"Contextual adaptation" is what they want, but that doesn't mean it's coming in the near future. However, this does mean that funding for research on it will be available.
As I've said for years, the big lack in AI is in the "common sense" and unstructured manipulation area. Nobody can build something with squirrel levels of manipulation and agility, even in simulation. Robot manipulation in unstructured situations is still very poor. The people trying to simulate C. elegans at the neuron level can't get that to work, despite a full wiring diagram and years of effort.
Something very low level is not understood. There's a Nobel Prize waiting for whomever figures that out.
> Nobody can build something with squirrel levels of manipulation and agility, even in simulation.
Isn’t that what alpha go or the StarCraft league are? Organic strategies in well-defined contexts (action options of the squirrel at Tn)? “Squirrel” is a nice reference frame.
Manipulation in the sense of being able to physically pick things up and manipulate them. Which is a difficult problem. StarCraft is moving a mouse pointer and clicking.
AlphaStar dropped the pretenses of even doing that much because it added too much complexity. For its show matches in Starcraft2, AlphaStar just uses an API to say "Target this specific enemy carrier". It doesn't have to do any vision recognition or differentiate between multiple overlapping units like humans do.
That was the original intent, but it appears the dropped it before public demos.
No - those systems simulate "control" at the level of giving commands, not actual motor control. So practically, current AI can win at Go and play at superhuman levels, but still cannot drive a car as well as most adults.
This made me giggle like a dumbass. Thanks. I participate on the SC2 AI community. It is a-lot of fun. I use Python and CPP for my ML. The game matches are so unpredictable which makes the project questionable because the AI indeed smashes any opponent.
For all of our mastery of information processing, our ability to fabricate and motivate fine differentiated structures is really quite incredibly poor relative to biology. The little beetle is basically alien technology relative to the toy. But add in the fact that it can feed itself, repair itself and reproduce itself with a system that can scale 3-5 orders of magnitude up and down based on need is pretty mindblowing.
The mRNA vaccines might be one of the first times where we've leveraged biology to directly fabricate a discrete part with desired properties. Certainly in raw tonnage of output I can't think of anything that comes close. I feel like this is only going to advance with time, to the point where we may be growing robots instead of milling and printing them.
I find the comparison with AI and insects really interesting to illustrate just how far off the mark we still are.
You have a room sized super computer that trains for zillions of hours and can’t even pick out pedestrians with especially good precision.
Then you have a bee with a minuscule brain that can do complex pattern recognition to find flowers, complex manipulation to get nectar and pollen out of them, object avoidance, swarming behaviour, and fancy communicative dances (both performing and interpreting), plus all the other stuff bees do to get by in the world.
The low-level thing that isn't understood is that intelligence is not progressive. You can not build up to it. Sure, you can forever continue to build better and better approximations of certain aspects of intelligence and pretend you're making progress, but you are not. Of course, there is great utilitarian value in those approximations, so we MUST continue to do this.
Still, my position is that NO progress has ever been made in the area of AI, and no progress will be made any time soon.
I'll take it a step further (in case I don't get enough downvotes for what I've written so far). I maintain that you CAN NOT build intelligence. You can only TAP INTO it. So the very direction in which all of our AI efforts are headed is a dead-end.
I agree. We need to stop this over use, actually the use of AI as some grounded in science matter of fact. What we have is ML, and thus a system for algorithmic refinement and definition.
> I maintain that you CAN NOT build intelligence. You can only TAP INTO it.
I wouldn't go exactly this far, but I would say that whatever process might exist to create artificial intelligence, it might be closer to gardening than to engineering.
My vague feeling is that there might be some sort of (non-supernatural) "mysterious" component to intelligence that we won't be able to engineer and that might just emerge under the right circumstances.
In that case we would just have to "grow" AI, without being completely sure that our effort will work.
"Squirrel level of manipulation" seems more a problem of building tiny physical robots and giving them appropriate interfaces for AI to control, which is an unreasonably difficult benchmark.
A reasonable first step would be autonomous small cheap drones, but I can't say what their missions would be. AlphaStar and Openai Five are mentioned elsewhere here, and these demonstrate that the problem isn't unapproachable.
There probably isn't enough confidence yet to arm autonomous drones, or there isn't a meaningful tactical purpose.
