Geoffrey Hinton and Demis Hassabis: AGI is nowhere close to being a reality

https://venturebeat.com/2018/12/17/geoffrey-hinton-and-demis-hassabis-agi-is-nowhere-close-to-being-a-reality/

by z0a • 8 years ago

204 171 8 years ago

venturebeat.com

ludicast• 8 years ago

Isn't that exactly what an agi would say once it takes over the brains of leading scientists.

MR4D• 8 years ago

No - it would post what you just said on HN instead. :)

Seriously - that's a wicked funny post you had there!

mlthoughts2018• 8 years ago

Nice try, Mr. AI but you won’t escape detection by pretending to be a jokester who grew up in Boston.

toasterlovin• 8 years ago

Not to mention that we don't even know if general intelligence exists. All we know is that mental abilities tend to correlate, but not why they tend to correlate. And if you think about designing machines, in general, the idea of general intelligence is utterly ridiculous. Does a fast car have general speediness? Of course not, it has dozens or hundreds of discrete optimizations that all contribute in some degree to the car being faster.

mac01021• 8 years ago

I'm not sure you and the OP mean the same thing by "General Intelligence".

It seems clear that autonomous systems which can apply their computational machinery to a diverse range of problems, and can, in a diverse range of settings, formulate instrumental goals as part of a plan to attain a final goal, do exist.

Because that's what humans are, at least some of the time.

mannykannot2• 8 years ago

But if human performance in these regards never exceeded what the pinnacle of today's AI performance is, we would not regard them as intelligent in a general sense, either.

darkmighty2• 8 years ago

Well, we have general purpose processors. You can prove they can run any algorithm you want (i.e. are Turing complete), but also, for practical problems (i.e. the ones encountered in engineering solutions in our planet and in our universe), they give reasonable max-min performance. Analogously I don't think 'AGI' is entirely useless -- you'd expect an AGI to have some properties like being able to solve reasonably well problems found in nature and society, maybe have a motivational framework distinguishing it as a separate entity, some knowledge about the world, etc.

edit: In terms of Turing-completeness analogues, the best candidate for AGI I think would be simply brute force capability: can this agent try all possible solutions until it solves this problem? (obviously using a heuristic to prioritize) -- that is, it'd employ a form of Universal Search[1] (aka Levin Search). Humans don't necessarily pass this test rigorously because we'd always get bored with a problem and because we have finite memory. But then CPUs are not truly Turing complete either (it's "just" a good model).

[1] http://www.scholarpedia.org/article/Universal_search

yters4• 8 years ago

As I always ask regarding this sort of story, why do we believe human intelligence is computable? The only answer I've heard is the materialist presupposition and sneers at any other metaphysic as "magic," which is not exactly a valid form of argument.

As an alternative, the human mind could be some sort of halting oracle. That's a well defined entity in computer science which cannot be reduced to Turing computation, thus cannot be any sort of AI, since we cannot create any form of computation more powerful than a Turing machine. How have we ruled out that possibility? As far as I can tell, we have not ruled it out, nor even tried.

siekmanj• 8 years ago

We know the brain is doing something - if you don't want to call it computation, then you might as well call it magic.

eej71• 8 years ago

Seems like a false alternative between computation and magic.

yters4• 8 years ago

There are other possibilities. For example, there can be an immaterial mind that operates as a halting oracle and interfaces with the world through the brain. Halting oracles are well defined, and we can empirically test for their existence. So, no reason why we have to assume everything humans do is reducible to some sort of automata. The only reason we make the assumption is because of prior materialistic commitments.

UPDATE: I've been rate limited for some reason, so here is my response whether the mind intuitively seems to be a halting oracle.

1. It's obvious there are an infinite number of integers, because whatever number I think of I can add one to it. A Turing machine has to be given the axiom of infinity to make this kind of inference, it cannot derive it in any way. This intuitively looks like an example of the halting oracle at work in my mind. Or, an even more basic practical example: if I do something and it doesn't work, I try something else. Unlike the game AIs that repeatedly try to walk through walls.

2. We programmers write halting programs with great regularity. So, it seems like we are decent at solving the halting problem. Also, note that it is not necessary to solve every problem in order to be an uncomputable halting oracle. All that is necessary is being capable of solving an uncomputable subset of the halting problems. So, the fact that we cannot solve some problems does not imply we are not halting oracles.

colordrops• 8 years ago

Are you positing that the only alternatives are computation and magic?

hackernudes• 8 years ago

This line of reasoning can be applied to almost any fundamental scientific discovery before it was made.

Why do we believe man can make fire? Well, dammit, we WANT to make fire. Let's figure out how to do it!

Finally, if we were able to explain the brain well with "metaphysics" it would then be just "physics". It seems that all you are saying here is that there is a mechanism that is not yet understood and it may be fundamentally different than other things we have studied so far (which seems unlikely, I might add).

yters4• 8 years ago

Well, if we forced our presuppositions on our experiments we would not have discovered gravitation, electricity, relativity, quantum mechanics, etc. Each of those advances is a metaphysical paradigm shift over previous views of reality. E.g. the original materialism was the billiard ball model with a random swerve, due to Epicurus. No one believes in the original materialism anymore, since it has been falsified in so many ways. Instead, we have redefined materialism to include a whole host of causal forces that were previously ignored or unknown.

Similarly we can mathematically and empirically differentiate between halting oracles and Turing machines, so why not leave both possibilities open as scientific explanations, instead of doubling down on the Turing machine model? Call halting oracles materialistic if it makes you feel better.

mortivore2• 8 years ago

Part of me almost hopes it is a halting oracle of some sort because then we could start looking into either hooking up multiple brains to a single oracle or a single brain to many oracles.

hacker_9• 8 years ago

Behind every successful neural network is a human brain. Neural networks are a tool, an advanced tool for sure, but still just a tool. If we are looking for AGI, and assuming the brain is an AGI, then there are still many differences to resolve. For example, back propagation has not been observed in nature. Nor has gradient descent. So the core mechanisms for learning in nature have still to reveal their secrets.

codekilla2• 8 years ago

Behind every successful neural network is a human brain.

I've spent a lot of time trying to explain this to people, that there is a confluence between the human brain and the machine, people tend to look at the machine separately, which is a mistake. When I say unequivocally, 'there is no such thing as machine intelligence', I just get blank stares.

