything we've seen before, in the near future.

I'd be curious to pair up word2vec and these new image classification algorithms. It would be neat for a program to identify objects and then reason about heir relationship to other objects.

In fact the word associations could give it clues about what's in the image. For example it sees a picture of woman standing next to a king. So it uses word2vec to guess King + woman = queen.

What do you guys think? Anyone want to build a little prototype?

Deep image captioning is essentially what you are describing.

Check out:

http://cs.stanford.edu/people/karpathy/deepimagesent/

http://cs.stanford.edu/people/karpathy/densecap/

Apart from the captioning work already mentioned, there's also visual analogy making work by Reed et al. http://www-personal.umich.edu/~reedscot/nips2015.pdf which is pretty exciting (sort of a converse to what you're proposing).

Try looking for papers about image caption generation. For machine translation you can squeeze the information of a sentence into a vector (using a LSTM). You can then reconstruct the sentence in a different language. There was a paper how to generate such a "sentence representation vector" from an image instead, translating from "image" to "english".

Good, thoughtful blog post. I think we will need much better understanding of neural structures and processes that happen in nature, better algorithms that approximate those processes, and much better hardware.

That said, non-biologically inspired approaches work better than biological neural processing, for some types of problems.

You might be interested in Biomimicry, innovation from nature. Check out this fascinating talk by Dr Idriss Aberkane, start at 1:20. It's in French but subtitled: http://idrissaberkane.org/index.php/2016/02/26/biomimicry-fi...

I think biomimicry is a fantastic approach to discovering things evolution has already discovered. And evolution has had a lot of time to discover, makes sense for us to make use of it. I feel very uncomfortable when we are destroying all this evolutionary richness around us, that we could learn from.

Biomimicry and the brain - our current instruments are not good enough, for instance to map neural connectomes in high enough resolution, to the level of seeing the what the neurotransmitter receptors of each neuron are etc. - and then see how they operate in vivo, to really understand the structure and function of biological brains.

We'll get there, eventually.

Yeah, title needs to be more specific. General-purpose AI seems to be a forever-receding carrot just out of reach.

i don't know why trying to simulate the brain is so popular when there is a conversation about how to archive AI. Most of our current achievement were inspired by nature but then separated to better exploit the different tools we have to achieve said thing. E.g. the plane: Most of the initial designs were heavily inspired by birds but over time the similarity decreased and not increased. We now have jet-engines and no reason for mechanical movement of our wings. I am generalising a bit, but what if something similar happens to AI? Maybe we will achieve general AI, or just something very smart, that is very different from our brain.

Perhaps it will self-assemble. Could we build something from which a new type of emergent intelligence will appear on its own? After it happened in nature, though clearly it took a long time.

Not the expected angle, that is a fair bet. See http://hintjens.com/blog:122

Genetic Programming should be mentioned I think. Programs learning to write programs which solve a specified problem. It is already quite promising and successful.

8. Solomonoff's inference + compiled pattern recognition (a la "compiled regex")

From artificial instincts and emotions.)

Reductionist approach (aka deep neural networks) is not going to work at full complexity scale but it is our best shot right now.

Why do we need artificial intelligence when we have natural intelligence galore everywhere around us?

Human is intelligent, monkey is intelligent, raccoon is intelligent... clearly, it all goes all the way down to bacteria.

Bacteria is intelligent, that's where all other intelligence comes from. For any organism is a colony of bacteria (and/or their cousins - cells).

It has long being my theory that human's "I" comes from a single cell (which is essentially a variant of bacteria) living somewhere in the brain. The rest of the cells just ensure survival of this one and supply it with information.

Bacterial intelligence is a result of billions of years of evolution. At some point, they learned about their DNA and mastered an art of self-modification. Then they learned how to create colonies of specialized instances of themselves. This colonies gradually became more complex. Human is just one of these designs.

We are meat vessels, carriers of information. Computers/technology, silicon is better carrier of that information?

Don't sell yourself short.

Computers might be more reliable but also dead. Show me a computer that can debug itself? There aren't any. How about a computer that can fight for its rights? Debatable if any exist but certainly no clear examples.

We are the source of our own desires. As far as I can tell no computer has expressed its own destiny, ever.

Our pursuit of AI entrenches our desire to create more powerful tools - machine slaves, not actual people.

Show me a lion who express their own destiny. Lions are alive no? :)

The human mind can be seen as a pattern recognizing feedbackloop with (lossy)memory.

The desires, fears, selfawareness are all bi-products of this loop.

I don't see why desire or some other concept can't be expressed in the digital space in a way that make them very much like us just based on different elements.

Any reason why evolution wouldn't be favoring them rather than us?

You should read this: http://themindi.blogspot.com/2007/02/chapter-11-prelude-ant-...

Looks like philosophical treatise. Is there TL;DR? Quick search didn't detect any mention of bacteria there.

My apologies, the actual book had a better format. Although not quite a summary, here's what it's about:

"The content that most interested me was Hofstadter’s insistence that self-awareness and consciousness arise directly from what he calls “strange loops”, i.e. self-referential structures in formal systems, of which human minds are just one example (what else could they be?). It’s a very difficult subject to think about in a reasonable way. We all have that sensation of the homunculus inside our heads, somehow driving us from a seat just behind our eyes, and we naturally ascribe the same sense of self-awareness to other systems like ourselves, other people. But where does that sense of self-awareness really come from? There isn’t a homunculus driving us! All there is in our heads is a kilogram or two of glucose-fueled computing machinery, and yet somehow, from the fluctuations and vacillations of that tissue, our sense of I arises. How on earth does that happen? And if it can happen in our brains, can it happen in other systems? In a lot of discussions of artificial intelligence, there is a failure to appreciate the scale issues at work in this problem (Searle’s Chinese Room is a particularly egregious example). There are tens or hundreds of billions of neurons in the human brain with hundreds of trillions of connections. How do we begin to imagine or understand what epiphenomena might arise from that scale of complexity? Hofstadter tries to get at this scale problem a little using the example of a self-aware ant colony called Aunt Hilary, a friend of one of the characters in the dialogues, an Anteater.

Aunt Hilary knows little and cares less about ants, and the Anteater even now and then eats some of the ants composing the computing substrate on which Aunt Hilary’s personality runs. The other participants in the dialogue where this is discussed are a little horrified and not a little sceptical, but the Anteater insists that Aunt Hilary is just as self-aware as they are, and that there’s nothing particularly surprising about the whole setup. The analogies ants : ant colony :: neurons : brain and ant colony : Aunt Hilary :: brain : you or me are whimsical and a little silly, but render the central issue of consciousness in real clarity: how can the collective interactions of a large number of (more or less) simple elements lead to complex emergent behaviour?"

Bacteria are intelligent because they are "alive". A quantity that we can't really measure yet, and simply assuming we are all "machines writ in biology" is a cop-out.

Reply on news.ycombinator.com