With regards to the specific problems described here, referencing how computers can make sense of communication between a listener and a speaker when each has different knowledge in the world and where strict and literal communication isn't always the most efficient, search for the "Rational Speech Acts (RSA) model" -- an active field in the literature in the past few years.
Yes, the trouble is that the work on performative acts (speech acts) only goes so deep, and then everything turns to confusion and ad hoc rationalizations presented like theories (I have a couple in mind). There'll be no means of implementing this body of knowledge in software -- or even taking it as a guide. So, if it were going to work out, you'd first have to solve the immensely challenging scientific problems, and then see about using the resulting knowledge to build something. As a person with some experience in the area (admittedly nothing published), I don't expect the former to occur. Ever.
Nor is the footstool problem a terribly hard one. The challenging bit is integrating the many and various point-solutions into an adequately coherent assembly.
I hear you. I would say that we are not omniscient, yet we know which footstool is intended in the context, and without any conscious effort. Human-like general intelligence is marvelous. And it's well out of reach, and will remain so.
I opened the link expecting a discussion of flawed metrics that fail to capture uncertainty or failure in replication of claimed studies, or complex architecture turning out to be shallow combinations of simpler ones.
Instead the post attacks a straw man argument about "General intelligence", something that no AI researcher is claiming.
Also the author is clearly clueless as evidenced by:
"Ask a medical researcher to estimate the odds that a Google-funded AI-driven cancer moonshot will succeed, and I expect you will be met with a snicker. '
Google employs several researchers [1] who are well versed in both medicine as well as ML. Who are able to competently judge capabilities and limits of their tools and challenges involved in doing medical research. Rather than creating straw mental images of "medical researcher" "snickering" at "moonshots", the author should to engage in some self reflection before writing about scientific fields he has zero knowledge of.
AND
"I promise you, Facebook did not just work it all out in PHP, and they will not do so in my lifetime or yours."
I won't comment on absurdity of the paragraph about FB or language & meaning. But if any reader was confused, Facebook AI team uses Torch/Lua (for deep learning) or Python (in FBLearnerFlow) and surely not PHP.
Right a moonshot is spectacular, a triumph of PR, a source of wonder and ultimately a really stupid and unbelievably wasteful way to spend a vast fortune. Moonshot really should be the byword of how /not/ to do science funding. Likewise anything beginning "war on..." The war on cancer is going about as well as the war on drugs. Progress is made elsewhere.
Moonshot odds weren't particularly long given the resourcing and desire and this makes it no less impressive as a feat of engineerning and resourcing. And no more sensible.
I'm not convinced this is a general principle like "war on..." Moonshots are intrinsically unlikely, and I suppose we shouldn't expect lots of ancillary benefits (if you get those, it's "fundamental research" instead of a moonshot).
But I don't think there's any requirement that moonshots be heavily financed, or otherwise carry net-negative expectation. They may not be the best social expenditure of funds, but those expenditures (typically attacking acute poverty) aren't realistically competing with moonshots for funding. Throwing a trickle of funding at long-odds projects seems like a worthwhile idea - I'll bet there are 100 different large-upside projects out there with a 1% chance of success.
Now that you mention it, I think the term 'moonshot' does imply the expectation of some kind of spin-offs being discovered along the way. After all, the original moon missions produced a whole range of useful inventions.
I agree that it shouldn't be a pejorative term. The bigger a company is, the closer to basic research they can afford to go in order to create new products and markets. Alphabet is one of the biggest so you'd expect them to be doing pretty much everything down to nuclear physics.
> Alphabet is one of the biggest so you'd expect them to be doing pretty much everything down to nuclear physics.
I've always had the sense that they operate on a strategy of "if it's technological and important, we should be doing it". Self-driving cars and longevity enhancement aren't terribly natural markets for Google, but any successes will be worth a fortune. It seems like "don't miss anything" is the basic plan.
>> Instead the post attacks a straw man argument about "General intelligence", something that no AI researcher is claiming.
Several of the Deep Learning people are on record making silly claims about the future of AI. I'll dig you up the links if you insist [1] but frex, Juergen Schmidhuber has made noises to the effect that deep nets have shown intuition and things like that. Jeoff Hinton has said similar things.
The reason for that is that the original AI project started out to create strong AI in the first place, so if you could not show that your theory has some sort of chance to lead to strong AI people would scoff at it. There's still a bit of that in academia and in industry and parts of the press have not yet gotten the memo that things have changed now, so you'll still hear lots of people making these claims, or having such claims drawn out of them in interviews.
But that's besides the point- which is that a lot of the very clearly stated goals of the industry require the development of systems with human-like intelligence. For instance, natural language processing. We have made some advances in recent years (for instance in things like part of speech tagging or even machine translation to an extent) but the parts of language that we really take for granted as humans, conveying meaning and being aware of some sort of context, those are completely outside our grasp for the time being.
So when Google pretends it can just throw a ton of processing power at, say, machine translation between any two languages any time, and get it to work well, you know that's just its marketing people knowing they can claim anything they like and only their engineering team will be in a position to know how far it is from the truth.
> Juergen Schmidhuber has made noises to the effect that deep nets have shown intuition and things like that.
The only thing wrong in this sentence is your sarcastic tone. Neural nets, deep or not, rely on intuition to produce output, so it is completely true that deep nets show intuition.
What is intuition? It's a feeling that something is of a certain way that is not obvious or immediately obvious/self evident. And translated to ML terms, it is some amount of certainty about something. Intuition is an intrinsic feature of any ML algorithm.
You are mixing feelings with emotions. Sadness is a feeling, but more specifically an emotion. A generic feeling is just amount of fitness to the pattern. The better the network is fit to recognize specific pattern (machine or human alike), it can be said the better "feeling" of the pattern it has.
By the way, emotions won't be that difficult to implement once the feeling part will work.
On a more serious note (and therefore a different comment) you're not even, like, factually correct. What you call "certainty", more commonly known as confidence, which you seem to think is a component of all machine learning algorithms [1], is actually a characteristic of probabilistic classifiers and makes sense primarily in a Bayesian belief network context in any case. If you train an SVM for instance, you would normally not learn much about its degree of confidence to its results (unless you explicitly output probabilities).
Also like I say it's called confidence not "certainty" and it absolutely refers to the degree of confidence of an observer of a given phenomenon, and not some mental state of a system that computes the probabilities by which that observer sets his or her degree of confidence.
Not to mention that in any other context, such terms as "belief", "confidence", and so on are only used metaphorically and - assuming a modicum of reason - only sparingly.
[1] You say that in the context of machine learning intuition is "some amount of certainty" and so "an intrinsic feature of any ML algorithm". Correct me if I misunderstood that.
Well, I'm not native English speaker, but certainty and confidence means pretty similar things, although probably I should have used confidence instead.
And for your criticism about all ML algorithms having degree of confidence - can you give me an example where an algorithm would not have it? And therefore would not have intuition?
ML are made for partial knowledge problems - means that they don't SOLVE a problem, they take prior experience and GUESS the answer. That is intuition by definition.
Here's the kind of claim I mean to attack in the polemic:
"The algorithm's understanding of language 'has moved from a 2-year-old infant to something close to an 8 or 10-year-old child,' said Amit Singhal, a Google Fellow, an honorific reserved for the company's top engineers."
The term refers to work on syntax (or phonology) that does not extend into semantics. So, basically anything Chomsky writes is on narrow syntax, whereas anything (say) Irene Heim or Barbara Partee writes is concerned with a broader notion of syntax (which includes semantics).
Personally, I'd argue the opposite: AI today is super impressive compared to what it was 15 or 30 years ago.
There used to be a time when one could rightfully dismiss new cool things AI with "meh, wake me up when an AI beats a professional Go player on a 19x19 board". Because doing so requires an AI that displays intuition and use of heuristics (Go moves tend to have both local/tactical and global/strategic consequences on the board) rather than one that merely brute forces its way through a decision tree.
The fact that one actually did recently is nothing short of game-changing in my view. It matters little that an AI doesn't demonstrably understand meaning or fully grasp intent. Plus, when one does, we'll arguably have (or not be too far from) a strong AI that, I would gather, won't be too impressed with our own human-level thinking. :-)
I argue that on the go board, it did understand meaning and intent. It had a very nuanced appreciation for go positions, and demonstrated creativity. The critics were shocked, even felt like it had a personality.
19x19 is the normal go board size, I've never heard of anyone playing on a bigger board. Computer players for smaller boards (13x13 and 9x9) have been very good for a while, it's a much easier game on this smaller boards.