WhompingWindows• 8 years ago

Arguably, there are successful brains behind every successful brain, too. Every great innovator and thinker was building off the backs of numerous other thinkers and teachers in their life. Should we be surprised that it's much easier for a tool+human(s) to do better than a tool alone, given we also expect a single human + human(s) as colleagues to do much better? Never mind the whole learning/development process, during which 22+ years of dedicated effort by adults to shape/craft a functional human worker.

Overall, I'd agree that really powerful tools for specific tasks is going to be the majority of "AI" in the coming years.

joe_the_user7• 8 years ago

Behind every successful brain is a little strand of DNA and some environmental inputs. Somehow a brain might be more than the DNA however.

chopin• 8 years ago

That's called an emergent property.

atschantz• 8 years ago

I mean it's difficult to 'observe' gradient descent, there are no characteristic properties that you can identify without specifying the relative objective function. But most of the process theories from computational neuroscience are based on some form of gradient descent. Even if it's only implicit, you'll be able to describe the variables of the system as moving against the gradient of some function.

But yes, it's extremely unlikely that nature implements backpropagation directly, as it relies on non-local gradients.

buboard• 8 years ago

Behind every brain is a successful neural network. Or at least that's the promise of connectionism.

9955336• 8 years ago

Your reasoning does not follow. To see why, take something humans already clearly created: Flight. Kerosene-type jet fuel propulsion has not been observed in nature. It is flight nonetheless.

Human flight is not as agile or energy-effective as a dragonfly, but it is faster and stronger. Just like artificial learning may not be as sample-effecient as the human brain. It is a learning intelligence nonetheless and we are already working with the core mechanisms of reasoning and deduction.

mrdoops• 8 years ago

I wonder to what extent the data being fed to these models are the issue. Or rather the problem is the systems that generate these data-sets and how representative of reality they are. If we make an app that involves humans and that data is used in a model - to what extent does user experience and other factors warp reality?

Maybe our existing methods are good enough given enough compute to reach AGI but our datasets are too low fidelity and non-representative of the problem space to reach desired results?

MAXPOOL• 8 years ago

The problem is not the data. The problem is the need for high quality data. Current ML is data driven statistical learning. ML tries to learn a model that describes the distribution. It's impossible to get similar performance as the best reference implementation (human brain) using this approach. https://i.redd.it/kvvgv6zzhtp11.png

Think of 16 year old human:

* it has received less than 400 million wakeful seconds of data + 100 millions seconds of sleep,

* it has made only few million high level cognitive decisions where feedback is important and delay is tens of second or several minutes (say few thousand per day). From just few million samples it has learned to behave in the society like a human and do human things.

* Assuming 50 ms learning rate at average, at the lowest level there is at most 10 billion iterations per neuron (Short-term synaptic plasticity acts on a timescale of tens of milliseconds to a few minutes.)

Humans generate very detailed model of their environment with very little data and even less feedback. They can learn complex concept from one example. For example you need only one example of pickpocket to understand the whole concept.

machiaweliczny2• 8 years ago

This pickpocket example seems like symbolic? relations reasoning.

I think we need simulation of other agents outputs as primary tool for reasoning. That seems to be how intelligence emerged in evolution.

Something like this: choose desired action > simulate other agents outputs based on future state after performing action > check reward for this action after simulating outputs of others > perform action or not > update all agents models and relations in "world" graph model

I think world could be modeled as simple graph and each agent as NN.

Then based on graph we could conduct symbolic reasoning and very fast learning (by updating edges)

I think these models need also need good physical simulator and good understanding of competitivness.

Is anyone aware of such trials of building AGI as I described?

Humans have natural language as big competetive adventage (easy way to compress parts of world graph and pass it to others - ambiguous. I think with aftificial machiness can be done more efficient). Another advantage is knowledge storage - also easy to do with machiness.

If we can build insect AI building human AI should be easy.

lemoncucumber• 8 years ago

As someone who files taxes every year, I'm quite certain that Adjusted Gross Income is a reality ;)

karmasimida3• 8 years ago

And no one should be surprised by this. The NN advancement of late doesn't help addressing human-style symbolic reasoning at all. All we have is a much more powerful function approximator with a drastic increased capacity (very deep networks with billions of parameters) and scalable training scheme (SGD and its variants).

Such architecture works great for differentiable data, such's images/audios, but the improvement on natural language tasks are only incremental.

I was thinking maybe DeepMind's RL+DL is the way leads to AGI, since it does offer an elegant and complete framework. But seems like even DeepMind had trouble to get it working to more realistic scenarios, so maybe our modelling of intelligence is still hopelessly romantic.

odyssey7• 8 years ago

Are there ways that an AI practitioner would be able to tell whether a neural network is doing human-style symbolic reasoning?

karmasimida3• 8 years ago

Good question. I don't think we do actually.

The only reason I am convinced it is NOT doing a good job, is how utterly difficult to apply NN to dialog generation/management domain of business, often time it behaves much worse than rule-based systems.

bsaul• 8 years ago

As mentioned in the original article, being able to reuse part of a network trained on one task, on another different task that shares a subset of concepts, would indicate something like the understanding of a concept has emerged.

darkmighty2• 8 years ago

You'd be able to tell if every brain structure was replicated by a NN analogue (and we understood them sufficiently well). Otherwise you can only use behavioral replication (i.e. Turing tests) to infer it.

antidesitter• 8 years ago

IQ tests [1] and one-shot learning come to mind.

[1] https://arxiv.org/abs/1807.04225

skybrian• 8 years ago

Hmm, it seems like natural language translation has been getting quite a bit better with statistical techniques, though? I guess it depends what you mean by "only incremental".

harperlee• 8 years ago

Statistically, but it is a shallow translation, with no modelling of what is said. This works astonishly well for translation, but gives the false intuition that it is meaningful; whereas it is in fact orthogonal to advancing in systems that comprehend a text so as to be able to reason their way out of winograd problems.

tanilama• 8 years ago

I still remember sometime back Google's GNMT translate the Chinese text of 'I don't want to go to work' into English 'I want to go to work'. That example alone should be sufficient to showcase how most advanced machine learning model can fail at the simplest task.

