This may be slightly OT, but is it unfair of me to assume that science reporting done by mainstream outlets is at worst so inaccurate and at best half true; to the point that when it involves subjects I know little about, I simply avoid the article entirely, and wait for a reasonably credible source to comment on the matter?
Kevin Kelly is an opponent of the most optimistic AI researchers. For him to use a title like that is a bit of a concession.
This is the closest you are going to get to something that is more-or-less accurate and that you can accept.
He is reporting that because he has to. Deep learning is not something anyone can dismiss. If Kevin Kelly thought he could downplay the massive importance of deep learning and similar technologies, or say that it wasn't AI, he would.
This is actually a fairly pessimistic article in terms of AI prediction, and yet, that title was absolutely justified.
AI is here and very powerful, and no one can deny it. People can still pretend, however, that it will never have human-like capabilities, or cannot become smarter than us (at least not in _our_ lifetimes). Give it five to ten years, human-like and beyond-human intelligence will be built, Kevin Kelly will be talking to it, and he will have to write an article about how he was wrong about everything (of course he won't admit that much).
> Give it five to ten years, human-like and beyond-human intelligence will be built
It is absolutely safe to say that human-like AI in five years is not going to happen. Even ten years seems highly unlikely. All leading researchers in the field agree on this.
If I had a human-level-AI capable machine delivered today to my basement, then it'd still take something like 5 years of training to bring it to a human-level intelligence.
Babies are born with all the hardware to have an adult-level intelligence, but converting that potential to actual skills requires months/years for each skill of active learning, experimentation and feedback, not only reading information from the web.
> Give it five to ten years, human-like and beyond-human intelligence will be built
That's probably a bit early. The Machine Intelligence Research Institute has collected various surveys of experts asked about this question. The one with the earliest estimates asked the question, “Assuming beneficial political and economic development and that no global catastrophe halts progress, by what year would you assign a 10%/50%/90% chance of the development of artificial intelligence that is roughly as good as humans (or better, perhaps unevenly) at science, mathematics, engineering and programming?” Of 19 replies, the median estimates for 10%, 50%, and 90% were 2025, 2035, and 2070, respectively."
If you asked those same experts in 2005 if a Jeopardy computer would beat the best human Jeopardy contestants ever in 2011, what percentage of them would go on the record with the correct answer?
Yes, it would have been hard for them to predict that anybody would decide to tackle that particular task. The question from GP is about a much broader range of tasks.
Watson was designed to answer Jeopardy questions. If Alex Trebek had used natural language to ask it anything other than a jeopardy question it would have failed spectacularly. ie, "So where are you from?".
It was tuned for that particular task, but the core tech is much more general - the current main application for which IBM is selling the Watson platform is healthcare, which is "a bit" different from answering Jeopardy questions.
Not really. "A runny nose, a fever, body aches" "What are symptoms of flu the with 75% accuracy?"
That's dumbed down of course. It's probably more like "This much of this enzyme concentration in blood, that thing in urine"
"What is a 17% chance of developing some condition in the next 5 years?"
Correct, but that's just because of the inherently unpredictable nature of future telling.
What are the alternatives? One could simply not engage in projections, which is difficult because they're so tempting. A third option is to listen to the non-experts, which doesn't seem more valuable...
This is something I struggle WRT economics. I know economists are little more than fortune tellers, but what are the other options?
A more useful question: How far away is Microsoft Middle Manager 3.0, from Yu's "How To Live Safely In A Science Fictional Universe"? AIs will take over when they start outperforming humans at management. When the most profitable companies are computer-run, capitalism has to let them take charge. As soon as computers can do management at all well, they may well be better at it than humans simply because they can communicate faster.
There's already a hedge fund with an AI on its board. Really.
>OUR MOST PREMIUM AI SERVICES WILL BE ADVERTISED AS CONSCIOUSNESS-FREE
I always assumed that breakthroughs in AI would come from philosophy than computer science. We don't even know if have the capability to define the word 'consciousness'.
If General Relativity could benefit from the application of pre-existing pure mathematics (Riemannian geometry), why shouldn't it be conceivable that computer science could benefit from (possibly ancient) philosophical ideas?
Because General Relativity primarily benefited from evidence that a new model was necessary. A-priori thinking can help to expand the space of possible answers, but it's terrible at locating uniquely correct answers within any given space.
I don't accept that premise. It took a while to verify the predictions of General Relativity. It certainly wasn't driven by experimental evidence. Einstein started work on GR with a thought experiment (according to Wikipedia) and of course there's Minkowski's famous observation that SR could have been inferred before Einstein from Galilean Relativity, spatial isotropy and the probable existence of a geometric invariant. In other words, pure math didn't get there, but once they all arrived it could really give insight. Check out Minkowski's famous "staircase wit" comment.