I think the parent's point was about whether the AlphaGo AI can generalize to a different board size (as a human player presumably can) rather than about what is the normal Go board size, or what board sizes are easier or harder.
The parent was suggesting that while a human might possess a general understanding of Go, performing comparably regardless of board size ("The human would play on the same level"), the AlphaGo AI in contrast might drop significantly in capability on a board of different size. In effect, the AlphaGo AI might just be a very well-fitted (perhaps even overfitted) machine-learned algorithm on a 19x19 board, lacking a fundamental "understanding" of the game that could generalize to a different board size.
Agreed. Look at speech recognition right up until a few years ago. Until Siri it just didn't work at all, and even Siri wasn't good enough to be useful when it first launched.
Now Google's voice search, and Amazon's Alexa are good enough to be actually useful. If you had asked an AI researcher 10 years ago if that would be the case they would probably have sniggered at you too.
Voice recognition has been okay for a long time, and while Siri/Cortana/OK Google are impressive displays for what they do, I don't really consider it AI. It's heavy keyword indexing combined with speech recognition, and all of the Voice Assistants on phones are still pretty limited in how they will interpret requests. That is, you can't really ask for a value judgement right now from the Assistants simply because they haven't really been written to do much else besides search for specific keywords and perform an action.
What I mean by this is suppose a scenario where you and friends are really hungry on a Friday and you're trying to think of what restaurant would have the shortest wait time. If you were to ask Siri, et. al. about it, unless the restaurant had a publicly accessible wait time that could be tapped for reference, the best the assistants could do is get you a list of open restaurants in yuor area. Your friends, however, probably can, with historical knowledge (times they've been there) and knowledge of current social trends (e.g., asian fusion is out, swedish fusion is in), could make a semi-educated guess as to which restaurant likely has the shortest wait. This is simply because we're recording this data subconsciously - the Assistants aren't because they've not been told to, nor do they even really have an understanding of it. If someone were to write up some parameters for the Assistants to consider it, they could start to do so (and honestly, probably should now that I think about it...).
But even 15ish years ago I was goofing around with voice control for my old boat anchor iMac and was pretty impressed as to how well it handled. It wasn't as fast as what Siri, et. al. do now, but the boat anchor didn't have an entire computing center to offload the processing workload to. There really isn't anything particularly special about what is happening right now with voice assistants except for the speed at which we get the results. It is very cool to be able to do it, but I would't call it AI. At the end of the day, you're still just having your vocalizations converted into a search query and getting the same limits to results you used to get in the past.
And to address the parent comment on Go "...requires an AI that displays intuition and use of heuristics", this is absolutely not in line with how the AlphaGo Researchers were describing it. Lee Sedol even started to get the hang of AlphaGo's prediction model and, if I remember correctly, it really didn't have any idea on what to do with a move that was a "wash". Lee's [1] victory came because of a move that was not evaluated correctly by AlphaGo - the move wasn't bad, in that AlphaGo couldn't capitalize on it, but it also wasn't "good" in that it didn't evaluate it as strong. AlphaGo's weakness was that it really didn't consider history at all when it came to moves, it only operated to rapidly assess the current condition of the board and try to determine the strongest path at the moment. It was a very impressive display to really hammer Lee Sedol like it did, but in interviews, Lee commented that he was starting to understand how AlphaGo made decisions, which was a problem in early games because he had no opponent to read, he had to understand an algorithm making decisions. This reddit thread has some okay breakdowns of how AlphaGo lost and how it decided moves: https://www.reddit.com/r/baduk/comments/4a7wl2/fascinating_i...
The entire point I'm trying to make is that while there are some very impressive Neural Network and Machine Learning tricks going on, there's also a lot going into making a really impressive curtain to hide all the computing behind. Right now, it&#x...
Speech recognition has been ok for a long time if you happened to have an accent that worked fine with it.
I don't. I speak English with a relatively broad Scandinavian accent. Up until a few years ago, no voice recognition system I tested managed to recognise what I said with a high enough precision to be usable.
This "suddenly" changed over the last few years. And it may very well be that the approaches changed (e.g. more heavy emphasis on statistics and large scale feeding of voice data for bad recognition to people to help improve the models; offloading to data centres), but it did change for me at least. E.g. I can now not just have it recognise valid English words, but Youtube will recognise channel names that are mixes of words and letters/numbers, and Amazon Fire will recognise every actor name or movie title I trow at it with a precision rate I would not have expected just a 3-4 years ago.
It's to the point where the Google Docs speech recognition is almost usable for me too.
The fact that voice recognition now understands your accented English is not a sign of its increased ability to understand different accents: only of its increased coverage of specific accents which now include yours.
Not sure if I'm explaining that well enough. What I mean is, say Siri had a 90% accuracy in English before; that would be because it was trained with a lot of samples of English of many different accents. It wasn't trained with enough Scandinavian accents so it didn't understand them very well. Eventually more Scandinavian accented English examples became available and Siri's accuracy with Scandinavian accented English improved as a consequence.
But that doesn't mean that Siri is good at understanding English with any novel accent. So its improvement is not an improvement in terms of quality, only in terms of quantity: it can do more of what it did before, but nothing new.
Which is not to say that speech recognition doesn't work or that small improvements like that are useless, or anything like that, it's just to say there are still limitations and unless you're very careful it's easy to misunderstand what those are.
Ha what? That's just provably untrue. Just find an old Mac and try and ask it the time. A strong memory of mine is me and some friends discovering OSX's voice recognition system and trying to get it to work. It worked about 1/10th of the time, and the 'Tell me a joke' function was a particular challenge because it required you to get two recognitions to work in a row ("Tell me a joke", [knock knock], "Who's there?").
Remember those jokes techie people told about voice-activated photo copiers. That's how implausible it was that there was good voice recognition.
My experience is simply different. I suspect it's something to do with what the other person who replied to me said and that I happen to be from the Midwest of the United States with only a slight accent to give away my origin. The system was likely tuned around speech patterns like my own.
But I'm not going to deny what I had at the time. I could tell my Mac on OS 9 to open "Browser" and it could. I could tell it "Open control panel" and it could. I could ask it the time, or say "schedule", and it could tell me the time or open up my calendar for me.
This is why I'm somewhat dismissive of what we have today. In no way am I suggesting that voice recognition is the same now as it was 15 years ago (nor did I), but the method of operation and results are the same; the voice tools just have a larger vocabularly and thankfully can understand many other languages. But Siri and other Assistants are still pretty much just looking for keywords, just like the happy Mac was on OS 9. Siri sounds more natural than Victoria did, but if I say "Show my me daily schedule" to both Victoria and Siri, they're both going to hear "Schedule" and bring up my calendar.
It is super-impressive, but peoples expectations are also massively higher because more and more people have become accustomed to things that were "crazy" 15-30 years ago.
Even so, to paraphrase William Gibson, the future is not very well distributed.
On one hand people expect miracles and are not very impressed with systems that are incredibly impressive. On the other hand I've been showing some a side-project that mostly just uses a basic Bayesian model that people get impressed by seemingly because it's in an area where they are used to poor results (I'm using it for a very basic text categorisation problem).
People in general appear to be very bad at extrapolating possibilities and mostly judge these things based on what they're used to rather than any assessment of what "ought" to be possible.
This is also true in research in the sense that the moment a problem is solved that previously felt unattainable, it gets somewhat dismissed as "weak AI". The goalposts are constantly moved and the progress to date is underestimated.
My friend did a simple python script that asked "give me a few words, with one being in some ways different than the others", and the script returned the word that was different.
So, for "dog, snake, sheep, tree, owl" it would respond with "tree". For "blue, green, sky, red, black" it would respond with "sky".
It worked by posting a query to google with all the words except the first one, then all the words except the second, etc. And then it compressed the results, and the word that resulted with the smallest compressed result when removed - was the odd one. Very nice use of Kolmogorov complexity estimation.
It still blows people' minds few years later, when I describe what it could do, but then I explain how it does it - and they react with "meh, a simple trick, not an AI".
That's how people react to AI in general - it's only AI if you don't understand it.
I think the point on this story is a bit different than you are interpreting it.
What you're basially describing is the same issue that magic tricks have, where it's fantastic until you know the secret which is often simple.
The difference is that with magic, no (good) magician is claiming they actually have magical powers, and that's why some people get really skeptical and dismissive of AI claims. (Not to say you or your friend are claiming you had strong AI, but more to the modern dsicussion of AI in general)
The problem really is that there are a lot of people commenting on AI who exist in the "know enough to be dangerous" category of technology. They don't fully understand things about computers or programming, but they know enough to sound like they do. This, combined with the fact that extraordinary claims and articles about AI is great publicity for businesses, there is a tendency towards exaggeration when it comes to the idea of AI. For example, I personally wouldn't call what Google, et. al., do as "AI", but as a subset of the field, focusing a lot on speech analysis. This isn't nearly as evocative as saying "AI Research", but it's much more accurate and I think would prompt more intelligent commentary and questions.