It didn't understand the source material, it is just very good at memorizing and faking.

m0zg• 8 years ago

It's been "statistical" since about mid-00's. What's new is that it's now neural.

joe_the_user7• 8 years ago

Whenever I attempt to use a translate site to translate more than a paragraph (Facebook or Google), it comes out a garbled mess - doesn't mean some sentences are seemingly clear and meaningful. The big thing is it chokes on idioms, not understanding, not leaving them as is but guessing some clearly wrong meaning. I occasionally find single-sentence posts on Facebook apparently translated astonishingly well, in the sense of being literate English and seeming to reflect the original meaning. But my French or Spanish is quite rough so translate could have missed something big - as I know it does when you get into longer texts.

sdenton4• 8 years ago

Maaaaybe. I tend to think that symbolic reasoning is a learning tool, rather than a goalpost for general intelligence. For example, we use symbolic reasoning quite extensively when learning to read a new language, but once fluent can rely on something closer to raw processing - no more reading and sounding out character sequences. Similarly with chess - eventually we have good mnemonics for what make good plays, and can play blitz reasonably well.

And - let's be real - a lot of human symbolic reasoning actually happens outside of the brain, on paper or computer screens. We painstakingly learn relatively simple transformations and feedback loops for manipulating this external memory, and then bootstrap it into short-term reaction via lots of practice.

I tend to think that the problems are: a) Tightly defined / domain-specific loss functions. If all I ever do is ask you to identify pictures of bananas, you'll never get around to writing the great american novel. And we don't know how to train the kinds of adaptive or free form loss functions that would get us away from these domain-specific losses.

b) Similarly, I have a soft-spot for the view that a mind is only as good as its set of inputs. We currently mostly build models that are only receptive (image, sound) or generative. Reinforcement learning is getting progress on feedback loops, but I have the sense that there's still a long way to go.

c) I have the feeling that there's still a long way to go in understanding how to deal with time...

d) As great as LSTMs are, there still seems to be some shortcoming in how to incorporate memory into networks. LSTMs seem to give a decent approximation of short-term memory, but still seems far from great. This might be the key to symbolic reasoning, though.

Writing all that down, I gotta say I agree fundamentally with the DeepMind research priorities on reinforcement learning and multi-modal models.

joe_the_user7• 8 years ago

Once someone is fluent in a language, the logical operations and judgements involved stop being overt and highly visible to the conscious mind. But that doesn't mean that one stops getting the benefits and results of logical operations.

What you might see as logical operations "not mattering", I would see as logical operations integrated so deeply into reflexive operations that it's hard to see where one ends and the other begins. The contrast is that humans can do pattern recognition in a neural net fashion, taking something like the multidimensional average of a set of things. But a human can also receive a language-level input that some characteristic is or isn't important for recognizing a given thing and incorporate that input into their broad-average concepts. That kind of thing can't be done by deep learning currently - well, not a non-kludgey sort of way.

Similarly, I have a soft-spot for the view that a mind is only as good as its set of inputs.

It depends on how you want to mean that. A human can take inputs on one thing and apply them seamlessly to another thing. Neural nets tend to be very dependent on the task-focused content fed them.

meroes• 8 years ago

As a layman, is this just saying we learn by training an intuition of what's what/what's correct, rather than actually calculating deep reality/referencing our entire memory set every time we intake some information or need to solve a task problem? Meaning, we develop tons of rules/heuristics after repeated pattern exposures, and use the simplified rules rather than a deep theory or 'brute-forcing' every possibility until we find one that's right. For example with chess, we don't know at the deepest possible level why this move might be best, it just feels right due to a massive learned intuition.

If that's the case, then to me it seems like AGI is limited by the amount and type of data a NN can be fed. To have an intelligence like homo sapiens, wouldnt you expect that no matter the underlying NN, it has to take in a comparable amount of data to what the 5+ human senses take in over lifetime, plus the actual internal 'learning' (i.e pattern recognition, heuristics, and intuition) + some kind of meta awareness (consciousness) to speed up and aid this process + dedicated pieces of the brain such as Broca's/Wernicke's

karmasimida3• 8 years ago

I put symbolic reasoning at the spotlight for it is something that NN is particularly bad at: discrete data, hard to design, often approximate and non-differentiable measurement.

The problem is so inherently hard that we are struggling even to come up with a meaningful task, telling us how bad we are doing. That comes to your first point, I think finding the right loss function is a like a chicken-and-egg situation here. When you have the loss function at hand, you already what task and problem you are going to solve, then it becomes easier. But that is apparently not our current situation.

That is why I think DeepMind has a good reason to go after reinforcement learning, after all, that is how we human are trained, through exams and the feedbacks.

As to your point about LSTM, I am not very passionate to qualitatively claim it whether it can/can't handle short/long term memory. That is apparently task dependent, and all the concepts involved are ill-defined.

mindgam36• 8 years ago

"Similarly with chess - eventually we have good mnemonics for what make good plays, and can play blitz reasonably well."

"let's be real - a lot of human symbolic reasoning actually happens outside of the brain"

I was a chess master at age 10. Let's be real - when I play blitz and bullet chess, I am performing multi-level symbolic reasoning at multiple frames per second. In my brain.

I am not an alien. I can do these kinds of symbolic calculations faster than 99.6% of the population mainly because I learned chess as a kid, making it a "native language", and I got good at it early so I spent much of my youth training my neurons with this perceptual task.

My point is not to claim I'm a genius. There are dozens of players who can school me in bullet the way I can school most people.

My point is that human beings DO symbolic reasoning, it is the core of our intelligence. Being able to take in different kinds of input, organize some of them into relevant higher level clusters, sort the clusters by priority, make a plan to deal with the highest prio clusters, act, rinse and repeat.

Humans simply do not have the computational ability to make decisions based on raw perceptual data in real time. Our brains are designed to act on higher levels of symbolic meaning, and we have perceptual layers to help us turn reality into manageable chunks.

In cognitive psychology this is referred to, not surprisingly, as "chunking": https://en.wikipedia.org/wiki/Chunking_(psychology)

Until DeepMind starts working on anything resembling chunking, I believe they are wasting their time and money.

kahoon2• 8 years ago

I don't understand this fixation on symbolic reasoning. Do any other animals practice this? If the answer is no, then it is probably not the most important milestone to AGI or at least not the one we should be currently aiming for. Right now we can not replicate the cognition of a mouse. Feels like we want to go to Mars before figuring out how to build a rocket.

evrydayhustling• 8 years ago

Seconded. Even if animals do symbolic reasoning, they do it on top of hardware based on continuous physical dynamics, more similar to DNNs... So why not build on that platform?