Another example from physics is the extent to which Schrodinger was influenced by Eastern philosophy in his formulation of quantum mechanics.
Is logic perception? Famously (I have to take them at their word) Quantum Physics describes behaviours which require the mathematician to abandon perception and manipulate the formula "blindly". Should we disbelieve them?
Also I thought science was (formally) the creation of beliefs based on falsifiable hypothesis supported by evidence? When did the news come that it's actually about excluding intuition and experience? I think quite a lot of grant applications are going to have to be rejected henceforth, but then, how shall we decide what to observe next?
I think good understanding of conciousness comes from the meditative tradition, and their definition for it is "the part of the brain that observes the other parts".
BTW we've already implemented consciousness in a robot: http://www.scientificamerican.com/article/automaton-robots-b... and it's useful. But AFAIK, i don't think it's used in any commercial system, although that's the kind of thing a commercial company won't advertise.
>Droids met the challenge of perceiving their self-image and reflecting on their own thoughts as part an effort to develop robots that are more adaptable in unpredictable situations
Who exactly is this entity doing the reflecting? .Isn't this a fundamental misunderstanding of what 'consciousness' means?
As a practical matter, especially for the robot ,it might not matter who is observing. The fact that one can observe oneself enables further capabilities like better feedback.
And by the way i think this observing robot has(or can have) an internal definition of self, and an observer beyond the self. It seems possible.
2. Philosophy recognizes that we have created a real AI.
Actually, 1 might have happened. I think that once we get a concrete understanding of consciousness, we will come to recognize that we have already made conscious systems.
It a carelessly-used word. We use it both for 'not unconscious' and for 'self-aware'. The 1st is obvious - unless its offline, the AI is not 'knocked out' or asleep so its 'conscious'.
The 2nd is harder. Maybe by close questioning you could find out. "If you were self-aware, what would you think about yourself?" etc.
From philosophy? Are you kidding? There's simply no way AI is ever going to come from a bunch of people arguing over what is "qualia" and what is "consciousness". There have been huge advances made in deep neural networks in the last few years. To the point where we now have human-competitive image recognition. That is computer science, and that is where AI is going to come from.
I think it will take a little while longer, maybe a few decades, and happen organically. At first, we'll have smart machines based on deep learning, possibly embedded in specialized hardware. We'll have smarter and smarter machines. Specialized robots doing basic tasks (e.g.: dish washing) are going to start roaming the world. They will be made increasingly capable as time goes on, and at some point, we'll begin to cross some sort of threshold where a robot's understanding of the world will be great enough. The robot will seem self-aware and "conscious" to an outside observer.
Computer scientists don't really need to care about the philosophy of consciousness, they care about building capable machines which can deal with the complexities world. At some point, to match or rival human capability, self-awareness is needed. That is, a model of the world that is "complete enough" and incorporates a model of you as an agent within this world. It will happen naturally, just as it did in nature.
CS isn't going to stumble onto AI, someone's going to have to theorize about it first. Much like CS doesn't "stumble" onto new equations for modeling physics, CS isn't going to accidentally build an AI.
A generalist AI must be immensely difficult to create, but won't the most difficult part be the design? I don't think it'll be CS researchers coming up with the design, either -- it'll be doctors, biologists, neurologists, etc.
In my opinion, CS is going to provide the basic components (neural networks, computer vision, planning engines, knowledge bases) and engineers will put them together in a useful way. I think the reason CS hasn't provided AI yet is because general intelligence (as implemented in the human brain) is not one magic universal algorithm, it's an assembly of many necessary components, some of which are specialized for specific purposes.
The human brain was created ad hoc through evolution, and its complexity is beyond what "pure" CS usually deals with, which is algorithms that can be described in one page or less. Creating a self-aware AI is going to take engineering work, experimentation and incremental improvement, rather than someone inventing an overarching design and going "yes, that's it, that will produce self-awareness".
I see that the term 'self-awareness' is being used a lot in this discussion. Can it be defined simply?
Is it basically a mechanism which divides intelligence into two parts one of which is 'observing' the other ?
To me, self-awareness is, as I said before, having a model of the world that is complete enough that it incorporates a model of yourself within this world as well (and possibly other agents you encounter in this world). I guess I would add that the model should be sophisticated enough to be able to model your interactions with the world and make some sensible predictions about the future. I suppose according to this definition, a smart thermostat could be considered self-aware, but the bar for self-awareness goes up with the complexity of the interactions you can have with the world.