When I read your story, I thought it was cool and was curious how you did it; I didn't ascribe AI attributes to it, however, and still found the technique clever as a result. When I hear a claim about "AI" from news currently, I find myself having to distinguish what they're actually talking about first. When there is an inevitable article about AI and people claim that AGI is right around the corner, that is when I get skeptical, and when I think that those who do know a bit about computing get very skeptical and dismissive, since they expected a bit more than more stuff related to just keyword searches. The simple fact is that, like magic tricks, our imagination for AI right now is a lot more impressive than the reality behind current AI research, and a lot of what is presented as AI is more focusing on a specific subset of AI with an impressive front-end to it.
This isn't to dimish the work that the researchers are doing, but clarity of language would be preferable to just the umbrella "AI" category.
I wonder if, as neuroscience research improves, intelligence itself will seem less "magic." As you say, good AI researchers shouldn't claim to have magical powers--as far as we can tell, only these lumps of meat in our head are magical!
I honestly hope so. Not sarcastic or demeaning, I just think that neuroscience is what is necessary for the mythical "AGI", since we don't have a strong notion. I don't believe for a second that it's unknowable, but right now I don't think we know.
I just published technology that some researchers called an 'AI pitch' about 'magic AI does everything' (https://news.ycombinator.com/item?id=12167097#12171264), even while they graciously pointed out some of its optimizations that add up to 'really good tech'. I appreciate all of the constructive and critical feedback, and I'm trying to bring it all together.
Probably the greatest conflict I feel in communicating about such R&D is the challenge of resolving two objectives:
(1) Explaining something technically complex to readers who are not generally interested in "what's under the hood", but still nonetheless appreciate high-level "outreach" about what's actually possible today
(2) Defining the marginal benefits of limited research to a small group of peers who prefer to think only in terms of "what's under the hood", and thus are offended by higher-level connotations that they may associate with hollow marketing rather than rigorous science
Every scientist who communicates to the general public faces the above challenge in some form, and thus struggles to maintain both clarity and integrity all around... One of my inspirations was Christopher Olah (colah.github.io), who I think strikes a wonderful balance of being very careful about jargon, while being very literal.
In any case, a pernicious problem that we face here, and perhaps a premise for the OP's article, is the highly subjective, controversial, and polarizing definitions of artificial intelligence. Too often we forget that what we imagine to be artificial intelligence may feel trivial to another person, and vice versa.
I'd consider this a trick because initially you'd believe the script is what's doing the intelligent part. I think this is an example of impressive AI, but your friend's script isn't the AI, Google is. Imagine using Twilio to text the question to a real human and await their response. The script isn't the "intelligent" entity in either of those cases, just a broker.
To me impressive would be if an AI could be told a story, such as The 3 little pigs and then be able to reason about it speculatively, e.g. answering questions such as "Why do you think this character did this and that?". This would should that the machine has developed an intelligence. So far, all AI can do is pattern recognition, which can be very useful for many tasks, but in my opinion it shouldn't be called "intelligence" because it isn't. Far from it.
I agree about the story. Text generation is already well within our capabilities, but it would be much more "intelligent" if the algorithm understood concepts of plot and characters to come up with a coherent piece of writing. There are already algorithms for summarising sporting and news events, but that's pretty formulaic, not creative like writing a story.
I want to be careful here; yes, semi-coherent text generation is possible under the current state of the art, but there is no evidence right now that it 'on the right track', as it were, to solving the harder problem of binding semantics and language that you allude to; it may be the case that our current techniques for text generation, inverses of our current techniques for text 'understanding' (and I use the term loosely) are just sufficiently powerful to pick up whatever signal is available from the surface-forms of language, but the information is too scarce, the entropy too high for that next big step of 'why'
This is actually one of the things AI is currently being developed for, and may not be that far off. Again, it won't be reasoning as we understand it, more a set of state tables that allow inferences.
But people are definitely trying to develop AI that can take a description of an event, and break it down into legal concepts such as "who was at fault here?"
I agree that that would be impressive, but beating a human being at Go is also very, very impressive, and had been considered a high benchmark for many years. I can't help but think that after a machine can do basic story analysis, there will still be people saying that this is just a trick, that doesn't show any "intelligent" analyses of the story.
AI has a long way to go, but I don't think we should dismiss intelligence as an either/or binary, and call things that are not "intelligence" mere "pattern recognition." Strategy games are widely accepted across cultures as a mark of intelligence when humans do them; you'd be very comfortable saying a human chess or Go grandmaster was intelligent. There was a long time when people were dismissing basic AIs as mere toys by saying they'll never beat a human at Go. There's a little goalpost-moving any time someone responds to an AI breakthrough by saying "You know what would be really impressive..."
The Go victory is definitely impressive, for what it was, but I think this is a pretty unfair comparison; it's an entirely different class of problem, and a strictly less difficult one at that (at least given current knowledge of the mind).
It seems likely to me that strategy games are accepted as an intelligence gauge partly because they're so hard for humans to do; thinking so many plys ahead by exploring and pruning across the game tree seems really hard for humans. For some reason, Natural Language and its semantic binding (I think the notion that these two things are often treated separately is part of a lot of the bigger deltas between expectation and reality when it comes to 'AI'. Honestly, they're probably the same thing anyway) doesn't seem really hard for us.
That doesn't strictly imply that 'strategy games' are strictly harder than 'natural language understanding'; in fact, as far as research goes, strategy games are easy to tackle, because we know all the rules; we designed them! And often, perfect information is available to all agents involved. It's about as black and white as problems get. Sure you can't brute-force Go before the heat death of the universe, but it still theoretically -can- be brute forced. The rules of the game (known and accepted by all parties) encode the algorithm, and it's just a matter of reverse engineering it step-by-step for a given situation. What a brain is doing when it speaks or listens or thinks is currently a much bigger, much more primitive mystery. As far as we know right now, you can't just brute force sentence permutations until you hit a correct answer to a question like the one posed by the parent.
So I don't call this goal-post moving (though acknowledge that the Vanishing Problem is systemic in AI Research and Pop Science); instead, I think this is more akin to the work in say, separating problems into classes of computational complexity. Finding the upper limits of things can be very helpful when we're dealing with such an old, deep, controversial problem.
For sure: as much as Go really was widely-talked about as an AI goal, real language processing has been the goal since, well, Turing. Even more importantly, it remains to be seen how far the current neural network approach extends, and it could still be only successful at certain tasks, with a bigger paradigm shift needed to crack language and so on. But I don't think it's worthwhile to make a hard, fundamental distinction between game-playing and language, such that one requires "real" intelligence and the other is something else.
After all, we possess a single machine that can both play Go and use language, and do far more things besides, and it really does look like the brain is re-using a lot of its structures for these different problems. You can theoretically solve Go, and you can't solve language, but the brain doesn't try to solve either (and neither does Deepmind); there's at least some hope that one paradigm can be applied to both these problems, and the reason we originally turned to neural networks was because of their (admittedly highly superficial) resemblance to our one positive case of intelligence.
So I do take issue with the idea that language understanding is "strictly" harder than board games; that's an artifact of the paradigm used. The idea that the number of rules and exceptions to rules is the 'objective' mark of a task's difficulty only applies to a certain kind of processing agent, and our one example of intelligence clearly uses a different one. For most of the past ten millennia, the only information processing agents on the planet had far more ease communicating with natural language than crushing at board games; ask some now to either write an essay or beat Lee Sedol 4-1, and they'll tell you which is easier. The fact that board game mastery is easier for our AI attempts than basic speech isn't a sign that they've secretly been easier this whole time; for a certain kind of processing agent, communicating in a poorly-defined space has been much easier than mastering a clearly-defined one, and if our AIs don't match that pattern, it's only another reason to wonder if we're on a different path.
The 'kind of processing agent' emphasis is a very good one, and I think worth bringing up in every one of these conversations.
That said, I want to niggle a little bit on some of these points
There certainly does seem to be component re-use in certain parts of the brain between seemingly disparate tasks. The left brain/right brain divide has been largely debunked, and I'm not entirely sure what the current state of the art is, but I'm not sure I'd go so far as to say that that re-use in this particular case (if it exists, I don't know) is anything other than a 'best approximation' (and maybe not even best) with the available hardware. Like using a fork to eat soup. Sure you can kinda do it, but you're always gonna be thinner than the fellow with the spoon.