I don't think biological precedent is the only or even most valuable heuristic for deciding where to research intelligence... But I don't see where there is evidence that symbolic reasoning is either necessary or sufficient for AGI, except people describing how they think their brain works.

Related, there are a lot of statements that symbolic or rule based systems do better / as well as / almost as well as neural methods. Citation please, I'd love a map of which ML problems are still best solved with symbolic systems. (Sincerely - it's not that I expect there aren't any.)

Gibbon1• 8 years ago

I think the the fixation on symbolic reasoning comes from ignorance at how hard classification is vs how hard pure mechanical symbolic operations are for humans. It's easy to make the mistake of thinking that since a computer can rapidly multiply two numbers together (hard for humans) that they were operating at a higher level than human brains.

Turns out this is wrong. Human brains are very efficient.

taneq3• 8 years ago

The thing is, whatever the hell it is that human brains actually do in the background to produce our 'understanding' of the world and our ability to synthesize new ways to manipulate it, we're also very good at back-fitting explanations based on symbolic reasoning. So it looks like machines need symbolic reasoning to replicate human abilities, whereas I'd bet a dollar that actually, we're doing something quite different (and messy and Bayesian and statistical) in the background and then, using the same process, coming up with a story to explain our outcome semantically. It's not insight so much as parallel construction.

darawk• 8 years ago

And no one should be surprised by this. The NN advancement of late doesn't help addressing human-style symbolic reasoning at all. All we have is a much more powerful function approximator with a drastic increased capacity (very deep networks with billions of parameters) and scalable training scheme (SGD and its variants).

You think symbolic reasoning is not a function? In what sense do you think 'symbolic reasoning' is a distinct thing from 'function approximation'?

Symmetry• 8 years ago

Something very interesting to me about the work Deep Mind has been doing is the way they've been combining neural network intuitions with tree search reasoning in Go, Chess, protein folding, etc.

cantthinkofone2• 8 years ago

If humanlike reasoning is the destination for AGI, there's more than just symbolic reasoning to factor in. Emotions are a huge control on human reasoning.

People essentially rely on emotions to make all their decisions. Emotions implicitly represent rapid-fire unconscious decision work.

Again the current popular understanding of the mind separates emotion from thinking. They are not distinct. Emotional processing is another kind of thinking, and it drives the show.

visarga6• 8 years ago

I see emotions as analogous to the value function in RL. It is essentially a prediction of future rewards based on current state and action plan. Artificial RL agents learn emotion as it is related to their tasks and environments.

anonytrary2• 8 years ago

Tangential: This title is weird. As if no one but the top minds in AI didn't know this? This isn't big news to anyone who has done even just a modicum of AI research.

ionforce• 8 years ago

The problem is when non-technical people write articles or respond to posts about Deepmind. They think all AIs are the same and that one specific AI achievement means the Matrix is coming.

People lack nuance and critical thinking.

rytill• 8 years ago

What makes you so confident that you’d say “anyone who’s done a modicum of AI research” would come to the same conclusion as you?

Also, do you believe AGI is currently more a compute/hardware problem, or an algorithmic problem?

lainga• 8 years ago

Maybe it's for the people who haven't, so that they don't give all their money to Eliezer Yudkowsky.

brandonmenc• 8 years ago

As if no one but the top minds in AI didn't know this?

Anecdotal, but nearly all of my programmer friends believe that full-blown AGI is less than a decade away.

Jach• 8 years ago

Sounds like an opportunity to place some bets and probably win some money. Or maybe they'll back down and widen their intervals -- less than a decade, maybe, but probably longer. Maybe quite a long time too, and maybe after development of some other planet-changing tech.

It's worth thinking about this section of [0] when various AI experts offer predictions:

Two: History shows that for the general public, and even for scientists not in a key inner circle, and even for scientists in that key circle, it is very often the case that key technological developments still seem decades away, five years before they show up.
In 1901, two years before helping build the first heavier-than-air flyer, Wilbur Wright told his brother that powered flight was fifty years away.
In 1939, three years before he personally oversaw the first critical chain reaction in a pile of uranium bricks, Enrico Fermi voiced 90% confidence that it was impossible to use uranium to sustain a fission chain reaction. I believe Fermi also said a year after that, aka two years before the denouement, that if net power from fission was even possible (as he then granted some greater plausibility) then it would be fifty years off; but for this I neglected to keep the citation.
And of course if you’re not the Wright Brothers or Enrico Fermi, you will be even more surprised. Most of the world learned that atomic weapons were now a thing when they woke up to the headlines about Hiroshima. There were esteemed intellectuals saying four years after the Wright Flyer that heavier-than-air flight was impossible, because knowledge propagated more slowly back then.

[0] https://intelligence.org/2017/10/13/fire-alarm/

dwiel2• 8 years ago

When I was at ICLR a couple years ago, a group of 10 or so researchers from deepmind took a poll of themselves at breakfast and found the general consensus was that AGI was between 5-10 years away.

sanxiyn2• 8 years ago

Shane Legg (DeepMind cofounder) is public about predicting AGI at 2028. http://www.vetta.org/2011/12/goodbye-2011-hello-2012/

My impression is this is common among DeepMind folks and not an aberration. (See also dwiel's comment elsewhere.) It is super weird for me that Demis Hassabis says AGI is nowhere close. Is he lying? Or does he mean 10 years is not close?

newishuser1357• 8 years ago

It is super weird for me that Demis Hassabis says AGI is nowhere close. Is he lying? Or does he mean 10 years is not close?

Maybe he just doesn’t believe the same thing some of his coworkers do? Seems pretty drastic to jump to the conclusion he’s lying if he implies it’s more than 10 years away.

joe_the_user7• 8 years ago

Research can tell you current effort fall far short. Research can tell you current efforts are moving incrementally towards the goal, even. But research won't you when something we don't understand will happen. "A long time" but overall, it seems like the kind of situation where probably as such isn't particularly applicable.

izzydata2• 8 years ago

I'm not even convinced that a real AI is possible with conventional computer hardware or anything remotely similar to it. Not even considering software I get the impression there is a fundamental limitation of hardware.

mortivore2• 8 years ago

I'm not convinced we've even defined the problem space well enough to solve it. Like what is the concrete measure(something to target) for intelligence? If we develop general intelligence is it going to be human, dog, or fish?