> Much like CS doesn't "stumble" onto new equations for modeling physics
Physicists come up with new equations for modeling physics. Metaphysicians make up elaborate reasons why they're actually intellectuals and not just on privileged-white-man welfare.
This article says the three breakthroughs were: parallel computation, big data and better algorithms.
I think we have a tendency to simplify scientific achievements. There's this romantic notion that real progress happens through eureka moments. Each of these "breakthroughs" are made up of many separate discoveries/inventions that, when aggregated, lead to modern artificial intelligence.
So the real question is - what types of new discoveries do we need to keep advancing AI? I'm not an expert in the field, but here are my guesses:
- Transfer learning is huge. How can we apply the data used from one problem to solve a different problem?
- Generalized pattern matching. Can we use the same algorithm that identifies separate objects in vision to identify different noises? Can we map these problems to the same dimensions?
- Better training sets. It takes years for a child to learn object permanence, much less speaking, listening and walking. What data sets can we feed a computer for it to learn about the real, non-virtual world around it?
- NLP based off of learning from real world datasets. Can we give a computer data from Google Glass and let it learn edge detection and words and then applying those words to the things it sees? Perhaps progress in NLP will come teaching a computer words the way you would a child instead of hard coding definitions. If you like Wittgenstein, I'd say it's a move away from Tracatus and towards Philosophical Investigations.
I'd like to know from AI experts on HN. This is the third wave of AI hype in 50 years. Will this too lead to an AI winter like the last two times or are we really onto something ?
This time there are applications and revenue. In the financial sector, huge amounts of revenue. In the first two "AI revolutions", the startups all went bust.
(I went through Stanford CS in 1985, just at the point that it was clear that expert systems were not going to be very useful. Much of the CS faculty back then was in denial about that.
More recently, I took a machine learning course. It was taught by someone from Black Rock Capital, not an academic.)
The machine learning course, given at Hacker Dojo about three years ago, was right out of Andrew Ng's videos, pre-Coursera, with his awful blackboard handwriting. Support vector machines, that sort of thing. Not enough graphics showing what all those functions are doing. Most of that stuff has a clear graphical representation, but in Ng's original course, you're supposed to imagine it.
A Stanford MSCS in the mid-1980s was heavy on mathematical logic. You could get through the whole curriculum without doing a floating point operation. I got discrete math, number theory (Knuth, vol. 2), proof of correctness, formal logic, proof of correctness, theory of programming languages, etc. Some graphics work, on a Xerox Alto. Dr. John's Mystery Hour, "Epistemological Problems in Artificial Intelligence", from John McCarthy. (He would describe a problem informally, then a miracle occurs, and it's in a predicate calculus notation where you just turn the crank to get the answer. It hadn't yet hit the logicians that getting the problem into the formalism is the hard part. Automatic symbolic math was in its infancy back then.)
"AS AIS DEVELOP, WE MIGHT HAVE TO ENGINEER WAYS TO PREVENT CONSCIOUSNESS IN THEM—OUR MOST PREMIUM AI SERVICES WILL BE ADVERTISED AS CONSCIOUSNESS-FREE."
No. Simply, no. We don't understand consciousness well enough in our own minds to understand how to stop that mechanism happening in another mind. As of yet, we have no mechanism that produces any of the insightful, creative, general intelligence that we see in humans. Even basic biological processes like walking over uneven terrain, flight by flapping of wings and picking up novel and oddly-shaped objects that haven't been seen before are challenges that we haven't even begun to master.
The hyperbole of these articles makes it seem as if creative machine intelligence is right around the corner. What we have done is make statistical pattern matching algorithms. They aren't learning in the way that a child learns through repetition. We simply don't know how general intelligence works well enough to do this.
I had to stop reading after that line. “Our most premium services will be advertised as consciousness-free” literally makes as much sense as a pre-Wright-Brothers article predicting that “our most premium flying machines will be advertised as free of feathers.” It’s just a non-sequitur given a grasp of the basic principles, whether it be flight or AI.
Deep learning is beyond statistical pattern matching. It does involve both supervised and unsupervised learning. Deep learning is currently the most successful technique, but not the most ambitious approach in AGI.
Google for 'AGI', 'deep learning', 'sparse autoencoder', 'hierarchical hidden Markov model', 'OpenCog', 'spiking neural network', 'Hierarchical Temporal Memory'
59 comments
[ 3.1 ms ] story [ 92.7 ms ] threadThis is the closest you are going to get to something that is more-or-less accurate and that you can accept.
He is reporting that because he has to. Deep learning is not something anyone can dismiss. If Kevin Kelly thought he could downplay the massive importance of deep learning and similar technologies, or say that it wasn't AI, he would.