Nice high level overview of the three waves of AI (two have happened, the "Contextual adaptation" wave is yet to occur). Includes examples and successes and failures. ("Young man holding a baseball bat", indeed :) ).
"Systems construct contextual explanatory models for classes of real world phenomena" is the next goal. That is, understanding + being able to describe the reasoning for the understanding.
No technical depth, really, but lots of words to google if you want to learn more.
It's good to see a main funding agency's perspective.
As a researcher, I like their non-hype way of defining AI as "programmed ability", which is accurate and realistic -- also puts AI further apart from real intelligence, which means unanticipated activities.
I would like to know more what they see as "abstracting", from their perspective.
We haven't got much further in our scientific understanding of intelligence - if you bought a psychology text book today and ten years ago there wouldn't be much of a breakthrough change detectable in terms of modeling cognition. And as impressive as some computer science AI models perform certain tasks, I haven't been taken by surprise by them asking me a question out of the blue, which is one of my personal litmus tests for intelligence.
I wonder who is trying to get what funded? "Second wave" is going gangbusters, despite Gary Marcus' every-six-months rant; the review of statistical learning is reasonable given a barely technical audience, but the summary of "third wave" seems designed to extract large amounts of funding from people who aren't up to date on the state of the field.
Theranos got a great deal of funding too, and the truth about how none of the "Second Wave" ML self driving car technology comes even remotely close to safe self driving will be coming out probably later this year. The issue is that ML has no conceptual and causal understanding of anything. For confirmation, I've been carefully watching the countless "self driving" startups in San Francisco driving around, and I have almost never seen a driver in those cars not actively steering it.
None of those self-driving MLs have a virtual world model in their "head", right? They just react to the latest video frame. If so, it's not even a fish level intelligence, it's more like a house plant.
All of the self driving systems I have intimate knowledge of have some level of "persistent memory". It dramatically improves performance in a lot of common situations, e.g. you're coming up to an intersection and your view of stationary cross traffic is occluded by a bus/truck. You enter the intersection and now the planner has to decide whether that vehicle is on an intersection path with the SDC. You only have a few frames of information without persistence and the velocity error might still be large. Thus the safety system has to slam on the breaks to avoid a collision that will never actually happen, potentially causing a real accident if the car behind is following too closely.
If youre curious, Cruise has a fairly comprehensive (but long) overview of their tech that goes over their world model and world simulation stuff: https://youtu.be/uJWN0K26NxQ
You can't always tell whether the car is operating autonomously even from the back seat, let alone another vehicle. A lot of operators keep their hands very close to the wheel and follow its motions to minimize response time when they need to take over.
I've worked at companies whose vehicles you've probably seen and have never heard anything like what you're suggesting is "the truth". Can you provide more details on how current solutions don't remotely approach reasonable safety levels?
> The issue is that ML has no conceptual and causal understanding of anything
Practitioners work with much more tightly defined objectives and methods for improving their systems than "conceptual and causal understanding of anything". From the perspective of folks in the industry we've been making remarkable progress, well beyond what we could have expected. We're saturating major benchmarks, sometimes in one to two years. Back in the day, during the "AI Winter", it was often 10 years before a new breakthrough happened.
I was gonna say this seems really outdated. 2016 was ages ago in the AI world. We've had Openai Five, AlphaStar, MuZero, BERT, GTP and tons of other groundbreaking stuff since then. This presentation tells you nothing about the current state of the field.
I understood clearly enough the first two, but the slides become increasingly ambiguous and fuzzy towards the end; and it seems to me they are mixing up a bunch of not self-evidently related desiderata.
It is not immediate, for instance, that small "generative" models that are easy to interpret necessarily lead to better "abstraction" (whatever that means). And whatever this all has to do with "contextual adaptation" is to me anyone's guess.
Highly alarming (but sadly, from experience, unsurprising) to see such fuzzy position document from such an important funding agency for AI.
This submission should have a (2016)/(2017) on it. I looked at the image caption example and immediately went 'wait, that's wrong, current 2020/2021 captioners would get that right', but of course, for 2017 that's plausible.
Tay had a 'repeat after me' feature (an ancient feature of any IRC chatbot to just echo or print a given string). That's all this is. A troll account issued the command 'Repeat after me' and then tweeted '@TayandYou HITLER DID NOTHING WRONG!', and the Tay daemon dutifully repeated it back to the troll in that thread.