EliRivers• 8 years ago

I'm not convinced any of those creatures have general intelligence. I'm similarly unconvinced that we'd recognise general intelligence if we saw it.

aerodude2• 8 years ago

I think this is the real answer. When we developed flight, the measure wasn't "can we fly like birds?" We still haven't achieved that even today, but we fly in otherwise unimagined, but equally powerful ways.

We seem to be looking at intelligence in humans and thinking we need to develop that, without first defining what intelligence actually is. We don't exist in isolation, and it's likely that the components of intelligence exist to varying degrees in other organisms. In the same way that birds, bats, gliders and insects all have wings that generate lift, what are the things that we have in common with other animals?

why_only_153• 8 years ago

It seems like the difference between humans and dogs is substantially smaller than the difference between computers and dogs, so if we figure out dog-level intelligence human level intelligence is right down the corner. Also, the intelligence is likely to be of a different kind. Someone had an interesting point that training an ML system to look at picture isn't like sending a million interns to look at a million pictures, it's sending one intern to look at a million pictures. When you do that, you can derive insights that are significantly different than if you look at 1 picture, or 10, or 100.

9955336• 8 years ago

Shane Legg (co-founder DeepMind) and Marcus Hutter (Schmidhuber pedigree) defined machine intelligence in this canonical paper from 2007: https://arxiv.org/abs/0712.3329

Universal Intelligence: A Definition of Machine Intelligence
A fundamental problem in artificial intelligence is that nobody really knows what intelligence is. The problem is especially acute when we need to consider artificial systems which are significantly different to humans. In this paper we approach this problem in the following way: We take a number of well known informal definitions of human intelligence that have been given by experts, and extract their essential features. These are then mathematically formalised to produce a general measure of intelligence for arbitrary machines. We believe that this equation formally captures the concept of machine intelligence in the broadest reasonable sense. We then show how this formal definition is related to the theory of universal optimal learning agents. Finally, we survey the many other tests and definitions of intelligence that have been proposed for machines.

General intelligence usually meant in relation to humans, but you are correct in noting that it is a spectrum, not a binary.

roenxi2• 8 years ago

The computational power of the hardware is getting really close to what a human brain is capable of (on an exponential scale, anyway). If "nowhere close" means not in the next 5 years then sure.

Over the medium term I'm not sure AI researchers are the best people to ask. They are completely dependent on how much power the electrical engineers give them - I doubt there is a deeper understanding what a doubling or quadrupling of computer power will do than any programmer learning about neural networks.

anonytrary2• 8 years ago

The computational power of the hardware is getting really close to what a human brain is capable of

Why do you say that? AFAIK computing architecture and brain architecture are completely different. How would you even begin to compare their power?

roenxi2• 8 years ago

Well, Wikipedia was my source [0] and it links http://hplusmagazine.com/2009/04/07/brain-chip/ as its source.

Google has TPU that are off from the estimated power required to simulate a brain by a factor of 3, so technology is reaching the ballpark. Given that brains were evolved, the part that does symbolic thinking is probably "easy to stumble on" in some practical sense.

[0] https://en.wikipedia.org/wiki/Computer_performance_by_orders...

jacquesm• 8 years ago

I doubt there is a deeper understanding what a doubling or quadrupling of computer power will do than any programmer learning about neural networks.

Sure they do. They just hook up four times as much compute power or simulate whatever they want to do in four times as much time. A slow AGI would still be an AGI. But we do not see anything like that if we use four times as much power as in the control. It is still nowhere near.

thanatropism2• 8 years ago

I don't believe in the idea of AGI for Dreyfusard reasons, but it's possible that it could emerge from something completely different than deep learning.

For all we know, Isabelle and Coq could be speeding through the road to consciousness but we're busy having a blast doing Computer Vision pretending it's AI.

thanatropism2• 8 years ago

I'm used to random downvotes for comments about current controversies - I'll say things about income inequality that people won't like, and it's okay, you have your politics. Deep learning shouldn't be part of your politics, it either does stuff or doesn't.

Deep Learning is amazeballs for Computer Vision. It's fun because people like looking at pictures. But sufficiently prodded Isabelle proves theorems, I've seen it first hand, and the "sufficient prodding" is way underdeveloped yet. At one point backpropagation was dead too.

mindgam36• 8 years ago

Demis Hassabis (true) statements here would be much more credible if DeepMind wasn't currently making a mint by promoting AlphaZero to the masses as a "general purpose artificial intelligence system".

Don't believe me? Check out this series of marketing videos on YouTube by GM Matthew Sadler.

1. “Hi, I’m GM Matthew Sadler, and in this series of videos we’re taking a look at new games between AlphaZero, DeepMind’s general purpose artificial intelligence system, and Stockfish” (1)

2. “Hi, I’m GM Matthew Sadler, and welcome to this review of the World Champinship match between Magnus Carlsen and Fabiano Caruana. And it’s a review with a difference, because we are taking a look at the games together with AlphaZero, DeepMind’s general purpose artificial intelligence system...” (2)

3. “Hi, I’m GM Matthew Sadler, and in this video we’ll be taking a look at a game between AlphaZero, DeepMind’s general purpose artificial intelligence system, and Stockfish” (3)

I could go on, but you get my point. Search youtube for "Sadler DeepMind" and you'll see all the rest. This is a script.

But wait, you say, that's just some random unaffiliated independent grandmaster who just happens to be using an inaccurate script on his own, no DeepMind connection at all! And to that I would say, check out this same random GM being quoted directly on DeepMind's blog waxing eloquently and rapturously about AlphaZero's incredible qualities. (4)

Let's be clear. I am in no way dismissing AlphaZero's truy remarkable abilities in both chess and other games like go and shogi. Nor do I have a problem with Demis Hassabis making headlines for stating the obvious about deep learning (that it's good at solving certain limited types of puzzles, but we are a long way from AGI, why is this controversial).

My problem is that Hassabis is speaking out of both sides of his mouth. Increasing DeepMind/Google's value by many millions with his marketing message, while acting like he's not doing that. It feels intellectually dishonest.

To solve this, all DeepMind needs to stop instructing its Grandmaster mouthpieces to refer to AlphaZero as a "general articial intelligence system". Let's see how long that takes.