That suggests to me then that using games as a metric of intelligence is actually a pretty poor way of getting at the deeper issue; there are too many other variables beyond "how much (general/amalgamated) smarts have you got?". It might be semi-convenient when you have cause to believe that the two models you're testing have largely similar structure (i.e., two humans), but it gets pretty noisy when you've got a less than solid idea how similar the two models are. Because there is an obvious, perfect solution to the problem, you end up with a kind of algorithmic continuum, with the system you're trying to model on one end and perfect brute force on the other. And you don't know how far along the continuum your model is; it might be pretty close to the system (the brain), in which case, hey! Progress. But it might be closer to the perfect solution, and thus closer to dedicated hardware for the problem at hand, instead of adapted hardware doing its best. It might have a spork, which makes it rather less worthy of admiration when it comes to soup-eating. Especially when you see it struggle with the steak.
So, in short; I think it actually -might- be a sign that strategy has secretly been easier this whole time. I'm not even sure it's that much of a secret. At least when it comes to the particular processing agents in question, the ones capable of general intelligence as we understand it. If you want to get into philosophical conversations about Hyperintelligent Shades of Blue, well, that's another story.
The article looks at an issue I think about quite often and occasionally have a chance to discuss. The complexity of real-world natural systems is at a magnitude that's hard to grasp and harder to describe. AI, as an attempt to replicate natural human intelligence is up against neural systems that have evolved for 100's of millions of years into numerous forms of astonishingly intricate structure and function.
In humans neuronal brain circuitry comprises an astronomical number of interactive elements, and each connection is itself remarkably adaptable with a large number of interfaces associated with a range of signalling modalities.
Consider that we have ~100 billion brain neurons and 100-500 trillion synaptic interconnections and each connection is modulated by many factors including endocrine and immune system input. Furthermore connections transmit and receive many kinds of signals in complex feedback relationships.
It's well beyond my ability to know how to determine an estimate of how many possible modes of computation can exist within such a marvelous creation, let alone begin to adequately understand how even a small part of it actually works.
I suppose that's the author's point, that what we regard as "simple" problems are in reality not at all simple. Computers can do remarkable things, and surely are among the most useful inventions of human beings. We should realize that working like a significant subset of what our brains routinely do isn't going to be easy for computing to achieve.
We don't need to match 1 artificial neuron for 1 biological neuron. An artificial neural network doesn't have to be implemented in a self-replicating body as our brain does. It doesn't have to carry its biological baggage into the computational aspects. Our brain are part of a process of evolution - the specific requirements for that goal impose huge restrictions on it. On the other hand, in a computer, we can have perfect neurons that don't age or degrade, with any structure we want, scalable to a higher degree. We have already seen in image recognition and go play super-human performance, I expect all domains to be conquered soon. Consider go - they started on this project only 2-3 years ago and already AI has overgrown humans. It doesn't need a lot of time to learn once it is created.
I see it as similar to flight, once we stopped trying to mimic nature exactly (flapping planes) we started to not only make progress, but went beyond what nature can do in flight (eg supersonic)
I think that the biggest breakthrough we'll have in AI will come from a similar break away from copying the brain.
Yes of course you are right that an "artificial neuron" wouldn't (and can't) correspond 1:1 to a biological unit. Main point isn't about aging or degradation because these system characteristics will vary with the nature of the system.
However the computational capability of biological neuronal circuits is not very well understood, but it surely is vast, considering the roles subserved for the purposes of the organism in which it is embedded.
Personally, like the author of the article, I believe it's naive to predict AI will surpass human brains insofar as the functions that define "human". I don't doubt that computing systems can perform many tasks humans can't easily accomplish, e.g., calculate pi to 1000 places.
But I don't regard that as the crux of the question. Determining "meaning" in social interactions, receptive decoding of emotional significance of subtle verbal expression are a couple of examples, but I'll concede they could be unfair.
I've been a keen student of human brain structure and function in relation to behavior for a number of decades, and so little is truly known about most all of it. We don't know how any of the several types of memory work, we are just beginning to sort out the mind-boggling complexities of how critical neurons receive and transmit signals.
Believe me, I have enormous respect for and great affection for the capabilities of our advancing technologies. I'm a real fan, the reason I very much enjoy HN. But in fundamental ways what we do with computers (and the other relevant technologies) is kind of orthogonal to our neurobiological systems. Conflating these can mislead us to our disadvantage because then what we'd call AI will not really replace "NI" (natural intelligence), imagining it will predisposes to disappointing failures.
That's my view of it based on experience, everything has limits, just as you said, including limits on what's humanly knowable, bounded by structural constraints of what we cannot know. Where that boundary lies is indeterminate but not infinite, we keep up the push until slamming into Nature's wall, regroup and turn a different way.
> Consider that we have ~100 billion brain neurons and 100-500 trillion synaptic interconnections and each connection is modulated by many factors including endocrine and immune system input. Furthermore connections transmit and receive many kinds of signals in complex feedback relationships.
This is exactly why estimates of AGI in the next 20 years are laughable. It isn't the neuron that we have to emulate. It is the communication between neurons that we have to emulate.
It is absolutely correct that a 1:1 artificial to natural neuron will not be the case. Current artificial neurons (all deep learning neurons) are nowhere near as complex as a natural neuron. It will more likely be 10-100:1 artificial neurons for each neuron.
Bandwidth to communicate information between neural processing will be the bottleneck (rather than operations per second). Bits per second is multiple orders of magnitude higher than operations per second in human intelligence and our available bandwidth is at least an order of magnitude lower than our operations per second. That means we are at least 10,000 fold farther in terms of raw processing tech than the most optimistic estimates of ~20 years.
10,000 fold. That is 13-14 doubling times or 26-28 years longer than the most optimistic estimates. That is assuming something like Moore's law holds for technology over the next 46-48 years. That is the time until the fastest supercomputers will be able to emulate one human mind assuming we have figured out the algorithms. 40-60 years is a more reasonable time frame for one human mind on one supercomputer. Everyone should have a pretty good feel for the worth of a prediction that will happen 50 years in the future. There will be at least two more AI winters.
A lot of the comments here seem to focus only on the progress we've made rather than what AI should be (if it was true to the name). With "AI", sure I can tell Siri to send a certain email to a certain person. That's pretty damn impressive when I think back to what was possible only a few years earlier. But it's not intelligent. Nowhere near. And honestly, it doesn't save you that much work. It just performs some preprogrammed tasks given some unambiguous input that follows a regular pattern. And even that trips up Siri a ton. Yes, it's cool that you can identify pictures or have a program read your handwriting, but none of that gets us any closer to a program that behaves like a human when I tell it to do something (instead of sending an email or text that I specify, it could draft one on its own; instead of showing me a Wikipedia article when I ask what something is it should explain to me in layman's terms and give me a summary). AI is impressive in terms of progress, but is it really impressive in terms of what it promises to bring? Is it really even progressing in the right direction? Personally, I don't think so.
It's going fast, if you read the papers on arXiv, but if you're checking Siri from time to time, I agree with you that it seems to be stagnating. We're not that far from agents that can much more meaningfully interact in language.
>> We're not that far from agents that can much more meaningfully interact in language.
Really, come on. That's all based on what? A few missives from Google and Facebook, and other companies that try to sell you their super tech?
We don't even know what the distance is from where we are now to "agents that can much more meaningfuly interact in language". Let alone how we're going to cross that distance.
> Yes, it's cool that you can identify pictures or have a program read your handwriting, but none of that gets us any closer to a program that behaves like a human when I tell it to do something
Yes it does. Humans didn't just emerge out of the aether. We evolved from animals that first had to perform simple tasks like that. Our brains are still heavily based on networks of simple pattern recognition neurons. We just don't know how they are structured, or have enough computing power. But it's definitely progress.
As proof, we are starting to get neural networks that can do complex language understanding tasks. RNNs are the state of the art at tasks like predicting the next word, and make interesting chatbots.
Obviously they aren't human level, but they only have 1,000 neurons. Humans have billions! They don't even have episodic memory yet, it's amazing they can do language at all. But research on that is just starting. Give it time.
>> we are starting to get neural networks that can do complex language understanding tasks.
Ah, please, don't say things like that.
We have neural networks that can build complex models of language. But those complex models are nowhere near performing well in "understanding tasks", whatever those are (processing meaning, you mean basically).
The use of ANNs in Natural Language Processing has not led to the same advances it did in, say, image and speech processing, and... and, well, image and speech processing.