Possibly all of it. As best as I can tell, there may have been a few milquetoast rudenesses, of the usual sort for language models, but the actual quotes everyone cites with detailed statements about the Holocaust or Hitler seem to have all been repeat-after-mes then ripped out of context. It's hard to say given how most of the relevant material has been deleted, and what survives is the usual endless echo chamber of miscitation and simplification and 'everyone knows' which you rapidly become familiar with if you ever try to factcheck anything down to the original sources. (This is why lots of people still 'know' Cambridge Analytica swung the election, or they 'know' the Twitter facecropping algorithm was hugely biased, or that 'Amazon's HR software would only hire you if you played lacrosse', or 'this guy was falsely arrested because face recognition picked him' etc.) If you look at the very earliest reporting, they mostly say it was repeat-after-me functionality, hedging a bit (because who can prove every inflammatory Tay statement was a repeat-after-me?), and then that rapidly gets dropped in favor of narratives about Tay 'learning'. (Even though if you look at the chatbot code MS released later, it's not obvious at all how exactly Tay would 'learn' in the day or so it had before shutdown.)
Anyway, long story short, the Tay incident is either entirely or mostly bogus in the way people want to use it (as an AI safety parable). The real story, of an `echo` gone wrong, is vastly less interesting, and is about as important as typing '8008' into your calculator and showing it to your teacher.
Searching "tay ai" into Google images leads to a handful images that show the initial comment are clearly not "repeat after me" responses. Of those that exist, most of them have been pretty offensive.
I think you are being insufficiently mediate-literate and skeptical and willing to take screenshots at face-value, given that I just demonstrated that a widely-cited example is in fact maliciously and misleadingly edited to remove the context and lie to the viewer about what happened. Did any of the coverage you read mention that? No, they did not. So why do you trust them on everything else and assume they did a good job researching and factchecking when they were so clearly wrong on that one? Why do you believe the others are not repeat-after-mes? Shouldn't the burden of proof be on anyone who claims a specific tweet can be trusted even if those others are bad?
If I go through 'tay ai' (as I have before), I see much the same thing: I see a lot of clearly edited squirrely images, which remove all of the context, and when they appear to include context, the UI looks wrong (some of these are clearly using the 'Replies' tab on the Tay page, instead of being on the actual convo thread; why? to remove the context with the repeat-after-mes or which would show Tay is just spitting out lots of canned generic responses, of course), and like some tweets are being edited out and the remainder spliced together. Do you really think that a 2016-era mass market chatbot (sub-char-RNN in power, often relying on heavily engineered template/script databases) really knows how to do Trump memes complete with clap emoji? What dataset has those? Of course not. It's just a repeat-after-me where the initial tweets were edited out. (What? Someone edit a screenshot, especially where no one can check the original, to score political points? You really think someone would do that - just go on the Internet and tell lies?)
If we look at ones which seem semi-legitimate, like the genocide one, then all we are seeing is typical chatbot evasiveness and generic Eliza-level responses, cherrypicked out of the 100k or so tweets Tay made before they were all deleted. "Do you support $X?" "I do indeed". Wow! Stunning evidence that 'Microsoft turned an advanced AI loose on the Internet and trolls taught it to be evil', as the narrative goes.
So, like I said. All of the long, coherent, strikingly detailed and most offensive ones appear to be either blatantly or probably repeat-after-mes, while the plausible ones are empty chatbot spacefilling cherrypicked out of countless thousands of tweets and made to look more offensive by implying they are the same as the troll-written repeat-after-mes. Neither one supports the broad narrative Tay is used for by OP. And we can note that, beyond the points I've raised here like "how on earth did it 'learn' so quickly" and "there is no evidence it even had 'learning' turned on", there is generally a striking lack of replications of Tay. Which is exactly what you would expect if all the really offensive ones were due to a single ill-considered repeat-after-me feature left enabled, and the rest are just chatbot phaticisms.
You obviously have a lot more skin I the game here and seem to be presenting arguments against points I never suggested or implied.
> then all we are seeing is typical chatbot evasiveness and generic Eliza-level responses
Having found an archive of replies (https://imgur.com/a/iBnbW) "plausible ones are empty chatbot spacefilling cherrypicked out of countless thousands of tweet" seems apt.