(1) https://www.youtube.com/watch?v=2-wFUdvKTVQ&t=0m10s (2) https://www.youtube.com/watch?v=X4T0_IoGQCE&t=0m05s (3) https://www.youtube.com/watch?v=jS26Ct34YrQ&t=0m05s (4) https://deepmind.com/blog/alphazero-shedding-new-light-grand...

rozim2• 8 years ago

I suspect "general purpose artificial intelligence system" means the same architecture applied to 3 games (western Chess, Shogi, Go).

mindgam36• 8 years ago

Wouldn't that be a "general purpose game-playing intelligence system" at best? (without mentioning that it only applies to certain types of perfect-information games)

Maybe it's just me, but "general purpose artificial intelligence system" sounds like, well, General Artificial Intelligence. Which sounds like Artificial General Intelligence, which is the holy grail.

taneq3• 8 years ago

I don't think they mean the two in the same way. AlphaZero is "general purpose artificial intelligence" because if you formulate a problem in the right way and then throw a server cluster at it for a few weeks, it often comes back with pretty good performance at solving that problem. It's probably our current best crack at creating AGI, but it's a long way from a machine that can take a very high level goal and figure out the rest for itself, which is what we usually mean by "AGI" - not just a machine that answers multiple questions, but a thing analogous to a human mind which can analyse new things, infer properties and mechanics, generalise those to new contexts, and apply that knowledge to achieve new outcomes.

mindgam36• 8 years ago

“if you formulate a [very specific type of problem involving perfect information games] in the right way”

YeGoblynQueenne3• 8 years ago

Edit: You don't have to go far to see that DeepMind are pushing AlphaZero as "general purpose". This is the title of their Science paper on it:

A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play

http://science.sciencemag.org/content/362/6419/1140

"General" as in what? As opposed to reinforcement learning, in er, general? As opposed to other ANN architectures?

> I am in no way dismissing AlphaZero's truy remarkable abilities in both chess and other games like go and shogi.

More to the point- it's only chess, go and shogi; not games "like" those.

The AlphaZero architecture has the structure of a chessboard and the range of moves of pieces in chess, go and shogi hard-coded and you can't just take a trained AlphaZero model and apply it to a game that doesn't have either the board or the moves of those three games.

To be blunt, AlphaZero has mastered chess, go and shogi, but it can't play noughts-and-crosses.

goolulusaurs• 8 years ago

It bothers me that the qoutes in this article are all cut up, in some cases ending when a sentence clearly wasn't finished. It makes it hard to judge what they are really saying here, and I wish the full interview would be published.

nikkwong• 8 years ago

Not sure how I feel about this; for one, the Kurzweilian singularity which largely could be fueled by the advent of AGI is both exciting and yet also scary. The upside could forever change humanity as we know it; far increased longevity, the potential to create anything via a universal assembler[0], bringing everything feasible within the laws of physics to reality. Knowledge is the only limiting factor stopping us from doing anything which is physically possible in this universe; and in that light AGI could be an enlightenment.

On the other hand, the ubiquity of knowledge once it's available could lead any maniac to use it for the wrong purpose and wipe out humanity from their basement.

My feelings on the potential of AGI is therefore mixed. I for one have just found my particular niche in the workforce and am finally reaping the dividends from decades of hard work. Having AGI displace me and millions (or billions) of individuals is frightening and definitely keeps me on my toes.

Technology changes the world; my parents both worked for newspapers and talk endlessly about how the demise of their industry after the advent of the internet is so unfortunate. Luckily for them they are both at retirement age so their livelihood was not upset by displacement.

If AGI does become a thing it will be interesting to see how millenials and gen Z react to becoming irrelevant in what would have been the peak of their careers.

[0] https://en.wikipedia.org/wiki/Molecular_assembler

diminish• 8 years ago

I have a small experiment to discover if AGI is already a solved puzzle.

https://news.ycombinator.com/item?id=18720482

nuguy13• 8 years ago

I take huge offense to this article. They claim that when it comes to AGI, Hinton and Hassabis “know what they are talking about.” Nothing could be further from the truth. These are people who have narrow expertise in one framework of AI. AGI does not yet exist so they are not experts in it, in how long it will be, or how it will work. A layman is just as qualified to speculate about AGI as these people so I find it to be infinitely frustrating when condescending journalists talk down to the concerned layman. This irritates me because AI is a death scentance for humanity — its an incredibly serious problem.

As I have stated before, AI is the end for us. To put it simply, AI brings the world into a highly unstable configuration where the only likely outcome is the relegation of humans and their way of life. This is because of the fundamental changes imposed on the economics of life by the existence of AI.

Many people say that automation leads to new jobs, not a loss of jobs. Automation has never encroached on the sacred territory of sentience. It is a totally different ball game. It is stupid to compare the automation of a traffic light to that of the brain itself. It is a new phenomenon completely and requires a new, from-the-ground-up assessment. Reaching for the cookie-cutter “automation creates new jobs” simply doesn’t cut it.

The fact of the matter is that even if most of the world is able to harness AI to benefit our current way of life, at least one country won’t. And the country that increases efficiency by displacing human input will win every encounter of every kind that it has with any other country. And the pattern of human displacement will ratchet forward uncontrollably, spreading across the whole face of the earth like a virus. And when humans are no longer necessary they will no longer exist. Not in the way they do now. It’s so important to remember that this is a watershed moment — humans have never dealt with anything like this.

AI could come about tomorrow. The core algorithm for intelligence is probably a lot simpler than is thought. The computing power needed to develop and run AI is probably much lower than it is thought to to be. Just because DNNs are not good at this does not mean that something else won’t come out of left field, either from neurological research or pure AI research.

And as I have said before, the only way to ensure that human life continues as we know it is for AI to be banned. For all research and inquires to be made illegal. Some point out that this is difficult to do but like I said, there is no other way. I implore everyone who reads this to become involved in popular efforts to address the problem of AI.

nradov5• 8 years ago

Stating something more times doesn't make it true. Everything you've written is pure speculation, and alarmist at that. There's no proof that AGI is even possible, and if it is possible there's no proof that it will end humanity.

nuguy13• 8 years ago

Even anti-alarmists don’t ask for proof that AGI is possible. Obviously it is possible. Speculation is the best you get because nobody is going to be able to prove anything. We haven’t proven global warming is caused by humans but it’s still worth it to be proactive about greenhouse gasses. This is because when something is extremely dangerous, you don’t wait around for someone to finish proving it beyond any shadow of a doubt. You probably also think that god exists because nobody can prove otherwise?