There's a ton of things about language that go well beyond building a model of language. Neural networks are nowhere near addressing meaning and context, or the way those are conveyed through language but are not the same as language.
In a way, NLP today, with ANNs or not, is still stuck in the same rut we were ages ago, where we assumed that syntax was the biggest part of language and solving that would solve language [1]. We got syntax down pat and it didn't do us much good. Neural Networks simply build more complex models of language than what we were able to build by hand-crafting grammars thirty years ago- but they're still just complex models of syntax. But syntax is just the surface of language and it's not even all of the surface.
No we're not starting to get nets that can do "complex understanding tasks". Not even ones that can do simple ones. No AI technique does "understanding" yet, except in very limited ways.
[1] Did we really assume that? I'm not even so sure anymore.
In general, language processing has been an important motivator and funding source for AI work. Remember that US Navy Signals Intelligence funded Noam Chomsky's early work on syntax. When early advances in basic linguistic research alongside early advances in computer hardware made machine translation seem plausible, money poured in. And it's pouring in again (now from private industry too).
This is a sound argument, but I think it's operating a level abstract enough to gloss over some of the niggling details that might make the parent's argument valid.
- We don't know that our brains are heavily based on networks of simple pattern recognition neurons. Some parts of them certainly seem to be; especially those parts that seem to correlate to certain tasks that our current ANNs are really good at solving (i.e., basic-to-intermediate image processing). The fundamental mechanism of the brain, if there is one, seems to involve neurons pretty significantly, but we're still not entirely sure what the whole picture is.
- The evolutionary argument is interesting and two-sided; yes it implies that any progress can be 'part of the evolutionary process', so to speak, but it doesn't guarantee that the particular route of the tree being discussed terminates in the desired outcome. It doesn't even guarantee it's on the same tree. It may be the case that image recognition as a task was fundamental to the evolution of high-level human behaviors. It may be the case that it works as a transitionary state, assuming other sets of solved problems unified under the same model, but when treated as foundational, it leads to an evolutionary dead end.
- Following on the 'humans didn't emerge out of the aether' argument from another angle, there's another important consideration to be made. State-of-the-art research is often starting with 'randomly'-wired nets, maybe some lightly-informed initial weights from previous models. It's a kind of tabula rasa, "if you wish to make an apple pie from scratch" picture. The human brain, however, seems to acquire a very specific structure (down to a pretty low level), that is pretty universal amongst, well, humans. So it's not entirely out of the question that some of the things we do aren't strictly learnable from a base, blank initial state; the evolutionary process may well have hardcoded structures pre-wired to know or learn these things. This, of course, is the fundamental principle behind Chomskyian universal grammars.
- I take issue with saying that we are starting to get NNs capable of complex natural language understanding tasks. We really aren't; they are basic-to-intermediate at best, with no real evidence to suggest that they are (A) doing anything interesting with respect to semantics beyond the already well-established distributional semantics strategies, and (B) are not going to dead-end in the near future, like the GOFAI techniques their advocates never tire of feeling superior to.
Aspiring to merely a human clone is a really sad goal in my opinion. Machiens already outshines us on so many tasks, and our lives are much better for it.
Agreed, but the AI industry is mostly built on the hype that computers will some day be at the mental capacity of a human. I think there's a lot of merit in what's been developed (image processing, language processing, handwriting and voice recognition, etc. are all pretty awesome), but at the end of the day we're no closer to the Promised Land of general intelligence. At the moment, it looks to me like the AI industry is growing much faster than the knowledge/technical developments in said field, which is quite simply a waste of manpower that could probably be better applied in other fields.
The last 2-3 years of progress have come about a decade earlier than most people expected. Many people have some catching up to do in determining exactly where we are.
Do we have general intelligence? No, but we have audio generation from pure video and can change an image to match any art style and reasonably good self driving cars.
The site appears to be dead, so haven't had the opportunity to actually read it.
But I recently picked up a "Machine Learning" project for a client. I don't entirely understand everything in the fullest detail, but I understand the basics. Yet despite my lack of knowledge I got something to work 'enough' in a day. Tensorflow makes getting a working DNN easy, and then wrappers like TFLearn make it nearly trivial.
The hardest part of the project was sourcing a decent amount of training data.
And I'm honestly somewhat ashamed of it. My friends think it's cool, but really I don't feel like I did anything. I just glued a library together.
But maybe that's the interesting bit, that frameworks have been developed to a point that even a rube like me can get something working in a day.
I think it's definitely a good thing to abstract away the nuts and bolts of a complicated ML algorithm for general use. Implementing the algorithms from scratch is really useful if you want to learn the details, but most of the time the "standard" implementation will do fine, and will probably perform more efficiently. Of course as with any tool, a conceptual understanding is essential to meaningfully interpret the results though.
I completely get what you're saying about not feeling like you did anything though. Once you understand the algorithms even at a basic level they stop feeling like magic, even though they'll still feel like that for other people.
Reminds me of that article going around HN a few days ago about the ease of AI programming (http://bjenik.com/AIBasedProgramming/) It had a lot of problems, but maybe we really are moving to an era where data analysis consists more of throwing a neural network at it.
There's a clear difference between "bring me an apple" and "bring me the apple". The latter requires that there is either only one apple available, or requires some special contextual knowledge to comply with.
80 comments
[ 2.9 ms ] story [ 128 ms ] threadIntelligence != Omniscience
Instead the post attacks a straw man argument about "General intelligence", something that no AI researcher is claiming.
Also the author is clearly clueless as evidenced by:
"Ask a medical researcher to estimate the odds that a Google-funded AI-driven cancer moonshot will succeed, and I expect you will be met with a snicker. '
Google employs several researchers [1] who are well versed in both medicine as well as ML. Who are able to competently judge capabilities and limits of their tools and challenges involved in doing medical research. Rather than creating straw mental images of "medical researcher" "snickering" at "moonshots", the author should to engage in some self reflection before writing about scientific fields he has zero knowledge of.
AND
"I promise you, Facebook did not just work it all out in PHP, and they will not do so in my lifetime or yours."
I won't comment on absurdity of the paragraph about FB or language & meaning. But if any reader was confused, Facebook AI team uses Torch/Lua (for deep learning) or Python (in FBLearnerFlow) and surely not PHP.
[1] http://www.nimh.nih.gov/about/dr-tom-insel-to-step-down-as-n...
Moonshot odds weren't particularly long given the resourcing and desire and this makes it no less impressive as a feat of engineerning and resourcing. And no more sensible.
But I don't think there's any requirement that moonshots be heavily financed, or otherwise carry net-negative expectation. They may not be the best social expenditure of funds, but those expenditures (typically attacking acute poverty) aren't realistically competing with moonshots for funding. Throwing a trickle of funding at long-odds projects seems like a worthwhile idea - I'll bet there are 100 different large-upside projects out there with a 1% chance of success.
I agree that it shouldn't be a pejorative term. The bigger a company is, the closer to basic research they can afford to go in order to create new products and markets. Alphabet is one of the biggest so you'd expect them to be doing pretty much everything down to nuclear physics.
I've always had the sense that they operate on a strategy of "if it's technological and important, we should be doing it". Self-driving cars and longevity enhancement aren't terribly natural markets for Google, but any successes will be worth a fortune. It seems like "don't miss anything" is the basic plan.
Several of the Deep Learning people are on record making silly claims about the future of AI. I'll dig you up the links if you insist [1] but frex, Juergen Schmidhuber has made noises to the effect that deep nets have shown intuition and things like that. Jeoff Hinton has said similar things.
The reason for that is that the original AI project started out to create strong AI in the first place, so if you could not show that your theory has some sort of chance to lead to strong AI people would scoff at it. There's still a bit of that in academia and in industry and parts of the press have not yet gotten the memo that things have changed now, so you'll still hear lots of people making these claims, or having such claims drawn out of them in interviews.
But that's besides the point- which is that a lot of the very clearly stated goals of the industry require the development of systems with human-like intelligence. For instance, natural language processing. We have made some advances in recent years (for instance in things like part of speech tagging or even machine translation to an extent) but the parts of language that we really take for granted as humans, conveying meaning and being aware of some sort of context, those are completely outside our grasp for the time being.
So when Google pretends it can just throw a ton of processing power at, say, machine translation between any two languages any time, and get it to work well, you know that's just its marketing people knowing they can claim anything they like and only their engineering team will be in a position to know how far it is from the truth.
_______
[1] please don't, I need to stop procrastinating.
The only thing wrong in this sentence is your sarcastic tone. Neural nets, deep or not, rely on intuition to produce output, so it is completely true that deep nets show intuition.