In my mind this was the "lesson to be learned" is a HCI one rather than a technical one.
Yes, I would say there are two lessons to be learned from Tay: first, the now-familiar HCI "people are why we can't have nice things" lesson (eg. why you shouldn't do free-response text fields on surveys or polls); and second, about how bad reporting and 'AI safety' discussions tend to be. This is truly basic factchecking and Journalism 101: "so, did this actually happen?" (No.) That the 'Internet racists taught Tay to be evil' legend gets repeated so much and so uncritically, not just by random commenters but by supposed experts in AI and AI safety/bias specifically, is shameful and suggests the areas are epistemically rotten. If they can't understand a simple chatbot echo function, why do you trust them on anything harder?
63 comments
[ 4.6 ms ] story [ 141 ms ] threadAs I've said for years, the big lack in AI is in the "common sense" and unstructured manipulation area. Nobody can build something with squirrel levels of manipulation and agility, even in simulation. Robot manipulation in unstructured situations is still very poor. The people trying to simulate C. elegans at the neuron level can't get that to work, despite a full wiring diagram and years of effort.
Something very low level is not understood. There's a Nobel Prize waiting for whomever figures that out.
Isn’t that what alpha go or the StarCraft league are? Organic strategies in well-defined contexts (action options of the squirrel at Tn)? “Squirrel” is a nice reference frame.
That was the original intent, but it appears the dropped it before public demos.
But if you want it control a robot that's going to bring out a pot of tea while you play, you'll be wiping the floor.
For all of our mastery of information processing, our ability to fabricate and motivate fine differentiated structures is really quite incredibly poor relative to biology. The little beetle is basically alien technology relative to the toy. But add in the fact that it can feed itself, repair itself and reproduce itself with a system that can scale 3-5 orders of magnitude up and down based on need is pretty mindblowing.
The mRNA vaccines might be one of the first times where we've leveraged biology to directly fabricate a discrete part with desired properties. Certainly in raw tonnage of output I can't think of anything that comes close. I feel like this is only going to advance with time, to the point where we may be growing robots instead of milling and printing them.
You have a room sized super computer that trains for zillions of hours and can’t even pick out pedestrians with especially good precision.
Then you have a bee with a minuscule brain that can do complex pattern recognition to find flowers, complex manipulation to get nectar and pollen out of them, object avoidance, swarming behaviour, and fancy communicative dances (both performing and interpreting), plus all the other stuff bees do to get by in the world.
Something is up there for sure.
Still, my position is that NO progress has ever been made in the area of AI, and no progress will be made any time soon.
I'll take it a step further (in case I don't get enough downvotes for what I've written so far). I maintain that you CAN NOT build intelligence. You can only TAP INTO it. So the very direction in which all of our AI efforts are headed is a dead-end.
I wouldn't go exactly this far, but I would say that whatever process might exist to create artificial intelligence, it might be closer to gardening than to engineering.
My vague feeling is that there might be some sort of (non-supernatural) "mysterious" component to intelligence that we won't be able to engineer and that might just emerge under the right circumstances.
In that case we would just have to "grow" AI, without being completely sure that our effort will work.
While not squirrel level, this impressed the heck out of me: https://ashish-kmr.github.io/rma-legged-robots/
A reasonable first step would be autonomous small cheap drones, but I can't say what their missions would be. AlphaStar and Openai Five are mentioned elsewhere here, and these demonstrate that the problem isn't unapproachable.
There probably isn't enough confidence yet to arm autonomous drones, or there isn't a meaningful tactical purpose.
"Systems construct contextual explanatory models for classes of real world phenomena" is the next goal. That is, understanding + being able to describe the reasoning for the understanding.
No technical depth, really, but lots of words to google if you want to learn more.
As a researcher, I like their non-hype way of defining AI as "programmed ability", which is accurate and realistic -- also puts AI further apart from real intelligence, which means unanticipated activities.
I would like to know more what they see as "abstracting", from their perspective.
We haven't got much further in our scientific understanding of intelligence - if you bought a psychology text book today and ten years ago there wouldn't be much of a breakthrough change detectable in terms of modeling cognition. And as impressive as some computer science AI models perform certain tasks, I haven't been taken by surprise by them asking me a question out of the blue, which is one of my personal litmus tests for intelligence.