This is actually a fairly pessimistic article in terms of AI prediction, and yet, that title was absolutely justified.
AI is here and very powerful, and no one can deny it. People can still pretend, however, that it will never have human-like capabilities, or cannot become smarter than us (at least not in _our_ lifetimes). Give it five to ten years, human-like and beyond-human intelligence will be built, Kevin Kelly will be talking to it, and he will have to write an article about how he was wrong about everything (of course he won't admit that much).
It is absolutely safe to say that human-like AI in five years is not going to happen. Even ten years seems highly unlikely. All leading researchers in the field agree on this.
Babies are born with all the hardware to have an adult-level intelligence, but converting that potential to actual skills requires months/years for each skill of active learning, experimentation and feedback, not only reading information from the web.
That's probably a bit early. The Machine Intelligence Research Institute has collected various surveys of experts asked about this question. The one with the earliest estimates asked the question, “Assuming beneficial political and economic development and that no global catastrophe halts progress, by what year would you assign a 10%/50%/90% chance of the development of artificial intelligence that is roughly as good as humans (or better, perhaps unevenly) at science, mathematics, engineering and programming?” Of 19 replies, the median estimates for 10%, 50%, and 90% were 2025, 2035, and 2070, respectively."
Other surveys and research are described here: http://intelligence.org/2013/05/15/when-will-ai-be-created/
What is expertise today, tomorrow is passé.
It was definitely built for one specific task.
That's dumbed down of course. It's probably more like "This much of this enzyme concentration in blood, that thing in urine" "What is a 17% chance of developing some condition in the next 5 years?"
What are the alternatives? One could simply not engage in projections, which is difficult because they're so tempting. A third option is to listen to the non-experts, which doesn't seem more valuable...
This is something I struggle WRT economics. I know economists are little more than fortune tellers, but what are the other options?
http://www.nytimes.com/interactive/2011/12/06/science/201112...
There's already a hedge fund with an AI on its board. Really.
5-10 years is the default prediction for everything nowadays... We've been 5 - 10 years away from holographic storage mediums for the last 20 years...
Also:
http://www.goodreads.com/quotes/65213-briefly-stated-the-gel...
I always assumed that breakthroughs in AI would come from philosophy than computer science. We don't even know if have the capability to define the word 'consciousness'.
So I would only expect a breakthrough from philosophy if philosophy is integrating new results from computer science (or neuroscience, etc).
Another example from physics is the extent to which Schrodinger was influenced by Eastern philosophy in his formulation of quantum mechanics.
It's a philosophical debate, but the short version is No, a google search isn't conscious. It's AI, or machine learning, but it isn't self aware.
[0] http://en.wikipedia.org/wiki/Chinese_room
So can my toaster. I don't think my toaster has consciousness.
Also I thought science was (formally) the creation of beliefs based on falsifiable hypothesis supported by evidence? When did the news come that it's actually about excluding intuition and experience? I think quite a lot of grant applications are going to have to be rejected henceforth, but then, how shall we decide what to observe next?
So it seems to me that philosophers and computer scientists share some key behaviors.
Based on your own theory of shared behavior, computer scientists are philosophers.
Therefore, quite a bit of philosophy was required for google search.
BTW we've already implemented consciousness in a robot: http://www.scientificamerican.com/article/automaton-robots-b... and it's useful. But AFAIK, i don't think it's used in any commercial system, although that's the kind of thing a commercial company won't advertise.
Who exactly is this entity doing the reflecting? .Isn't this a fundamental misunderstanding of what 'consciousness' means?
And by the way i think this observing robot has(or can have) an internal definition of self, and an observer beyond the self. It seems possible.
1. We create a "real" AI -- not necessarily a GAI
2. Philosophy recognizes that we have created a real AI.
Actually, 1 might have happened. I think that once we get a concrete understanding of consciousness, we will come to recognize that we have already made conscious systems.
The 2nd is harder. Maybe by close questioning you could find out. "If you were self-aware, what would you think about yourself?" etc.
I think it will take a little while longer, maybe a few decades, and happen organically. At first, we'll have smart machines based on deep learning, possibly embedded in specialized hardware. We'll have smarter and smarter machines. Specialized robots doing basic tasks (e.g.: dish washing) are going to start roaming the world. They will be made increasingly capable as time goes on, and at some point, we'll begin to cross some sort of threshold where a robot's understanding of the world will be great enough. The robot will seem self-aware and "conscious" to an outside observer.
Computer scientists don't really need to care about the philosophy of consciousness, they care about building capable machines which can deal with the complexities world. At some point, to match or rival human capability, self-awareness is needed. That is, a model of the world that is "complete enough" and incorporates a model of you as an agent within this world. It will happen naturally, just as it did in nature.