What is intuition? It's a feeling that something is of a certain way that is not obvious or immediately obvious/self evident. And translated to ML terms, it is some amount of certainty about something. Intuition is an intrinsic feature of any ML algorithm.
Do you see the problem with this line of thought? You're now arguing that machine learning algorithms have feelings.
Well, I got feelings too and your assumption I'm being sarcastic makes me really, really sad T_T
By the way, emotions won't be that difficult to implement once the feeling part will work.
This is nonsense and the rest of your comment is nonsense on stilts.
Also like I say it's called confidence not "certainty" and it absolutely refers to the degree of confidence of an observer of a given phenomenon, and not some mental state of a system that computes the probabilities by which that observer sets his or her degree of confidence.
Not to mention that in any other context, such terms as "belief", "confidence", and so on are only used metaphorically and - assuming a modicum of reason - only sparingly.
[1] You say that in the context of machine learning intuition is "some amount of certainty" and so "an intrinsic feature of any ML algorithm". Correct me if I misunderstood that.
And for your criticism about all ML algorithms having degree of confidence - can you give me an example where an algorithm would not have it? And therefore would not have intuition?
ML are made for partial knowledge problems - means that they don't SOLVE a problem, they take prior experience and GUESS the answer. That is intuition by definition.
"The algorithm's understanding of language 'has moved from a 2-year-old infant to something close to an 8 or 10-year-old child,' said Amit Singhal, a Google Fellow, an honorific reserved for the company's top engineers."
Here's what shapes my thinking about medicine, moonshots, and Google: https://www.statnews.com/2016/06/06/google-star-trek-fiction...
About PHP - I am aware that no one would actually try to implement this kind of thing in PHP. I was going for snark. I probably shouldn't have.
There used to be a time when one could rightfully dismiss new cool things AI with "meh, wake me up when an AI beats a professional Go player on a 19x19 board". Because doing so requires an AI that displays intuition and use of heuristics (Go moves tend to have both local/tactical and global/strategic consequences on the board) rather than one that merely brute forces its way through a decision tree.
The fact that one actually did recently is nothing short of game-changing in my view. It matters little that an AI doesn't demonstrably understand meaning or fully grasp intent. Plus, when one does, we'll arguably have (or not be too far from) a strong AI that, I would gather, won't be too impressed with our own human-level thinking. :-)
The parent was suggesting that while a human might possess a general understanding of Go, performing comparably regardless of board size ("The human would play on the same level"), the AlphaGo AI in contrast might drop significantly in capability on a board of different size. In effect, the AlphaGo AI might just be a very well-fitted (perhaps even overfitted) machine-learned algorithm on a 19x19 board, lacking a fundamental "understanding" of the game that could generalize to a different board size.
Now Google's voice search, and Amazon's Alexa are good enough to be actually useful. If you had asked an AI researcher 10 years ago if that would be the case they would probably have sniggered at you too.
Voice recognition has been okay for a long time, and while Siri/Cortana/OK Google are impressive displays for what they do, I don't really consider it AI. It's heavy keyword indexing combined with speech recognition, and all of the Voice Assistants on phones are still pretty limited in how they will interpret requests. That is, you can't really ask for a value judgement right now from the Assistants simply because they haven't really been written to do much else besides search for specific keywords and perform an action.
What I mean by this is suppose a scenario where you and friends are really hungry on a Friday and you're trying to think of what restaurant would have the shortest wait time. If you were to ask Siri, et. al. about it, unless the restaurant had a publicly accessible wait time that could be tapped for reference, the best the assistants could do is get you a list of open restaurants in yuor area. Your friends, however, probably can, with historical knowledge (times they've been there) and knowledge of current social trends (e.g., asian fusion is out, swedish fusion is in), could make a semi-educated guess as to which restaurant likely has the shortest wait. This is simply because we're recording this data subconsciously - the Assistants aren't because they've not been told to, nor do they even really have an understanding of it. If someone were to write up some parameters for the Assistants to consider it, they could start to do so (and honestly, probably should now that I think about it...).
But even 15ish years ago I was goofing around with voice control for my old boat anchor iMac and was pretty impressed as to how well it handled. It wasn't as fast as what Siri, et. al. do now, but the boat anchor didn't have an entire computing center to offload the processing workload to. There really isn't anything particularly special about what is happening right now with voice assistants except for the speed at which we get the results. It is very cool to be able to do it, but I would't call it AI. At the end of the day, you're still just having your vocalizations converted into a search query and getting the same limits to results you used to get in the past.
And to address the parent comment on Go "...requires an AI that displays intuition and use of heuristics", this is absolutely not in line with how the AlphaGo Researchers were describing it. Lee Sedol even started to get the hang of AlphaGo's prediction model and, if I remember correctly, it really didn't have any idea on what to do with a move that was a "wash". Lee's [1] victory came because of a move that was not evaluated correctly by AlphaGo - the move wasn't bad, in that AlphaGo couldn't capitalize on it, but it also wasn't "good" in that it didn't evaluate it as strong. AlphaGo's weakness was that it really didn't consider history at all when it came to moves, it only operated to rapidly assess the current condition of the board and try to determine the strongest path at the moment. It was a very impressive display to really hammer Lee Sedol like it did, but in interviews, Lee commented that he was starting to understand how AlphaGo made decisions, which was a problem in early games because he had no opponent to read, he had to understand an algorithm making decisions. This reddit thread has some okay breakdowns of how AlphaGo lost and how it decided moves: https://www.reddit.com/r/baduk/comments/4a7wl2/fascinating_i...
The entire point I'm trying to make is that while there are some very impressive Neural Network and Machine Learning tricks going on, there's also a lot going into making a really impressive curtain to hide all the computing behind. Right now, it&#x...
I don't. I speak English with a relatively broad Scandinavian accent. Up until a few years ago, no voice recognition system I tested managed to recognise what I said with a high enough precision to be usable.
This "suddenly" changed over the last few years. And it may very well be that the approaches changed (e.g. more heavy emphasis on statistics and large scale feeding of voice data for bad recognition to people to help improve the models; offloading to data centres), but it did change for me at least. E.g. I can now not just have it recognise valid English words, but Youtube will recognise channel names that are mixes of words and letters/numbers, and Amazon Fire will recognise every actor name or movie title I trow at it with a precision rate I would not have expected just a 3-4 years ago.
It's to the point where the Google Docs speech recognition is almost usable for me too.
Not sure if I'm explaining that well enough. What I mean is, say Siri had a 90% accuracy in English before; that would be because it was trained with a lot of samples of English of many different accents. It wasn't trained with enough Scandinavian accents so it didn't understand them very well. Eventually more Scandinavian accented English examples became available and Siri's accuracy with Scandinavian accented English improved as a consequence.
But that doesn't mean that Siri is good at understanding English with any novel accent. So its improvement is not an improvement in terms of quality, only in terms of quantity: it can do more of what it did before, but nothing new.
Which is not to say that speech recognition doesn't work or that small improvements like that are useless, or anything like that, it's just to say there are still limitations and unless you're very careful it's easy to misunderstand what those are.
Ha what? That's just provably untrue. Just find an old Mac and try and ask it the time. A strong memory of mine is me and some friends discovering OSX's voice recognition system and trying to get it to work. It worked about 1/10th of the time, and the 'Tell me a joke' function was a particular challenge because it required you to get two recognitions to work in a row ("Tell me a joke", [knock knock], "Who's there?").
Remember those jokes techie people told about voice-activated photo copiers. That's how implausible it was that there was good voice recognition.
But I'm not going to deny what I had at the time. I could tell my Mac on OS 9 to open "Browser" and it could. I could tell it "Open control panel" and it could. I could ask it the time, or say "schedule", and it could tell me the time or open up my calendar for me.
This is why I'm somewhat dismissive of what we have today. In no way am I suggesting that voice recognition is the same now as it was 15 years ago (nor did I), but the method of operation and results are the same; the voice tools just have a larger vocabularly and thankfully can understand many other languages. But Siri and other Assistants are still pretty much just looking for keywords, just like the happy Mac was on OS 9. Siri sounds more natural than Victoria did, but if I say "Show my me daily schedule" to both Victoria and Siri, they're both going to hear "Schedule" and bring up my calendar.
Even so, to paraphrase William Gibson, the future is not very well distributed.
On one hand people expect miracles and are not very impressed with systems that are incredibly impressive. On the other hand I've been showing some a side-project that mostly just uses a basic Bayesian model that people get impressed by seemingly because it's in an area where they are used to poor results (I'm using it for a very basic text categorisation problem).