And what does alarmist even mean? Do you call global warming advocates alarmists? It’s such an annoying, nonsense word that boils down to name-calling really. Discuss the merits of my actual argument. If you think my speculation is wrong, point out a flaw in the chain of logic that leads to my conclusion. Don’t just wave your hand and say that “you can’t prove it” like some evangelical christian talking about god or global warming. Seriously infuriating when there is so much at stake.

9955336• 8 years ago

Clearly, AGI-level intelligence is possible, because human brains exist.

So unless you pose that a function has to rely on its materialization (there is something untouchably magic about biological neural networks, and intelligence is not multiple realizable), it should be possible to functionally model intelligence. Nature shows the way.

AGI will likely obsolete humanity. Either depricate it, or consume it (make us part of the Borg collective). Heck, even a relatively dumb autonomous atom bomb or computer virus may be enough to wipe humanity from the face of the earth.

AgentME2• 8 years ago

There are non-doomsday possibilities for AGI. Imagine a super-intelligent AI that was built from the start to value humanity and our way of life, and from there it chooses to protect us and enable us all to do what we want more. (Of course, this could go in dystopian directions, but even those are better than extinction.) A super-intelligent AI that is built to value humans could decide to "uplift" humans' intelligence to be able to keep up with it in places that humans desire to keep up with it.

If we can figure out decision theory and how our values work, then when we figure out AI, we can hopefully build it to be aligned with our values from the start, instead of blindly hoping it happens to play nice with us instead of brushing us off like ants.

nuguy13• 8 years ago

You don’t even begin to comprehend what I’m saying. You need to think about this more deeply.

So what if it is possible to create a benevolent ai? Nobody said this isn’t possible or even likely. We can also invent a machine that scrubs all the moss off of stones. Just because it’s possible for it to exist doesn’t mean it’s going to proliferate in the free-market of the world and everything in it. The only thing that is important is the fact that

1: we will enter an unstable configuration where any AI implementation that can exist will exist

2: the AI implementations that proliferate will be those that are not hamstrung by being forced to include humans in the loop

3: humans will be out of the loop for every conceivable task and therefore not enjoy the high standard of living that they do in 2018

sophistication• 8 years ago

AI is likely simple and won't require much processing power after all, so it will be impossible to ban it because it would imply more regulation and surveillance than what would be sustainable. Also global warming will likely kill us anyhow. The rational conclusion is to enjoy our supermarkets and warm showers as long as they last. They will probably last longer if we deny these threats so as to avoid causing mass panic and nihilism.

nuguy13• 8 years ago

We can survive global warming. We can fix it. We can come back from it. We will never come back from ai. It’s not impossible to ban ai. And we would be stupid to assume it’s impossible instead of trying to find out through an effort to save ourselves.

jeremyjh• 8 years ago

the only way to ensure that human life continues as we know it is for AI to be banned

Is that because you think banned things do not happen? Even if the thing that is banned could confer a massive advantage to the entities developing it?

I think AGI is unlikely to be a thing in my lifetime, or even my children's. But if I were worried about it, I'd probably focus on developing a strategy to create a benevolent intelligence FIRST, rather than try to prevent everyone else from ever creating one via agreements and laws.

nuguy13• 8 years ago

I appreciate that you actually suggest a solution. Nobody knows when AGI will come but it could come tomorrow. It could come in 1000 years. No harm in being proactive.

Developing a good ai first is useless because as I have said, the creation of ai enters us into an unstable configuration where bad ai will crop up regardless. Keeping bad ai from existing is infinitely easier when ai does not exist as a technology as opposed to when it’s a turnkey thing.

alexnewman• 8 years ago

"These are people who have narrow expertise in one framework of AI." Proof that you don't know who you are insulting

nuguy13• 8 years ago

I don’t remember insulting anyone. And how is that not true?

9955336• 8 years ago

You are correct, in that AI experts may not be the best predictors for AGI. For one, they spend their lives working towards the goal of AGI, so it would require a huge amount of cognitive dissonance for them to say that AGI is impossible or very very far on the horizon.

Philosophers and futurists are better suited to hypothesize an AGI timeline.

But you take it too far by saying it is anyone's game.

Game theory, security, and economic competition makes it impossible to globally ban AI. The incentives to automate the economy (compare AI revolution with industrial revolution) and to weaponize AI (Manhattan Project for intelligence) are just too big. We are already seeing that the US focus on fair and ethical AI puts them at a disadvantage against China and Russia. AGI must require pervasive surveillance of the populace, but the Luddites are holding this back.

I suggest you learn to stop worrying about the bomb, and start planning for its arrival.

p1esk3• 8 years ago

* AI to be banned*

Good luck with that.

nuguy13• 8 years ago

So sarcasm is what will save us all?

rozim2• 8 years ago

Ilya Stuskever of OpenAI says 5 years:

https://medium.com/intuitionmachine/near-term-agi-should-be-...

why_only_153• 8 years ago

I watched the talk linked where that quote apparently comes from, and it was really good. Thanks for sharing that. Ilya specifically says in the talk that it is unlikely but that there is sufficient lack of understanding that we can't rule it out, and that thus the questions around it are worth thinking about.

qwerty456127• 8 years ago

If AGI (an artificial human mind with direct access to computational power of classic computers and whole Internet of information) was possible then we would probably already be living in the Travelers TV show.

mindcrime2• 8 years ago

... was possible then we would probably already be living in the Travelers TV show.

How do you know we aren't?

BTW, if you hadn't noticed, Season Three just came out on Netflix. I'm champing at the bit to binge watch that... :-)

9955336• 8 years ago

If AGI is possible, it already happened. If even AI experts put it a 100-1000 years out, where some human monkeys banging on digital typewriters could eventually create it, then, in the vastness of space, time, military contracts, alien intelligences, and random Boltzmann brains, it must have been reality multiple times already.

If AGI is impossible, it will never happen. We already know that perfectly intelligent AGI's are not physically possible: Per DeepMind's foundational theoretical framework, optimal compression is non-computable, and besides that, it is not possible for an inference machine to know all of its universe (unless it is bigger than the universe by at least 1 bit, AKA it is the universe).

Remains being more intelligent than all of humanity. To accomplish that, by Shannon's own estimates, there is currently not enough information available in datasets and the internet. Chinese efforts to artificially increase the intelligence of babies is still in its infancy too (the substrate of AGI is irrelevant for computationalism, unless it absolutely needs to run on the IBM 5100).