A generalist AI must be immensely difficult to create, but won't the most difficult part be the design? I don't think it'll be CS researchers coming up with the design, either -- it'll be doctors, biologists, neurologists, etc.
The human brain was created ad hoc through evolution, and its complexity is beyond what "pure" CS usually deals with, which is algorithms that can be described in one page or less. Creating a self-aware AI is going to take engineering work, experimentation and incremental improvement, rather than someone inventing an overarching design and going "yes, that's it, that will produce self-awareness".
Physicists come up with new equations for modeling physics. Metaphysicians make up elaborate reasons why they're actually intellectuals and not just on privileged-white-man welfare.
I think we have a tendency to simplify scientific achievements. There's this romantic notion that real progress happens through eureka moments. Each of these "breakthroughs" are made up of many separate discoveries/inventions that, when aggregated, lead to modern artificial intelligence.
So the real question is - what types of new discoveries do we need to keep advancing AI? I'm not an expert in the field, but here are my guesses:
- Transfer learning is huge. How can we apply the data used from one problem to solve a different problem?
- Generalized pattern matching. Can we use the same algorithm that identifies separate objects in vision to identify different noises? Can we map these problems to the same dimensions?
- Better training sets. It takes years for a child to learn object permanence, much less speaking, listening and walking. What data sets can we feed a computer for it to learn about the real, non-virtual world around it?
- NLP based off of learning from real world datasets. Can we give a computer data from Google Glass and let it learn edge detection and words and then applying those words to the things it sees? Perhaps progress in NLP will come teaching a computer words the way you would a child instead of hard coding definitions. If you like Wittgenstein, I'd say it's a move away from Tracatus and towards Philosophical Investigations.
Heres an interesting AI company I've found : http://www.celaton.com/solutions
It's already begun.
(I went through Stanford CS in 1985, just at the point that it was clear that expert systems were not going to be very useful. Much of the CS faculty back then was in denial about that.
More recently, I took a machine learning course. It was taught by someone from Black Rock Capital, not an academic.)
A Stanford MSCS in the mid-1980s was heavy on mathematical logic. You could get through the whole curriculum without doing a floating point operation. I got discrete math, number theory (Knuth, vol. 2), proof of correctness, formal logic, proof of correctness, theory of programming languages, etc. Some graphics work, on a Xerox Alto. Dr. John's Mystery Hour, "Epistemological Problems in Artificial Intelligence", from John McCarthy. (He would describe a problem informally, then a miracle occurs, and it's in a predicate calculus notation where you just turn the crank to get the answer. It hadn't yet hit the logicians that getting the problem into the formalism is the hard part. Automatic symbolic math was in its infancy back then.)
No. Simply, no. We don't understand consciousness well enough in our own minds to understand how to stop that mechanism happening in another mind. As of yet, we have no mechanism that produces any of the insightful, creative, general intelligence that we see in humans. Even basic biological processes like walking over uneven terrain, flight by flapping of wings and picking up novel and oddly-shaped objects that haven't been seen before are challenges that we haven't even begun to master.
The hyperbole of these articles makes it seem as if creative machine intelligence is right around the corner. What we have done is make statistical pattern matching algorithms. They aren't learning in the way that a child learns through repetition. We simply don't know how general intelligence works well enough to do this.
I misread this as: "our most premium ad services will be advertised as conscience-free." :)
Insight, creativity, and general intelligence are all different and active areas of research.
There is quite a lot of progress in creativity. Google for 'creative software'.
Insight can mean many things, but Watson can now provide insights into cancer diagnosis.
Consciousness is actually fairly well understood in terms of attention, focus, and other aspects.
Artificial general intelligence is a very active field seeing quite a bit of progress.
Here is a bird that flies with flapping wings: https://www.indiegogo.com/projects/bionic-bird-the-flying-ap... (By the way, that is completely unrelated to AI).
Walking over uneven terrain, (also completely unrelated to artificial general intelligence): https://www.youtube.com/watch?v=uVG4J29JZI0 https://www.youtube.com/watch?v=W1czBcnX1Ww
Deep learning is beyond statistical pattern matching. It does involve both supervised and unsupervised learning. Deep learning is currently the most successful technique, but not the most ambitious approach in AGI.
Google for 'AGI', 'deep learning', 'sparse autoencoder', 'hierarchical hidden Markov model', 'OpenCog', 'spiking neural network', 'Hierarchical Temporal Memory'
[...]
3. Better algorithms
To be honest, I find the article a bit missing because of insights like this.