People in general appear to be very bad at extrapolating possibilities and mostly judge these things based on what they're used to rather than any assessment of what "ought" to be possible.
So, for "dog, snake, sheep, tree, owl" it would respond with "tree". For "blue, green, sky, red, black" it would respond with "sky".
It worked by posting a query to google with all the words except the first one, then all the words except the second, etc. And then it compressed the results, and the word that resulted with the smallest compressed result when removed - was the odd one. Very nice use of Kolmogorov complexity estimation.
It still blows people' minds few years later, when I describe what it could do, but then I explain how it does it - and they react with "meh, a simple trick, not an AI".
That's how people react to AI in general - it's only AI if you don't understand it.
What you're basially describing is the same issue that magic tricks have, where it's fantastic until you know the secret which is often simple.
The difference is that with magic, no (good) magician is claiming they actually have magical powers, and that's why some people get really skeptical and dismissive of AI claims. (Not to say you or your friend are claiming you had strong AI, but more to the modern dsicussion of AI in general)
The problem really is that there are a lot of people commenting on AI who exist in the "know enough to be dangerous" category of technology. They don't fully understand things about computers or programming, but they know enough to sound like they do. This, combined with the fact that extraordinary claims and articles about AI is great publicity for businesses, there is a tendency towards exaggeration when it comes to the idea of AI. For example, I personally wouldn't call what Google, et. al., do as "AI", but as a subset of the field, focusing a lot on speech analysis. This isn't nearly as evocative as saying "AI Research", but it's much more accurate and I think would prompt more intelligent commentary and questions.
When I read your story, I thought it was cool and was curious how you did it; I didn't ascribe AI attributes to it, however, and still found the technique clever as a result. When I hear a claim about "AI" from news currently, I find myself having to distinguish what they're actually talking about first. When there is an inevitable article about AI and people claim that AGI is right around the corner, that is when I get skeptical, and when I think that those who do know a bit about computing get very skeptical and dismissive, since they expected a bit more than more stuff related to just keyword searches. The simple fact is that, like magic tricks, our imagination for AI right now is a lot more impressive than the reality behind current AI research, and a lot of what is presented as AI is more focusing on a specific subset of AI with an impressive front-end to it.
This isn't to dimish the work that the researchers are doing, but clarity of language would be preferable to just the umbrella "AI" category.
Probably the greatest conflict I feel in communicating about such R&D is the challenge of resolving two objectives:
(1) Explaining something technically complex to readers who are not generally interested in "what's under the hood", but still nonetheless appreciate high-level "outreach" about what's actually possible today
(2) Defining the marginal benefits of limited research to a small group of peers who prefer to think only in terms of "what's under the hood", and thus are offended by higher-level connotations that they may associate with hollow marketing rather than rigorous science
Every scientist who communicates to the general public faces the above challenge in some form, and thus struggles to maintain both clarity and integrity all around... One of my inspirations was Christopher Olah (colah.github.io), who I think strikes a wonderful balance of being very careful about jargon, while being very literal.
In any case, a pernicious problem that we face here, and perhaps a premise for the OP's article, is the highly subjective, controversial, and polarizing definitions of artificial intelligence. Too often we forget that what we imagine to be artificial intelligence may feel trivial to another person, and vice versa.
But people are definitely trying to develop AI that can take a description of an event, and break it down into legal concepts such as "who was at fault here?"
AI has a long way to go, but I don't think we should dismiss intelligence as an either/or binary, and call things that are not "intelligence" mere "pattern recognition." Strategy games are widely accepted across cultures as a mark of intelligence when humans do them; you'd be very comfortable saying a human chess or Go grandmaster was intelligent. There was a long time when people were dismissing basic AIs as mere toys by saying they'll never beat a human at Go. There's a little goalpost-moving any time someone responds to an AI breakthrough by saying "You know what would be really impressive..."
It seems likely to me that strategy games are accepted as an intelligence gauge partly because they're so hard for humans to do; thinking so many plys ahead by exploring and pruning across the game tree seems really hard for humans. For some reason, Natural Language and its semantic binding (I think the notion that these two things are often treated separately is part of a lot of the bigger deltas between expectation and reality when it comes to 'AI'. Honestly, they're probably the same thing anyway) doesn't seem really hard for us.
That doesn't strictly imply that 'strategy games' are strictly harder than 'natural language understanding'; in fact, as far as research goes, strategy games are easy to tackle, because we know all the rules; we designed them! And often, perfect information is available to all agents involved. It's about as black and white as problems get. Sure you can't brute-force Go before the heat death of the universe, but it still theoretically -can- be brute forced. The rules of the game (known and accepted by all parties) encode the algorithm, and it's just a matter of reverse engineering it step-by-step for a given situation. What a brain is doing when it speaks or listens or thinks is currently a much bigger, much more primitive mystery. As far as we know right now, you can't just brute force sentence permutations until you hit a correct answer to a question like the one posed by the parent.
So I don't call this goal-post moving (though acknowledge that the Vanishing Problem is systemic in AI Research and Pop Science); instead, I think this is more akin to the work in say, separating problems into classes of computational complexity. Finding the upper limits of things can be very helpful when we're dealing with such an old, deep, controversial problem.
After all, we possess a single machine that can both play Go and use language, and do far more things besides, and it really does look like the brain is re-using a lot of its structures for these different problems. You can theoretically solve Go, and you can't solve language, but the brain doesn't try to solve either (and neither does Deepmind); there's at least some hope that one paradigm can be applied to both these problems, and the reason we originally turned to neural networks was because of their (admittedly highly superficial) resemblance to our one positive case of intelligence.
So I do take issue with the idea that language understanding is "strictly" harder than board games; that's an artifact of the paradigm used. The idea that the number of rules and exceptions to rules is the 'objective' mark of a task's difficulty only applies to a certain kind of processing agent, and our one example of intelligence clearly uses a different one. For most of the past ten millennia, the only information processing agents on the planet had far more ease communicating with natural language than crushing at board games; ask some now to either write an essay or beat Lee Sedol 4-1, and they'll tell you which is easier. The fact that board game mastery is easier for our AI attempts than basic speech isn't a sign that they've secretly been easier this whole time; for a certain kind of processing agent, communicating in a poorly-defined space has been much easier than mastering a clearly-defined one, and if our AIs don't match that pattern, it's only another reason to wonder if we're on a different path.
That said, I want to niggle a little bit on some of these points
There certainly does seem to be component re-use in certain parts of the brain between seemingly disparate tasks. The left brain/right brain divide has been largely debunked, and I'm not entirely sure what the current state of the art is, but I'm not sure I'd go so far as to say that that re-use in this particular case (if it exists, I don't know) is anything other than a 'best approximation' (and maybe not even best) with the available hardware. Like using a fork to eat soup. Sure you can kinda do it, but you're always gonna be thinner than the fellow with the spoon.
That suggests to me then that using games as a metric of intelligence is actually a pretty poor way of getting at the deeper issue; there are too many other variables beyond "how much (general/amalgamated) smarts have you got?". It might be semi-convenient when you have cause to believe that the two models you're testing have largely similar structure (i.e., two humans), but it gets pretty noisy when you've got a less than solid idea how similar the two models are. Because there is an obvious, perfect solution to the problem, you end up with a kind of algorithmic continuum, with the system you're trying to model on one end and perfect brute force on the other. And you don't know how far along the continuum your model is; it might be pretty close to the system (the brain), in which case, hey! Progress. But it might be closer to the perfect solution, and thus closer to dedicated hardware for the problem at hand, instead of adapted hardware doing its best. It might have a spork, which makes it rather less worthy of admiration when it comes to soup-eating. Especially when you see it struggle with the steak.
So, in short; I think it actually -might- be a sign that strategy has secretly been easier this whole time. I'm not even sure it's that much of a secret. At least when it comes to the particular processing agents in question, the ones capable of general intelligence as we understand it. If you want to get into philosophical conversations about Hyperintelligent Shades of Blue, well, that's another story.
In humans neuronal brain circuitry comprises an astronomical number of interactive elements, and each connection is itself remarkably adaptable with a large number of interfaces associated with a range of signalling modalities.
Consider that we have ~100 billion brain neurons and 100-500 trillion synaptic interconnections and each connection is modulated by many factors including endocrine and immune system input. Furthermore connections transmit and receive many kinds of signals in complex feedback relationships.
It's well beyond my ability to know how to determine an estimate of how many possible modes of computation can exist within such a marvelous creation, let alone begin to adequately understand how even a small part of it actually works.