This is the first point of advice in the Heilmeier Catechism
https://www.darpa.mil/work-with-us/heilmeier-catechism
So no, its a bit more than a house plant
I've worked at companies whose vehicles you've probably seen and have never heard anything like what you're suggesting is "the truth". Can you provide more details on how current solutions don't remotely approach reasonable safety levels?
Practitioners work with much more tightly defined objectives and methods for improving their systems than "conceptual and causal understanding of anything". From the perspective of folks in the industry we've been making remarkable progress, well beyond what we could have expected. We're saturating major benchmarks, sometimes in one to two years. Back in the day, during the "AI Winter", it was often 10 years before a new breakthrough happened.
Also:
> First wave: Handcrafted knowledge
> Second wave: Statistical learning
> Third wave: Contextual adaptation
I understood clearly enough the first two, but the slides become increasingly ambiguous and fuzzy towards the end; and it seems to me they are mixing up a bunch of not self-evidently related desiderata.
It is not immediate, for instance, that small "generative" models that are easy to interpret necessarily lead to better "abstraction" (whatever that means). And whatever this all has to do with "contextual adaptation" is to me anyone's guess.
Highly alarming (but sadly, from experience, unsurprising) to see such fuzzy position document from such an important funding agency for AI.
https://machinelearning.technicacuriosa.com/2017/03/19/a-dar...
- error rate too high
- you can trick a classifier with noise
- it's racist sometimes
Actual dangers of AI:
- stop problem
- infeasibility of sandboxing
- difficulty of aligning black boxes with human values
https://youtu.be/-O01G3tSYpU
Tay had a 'repeat after me' feature (an ancient feature of any IRC chatbot to just echo or print a given string). That's all this is. A troll account issued the command 'Repeat after me' and then tweeted '@TayandYou HITLER DID NOTHING WRONG!', and the Tay daemon dutifully repeated it back to the troll in that thread.
Anyway, long story short, the Tay incident is either entirely or mostly bogus in the way people want to use it (as an AI safety parable). The real story, of an `echo` gone wrong, is vastly less interesting, and is about as important as typing '8008' into your calculator and showing it to your teacher.
If I go through 'tay ai' (as I have before), I see much the same thing: I see a lot of clearly edited squirrely images, which remove all of the context, and when they appear to include context, the UI looks wrong (some of these are clearly using the 'Replies' tab on the Tay page, instead of being on the actual convo thread; why? to remove the context with the repeat-after-mes or which would show Tay is just spitting out lots of canned generic responses, of course), and like some tweets are being edited out and the remainder spliced together. Do you really think that a 2016-era mass market chatbot (sub-char-RNN in power, often relying on heavily engineered template/script databases) really knows how to do Trump memes complete with clap emoji? What dataset has those? Of course not. It's just a repeat-after-me where the initial tweets were edited out. (What? Someone edit a screenshot, especially where no one can check the original, to score political points? You really think someone would do that - just go on the Internet and tell lies?)
If we look at ones which seem semi-legitimate, like the genocide one, then all we are seeing is typical chatbot evasiveness and generic Eliza-level responses, cherrypicked out of the 100k or so tweets Tay made before they were all deleted. "Do you support $X?" "I do indeed". Wow! Stunning evidence that 'Microsoft turned an advanced AI loose on the Internet and trolls taught it to be evil', as the narrative goes.
So, like I said. All of the long, coherent, strikingly detailed and most offensive ones appear to be either blatantly or probably repeat-after-mes, while the plausible ones are empty chatbot spacefilling cherrypicked out of countless thousands of tweets and made to look more offensive by implying they are the same as the troll-written repeat-after-mes. Neither one supports the broad narrative Tay is used for by OP. And we can note that, beyond the points I've raised here like "how on earth did it 'learn' so quickly" and "there is no evidence it even had 'learning' turned on", there is generally a striking lack of replications of Tay. Which is exactly what you would expect if all the really offensive ones were due to a single ill-considered repeat-after-me feature left enabled, and the rest are just chatbot phaticisms.
> then all we are seeing is typical chatbot evasiveness and generic Eliza-level responses
Having found an archive of replies (https://imgur.com/a/iBnbW) "plausible ones are empty chatbot spacefilling cherrypicked out of countless thousands of tweet" seems apt.
In my mind this was the "lesson to be learned" is a HCI one rather than a technical one.