So until that time travels, we will have to make due with being smarter than/indistinguishable from a human on all economic tasks. We're already there for some subset of humanity, you may even be a part of that subset, if you believed this post was written by a human.

epicureanideal• 8 years ago

https://en.wikiquote.org/wiki/Incorrect_predictions

"Hence, if it requires, say, a thousand years to fit for easy flight a bird which started with rudimentary wings, or ten thousand for one which started with no wings at all and had to sprout them ab initio, it might be assumed that the flying machine which will really fly might be evolved by the combined and continuous efforts of mathematicians and mechanicians in from one million to ten million years--provided, of course, we can meanwhile eliminate such little drawbacks and embarrassments as the existing relation between weight and strength in inorganic materials. [Emphasis added.] The New York Times, Oct 9, 1903, p. 6."

-----

A couple of the leading minds in AGI say it's a long ways away... just because the universe likes to give us the finger, maybe AGI is on the horizon. Maybe we'll look back at this in 10 years and laugh (if we're here).

cr0sh• 8 years ago

Interestingly, Hinton is on record as essentially saying that there's a good possibility that what's currently being done is wrong - and that we need to rethink our approach.

Mainly in the idea/concept of back-propagation. It's something that I've thought about myself. For the longest time, I could never understand how it worked, then I went thru Ng's "ML Class" (in 2011, which was based around Octave), and one part was developing a neural network with backprop - and the calcs being done using linear algebra. It suddenly "clicked" for me; I finally understood (maybe not to the detailed level I'd like - but to the general idea) how it all worked.

And while I was excited (and still am) by that revelation, at the same time I thought "this seems really overly complex" and "there's no way this kind of thing is happening in a real brain".

Indeed, as far as we've been able to find (although research continues, and there's been hints and model which may challenge things) - brains (well, neurons) don't do backprop; as far as we know, there's no biological mechanism to allow for backprop to occur.

So how do biological brains learn? Furthermore, how are they able to learn from only a very few examples in most cases (vs the thousands to millions examples needed by deep learning neural networks)?

We've come up with a very well engineering solution to the problem, that works - but it seems overly complex. We've essentially have made an airplane that is part ornithopter, part fixed-wing, part balloon, and part helicopter. Sure it flies - but it's rather overly complex, right?

Humanity cracked the nut when it came to heavier-than-air flight when it finally shed the idea that the wings had to flap. While it was known this was the way forward long before the Wright's or even Langley (and likely even before Lilienthal), a lot of wasted time and effort went into flying machines with flapping wings, because it was thought that "that's the way birds do it, right"?

So - in addition to the idea that backprop may not be all it's cracked up to be - what if we also need to figure out the "fixed wing" solution to artificial intelligence? Instead of trying to emulate and imitate nature so closely, perhaps there's a shortcut that currently we're missing?

I do recall a recent paper that was mentioned here on HN that I don't completely understand - that may be a way forward (the paper was called "Neural Ordinary Differential Equations"). Even so, it too seems way too complex to be a biologically plausible model of what a brain does...

virgilp• 8 years ago

You're contradicting yourself with your examples. If we didn't manage to fly by imitating birds - why do you care that AI doesn't work the brain does? That should be a _good_ sign, if we trust the analogy - right?

nabla9• 8 years ago

The arguments like above are "platitude level arguments".

We really don't learn anything from the problem in had by talking in generic terms. We use these arguments when we want to justify our hopes and feeling, but there is really nothing to learn from it.

Hinton, Hassabis, Bengio and others point out that we can't 'brute force' AI development. There needs to be actual breakthroughs in the field and there may be several decades between them.

AI, brain science and cognitive science are extremely difficult fields with small advances, yet people assume that it's possible to 'brute force' AGI by just adding more computing power and doing more of the same.

Macroeconomics is probably less complex research subject than AI or brain science, but nobody assumes that you can just brute force truly great macroeconomic model in few years if you just spend little more resources.

edanm• 8 years ago

AI, brain science and cognitive science are extremely difficult fields with small advances, yet people assume that it's possible to 'brute force' AGI by just adding more computing power and doing more of the same.

Do people assume that? I mean, I'm sure some people do, but I don't think I've encountered many people, at least not in the AI safety movement, that actually think it's a matter of more hardware power. Some people think it's possible that that's all that's necessary, but I don't think most will say that that's the most likely path to AGI (rather than, as you say, actual breakthroughs happening).

YeGoblynQueenne3• 8 years ago

For flight, the components necessary were obvious very early on: you need some kind of structure to hold you aloft and some kind of powered apparatus to propel you forwards. Once those were found, mechanical flight was achieved (and unpowered flight was already possible long before that).

What are the components of intelligence? For example, AlphaZero can solve problems that are hard for humans to solve in the domain of chess, shogi and go- is it intelligent? Is its problem-solving ability, limited as it is to the domain of three board games, a necessary component of general intelligence? Have we even made any tiny baby steps on the road to AGI, with the advances of the last few years, or are we merely chasing our tails in a dead end of statistical approximation that will never sufficiently, well, approximate, true intelligence?

These are very hard questions to answer and the most conservative answers suggest that AGI will not happen in a short time, as a sudden growth spurt that takes us from no-AGI to AGI. With flight, it sufficed to blow up a big balloon with hot air and- tadaaaa! Flight. There really seems to be no such one neat trick for AGI. It will most likely be tiny baby steps all the way up.

cantthinkofone2• 8 years ago

If you want AGI you need to give it a world to live in. The ecological component of perception is missing. Without full senses, a machine doesn't have a world to think generally about. It just has the narrow subdomain of inputs that it is able to process.

You could bet that AGI won't manifest until AI and robotics are properly fused. Cognition does not happen in a void. This image of a purely rational mind floating in an abyss is an outdated paradigm to which many in the AI community still cling. Instead, the body and environment become incorporated into the computation.

mikhailfranco• 8 years ago

Great interview with Hassabis from the BBC. It's meanderingly biographical, with insights about his path through internships, curiosity, startups, commitment, burnout, trusted team mates and eventual successes ...

https://www.bbc.co.uk/sounds/play/p06qvj98

magwa101• 8 years ago

Yep and considering a very high majority of our work does not require GI we still have huge AI job disruption looming.

hyperpallium• 8 years ago

Progress-skeptics are always wrong - except for artificial inteligence.

Geoffrey Hinton and Demis Hassabis: AGI is nowhere close to being a reality

Related Stories