I suppose that's the author's point, that what we regard as "simple" problems are in reality not at all simple. Computers can do remarkable things, and surely are among the most useful inventions of human beings. We should realize that working like a significant subset of what our brains routinely do isn't going to be easy for computing to achieve.
I think that the biggest breakthrough we'll have in AI will come from a similar break away from copying the brain.
Inspired by nature, not dictated by it.
However the computational capability of biological neuronal circuits is not very well understood, but it surely is vast, considering the roles subserved for the purposes of the organism in which it is embedded.
Personally, like the author of the article, I believe it's naive to predict AI will surpass human brains insofar as the functions that define "human". I don't doubt that computing systems can perform many tasks humans can't easily accomplish, e.g., calculate pi to 1000 places.
But I don't regard that as the crux of the question. Determining "meaning" in social interactions, receptive decoding of emotional significance of subtle verbal expression are a couple of examples, but I'll concede they could be unfair.
I've been a keen student of human brain structure and function in relation to behavior for a number of decades, and so little is truly known about most all of it. We don't know how any of the several types of memory work, we are just beginning to sort out the mind-boggling complexities of how critical neurons receive and transmit signals.
Believe me, I have enormous respect for and great affection for the capabilities of our advancing technologies. I'm a real fan, the reason I very much enjoy HN. But in fundamental ways what we do with computers (and the other relevant technologies) is kind of orthogonal to our neurobiological systems. Conflating these can mislead us to our disadvantage because then what we'd call AI will not really replace "NI" (natural intelligence), imagining it will predisposes to disappointing failures.
That's my view of it based on experience, everything has limits, just as you said, including limits on what's humanly knowable, bounded by structural constraints of what we cannot know. Where that boundary lies is indeterminate but not infinite, we keep up the push until slamming into Nature's wall, regroup and turn a different way.
This is exactly why estimates of AGI in the next 20 years are laughable. It isn't the neuron that we have to emulate. It is the communication between neurons that we have to emulate.
It is absolutely correct that a 1:1 artificial to natural neuron will not be the case. Current artificial neurons (all deep learning neurons) are nowhere near as complex as a natural neuron. It will more likely be 10-100:1 artificial neurons for each neuron.
Bandwidth to communicate information between neural processing will be the bottleneck (rather than operations per second). Bits per second is multiple orders of magnitude higher than operations per second in human intelligence and our available bandwidth is at least an order of magnitude lower than our operations per second. That means we are at least 10,000 fold farther in terms of raw processing tech than the most optimistic estimates of ~20 years.
10,000 fold. That is 13-14 doubling times or 26-28 years longer than the most optimistic estimates. That is assuming something like Moore's law holds for technology over the next 46-48 years. That is the time until the fastest supercomputers will be able to emulate one human mind assuming we have figured out the algorithms. 40-60 years is a more reasonable time frame for one human mind on one supercomputer. Everyone should have a pretty good feel for the worth of a prediction that will happen 50 years in the future. There will be at least two more AI winters.
Really, come on. That's all based on what? A few missives from Google and Facebook, and other companies that try to sell you their super tech?
We don't even know what the distance is from where we are now to "agents that can much more meaningfuly interact in language". Let alone how we're going to cross that distance.
Yes it does. Humans didn't just emerge out of the aether. We evolved from animals that first had to perform simple tasks like that. Our brains are still heavily based on networks of simple pattern recognition neurons. We just don't know how they are structured, or have enough computing power. But it's definitely progress.
As proof, we are starting to get neural networks that can do complex language understanding tasks. RNNs are the state of the art at tasks like predicting the next word, and make interesting chatbots.
Obviously they aren't human level, but they only have 1,000 neurons. Humans have billions! They don't even have episodic memory yet, it's amazing they can do language at all. But research on that is just starting. Give it time.
Ah, please, don't say things like that.
We have neural networks that can build complex models of language. But those complex models are nowhere near performing well in "understanding tasks", whatever those are (processing meaning, you mean basically).
The use of ANNs in Natural Language Processing has not led to the same advances it did in, say, image and speech processing, and... and, well, image and speech processing.
There's a ton of things about language that go well beyond building a model of language. Neural networks are nowhere near addressing meaning and context, or the way those are conveyed through language but are not the same as language.
In a way, NLP today, with ANNs or not, is still stuck in the same rut we were ages ago, where we assumed that syntax was the biggest part of language and solving that would solve language [1]. We got syntax down pat and it didn't do us much good. Neural Networks simply build more complex models of language than what we were able to build by hand-crafting grammars thirty years ago- but they're still just complex models of syntax. But syntax is just the surface of language and it's not even all of the surface.
No we're not starting to get nets that can do "complex understanding tasks". Not even ones that can do simple ones. No AI technique does "understanding" yet, except in very limited ways.
[1] Did we really assume that? I'm not even so sure anymore.
Not so. Academic work in natural language processing has been very active for decades. Check out NL SOAR, which is an awesome attempt from the early 90s by top CS/AI scholars: http://www.cs.cmu.edu/afs/cs/project/soar/utc/nl/doc/nl-soar...
In general, language processing has been an important motivator and funding source for AI work. Remember that US Navy Signals Intelligence funded Noam Chomsky's early work on syntax. When early advances in basic linguistic research alongside early advances in computer hardware made machine translation seem plausible, money poured in. And it's pouring in again (now from private industry too).
- We don't know that our brains are heavily based on networks of simple pattern recognition neurons. Some parts of them certainly seem to be; especially those parts that seem to correlate to certain tasks that our current ANNs are really good at solving (i.e., basic-to-intermediate image processing). The fundamental mechanism of the brain, if there is one, seems to involve neurons pretty significantly, but we're still not entirely sure what the whole picture is.
- The evolutionary argument is interesting and two-sided; yes it implies that any progress can be 'part of the evolutionary process', so to speak, but it doesn't guarantee that the particular route of the tree being discussed terminates in the desired outcome. It doesn't even guarantee it's on the same tree. It may be the case that image recognition as a task was fundamental to the evolution of high-level human behaviors. It may be the case that it works as a transitionary state, assuming other sets of solved problems unified under the same model, but when treated as foundational, it leads to an evolutionary dead end.
- Following on the 'humans didn't emerge out of the aether' argument from another angle, there's another important consideration to be made. State-of-the-art research is often starting with 'randomly'-wired nets, maybe some lightly-informed initial weights from previous models. It's a kind of tabula rasa, "if you wish to make an apple pie from scratch" picture. The human brain, however, seems to acquire a very specific structure (down to a pretty low level), that is pretty universal amongst, well, humans. So it's not entirely out of the question that some of the things we do aren't strictly learnable from a base, blank initial state; the evolutionary process may well have hardcoded structures pre-wired to know or learn these things. This, of course, is the fundamental principle behind Chomskyian universal grammars.
- I take issue with saying that we are starting to get NNs capable of complex natural language understanding tasks. We really aren't; they are basic-to-intermediate at best, with no real evidence to suggest that they are (A) doing anything interesting with respect to semantics beyond the already well-established distributional semantics strategies, and (B) are not going to dead-end in the near future, like the GOFAI techniques their advocates never tire of feeling superior to.
Do we have general intelligence? No, but we have audio generation from pure video and can change an image to match any art style and reasonably good self driving cars.
I missed that one. Link?
https://news.mit.edu/2016/artificial-intelligence-produces-r...
But I recently picked up a "Machine Learning" project for a client. I don't entirely understand everything in the fullest detail, but I understand the basics. Yet despite my lack of knowledge I got something to work 'enough' in a day. Tensorflow makes getting a working DNN easy, and then wrappers like TFLearn make it nearly trivial.
The hardest part of the project was sourcing a decent amount of training data.
And I'm honestly somewhat ashamed of it. My friends think it's cool, but really I don't feel like I did anything. I just glued a library together.
But maybe that's the interesting bit, that frameworks have been developed to a point that even a rube like me can get something working in a day.
I completely get what you're saying about not feeling like you did anything though. Once you understand the algorithms even at a basic level they stop feeling like magic, even though they'll still feel like that for other people.
No one in the scientific literature has a working account of what the English word ‘the’ means.
Aye. But, strong AI is just around the corner. No worries.
"Bring me apple" vs "Bring me that apple".
I still don't understand the need for "a". Surely if you don't qualify a noun at all you mean "any".
I guess "the" is like "distinct" in sql, so it may be useful to catch "too many rows" errors.
It was a joke, BTW.
That's original Russellian theory, in a nutshell (http://www.uvm.edu/~lderosse/courses/lang/Russell(1905).pdf). None of its many variants cover the data.
>> It was a joke
Fair enough!
tay_ai: okay