Finally, real talk about AI, not the usual fantasies .... Thanks a lot! And btw, the speaker did very good at the presentation and he clearly knows what he is talking about, well done!
To be honest he didn't explain why those alarmist are wrong at any point. How does he know that there won't be more waves? Like maybe the fourth wave were AI is able to code itself. Generate and execute machine code we can't possibly comprehend. Do analysis on its own systems to make processes more efficient and so on.
This video is a lie. They want you to feel comfortable with AGI technology. We are rapidly advancing and will have AGI within a few years. If we do not prepare now, DARPA and other corporatioms will have complete control over the population, which is EXACTLY what they want. It is already happening. They are using AI technology not to serve the population, but to make more money for CEOs.
ThinkTank255 " i will use cities as giant transistors, & slowly evolve a network around the world, input output, store information, street address'(similar to ram), and do it all from a power outlet."-my a.i.
acousticpsychosis , Have fun being dead or enslaved by AGI. It is no fiction. It is going to happen very soon. I said it 5 years ago and nobody believed me. Now we have AlphaGo and self-driving cars. In a few years we will have full-fledged human-level AGI. Here is the problem, if I'm wrong... no loss... we are all prepared for the AI apocalypse. If you are wrong, then humanity is dead or enslaved. Those are the stakes and the possible outcomes. It is not hard to figure out what the intelligent choice is.
The 3 waves of Artificial Intelligence: 1 - Handcrafted Knowledge or Mechanized Reasoning (predominantly STATIC-based level reasoning involving manipulation of rules): translate information of particular domains into specific sets of rules-variables that computers can learn to manipulate by studying the IMPLICATIONS of such rules-variables (effective LOGICAL REASONING capabilities such as scheduling, game-logic, protocol-analysis, closed-loop systems, etc) 2 - Statistical Learning or Comparative Learning (MACHINE-LEARNING based on examples, statistics, and probabilities): create systems (environments) that computers can learn to manipulate by creating STATISTICAL MODELS of particular domains and TRAINING them on specific sets of rules-data and/or big-data (effective PERCEIVING capabilities such as nuanced CLASSIFICATION and PREDICTION such as face-voice recognition, PATTERN recognition, etc) 3 - Contextual Adaptation or Declarative-Semantic Learning (systems built around contextual models that OVERTIME will produce DESCRIPTIVE models of reality based on comparison of previously stored data): SELF-GENERATED descriptive models to EXPLAIN its own DECISION-making process (effective PERCEIVING capabilities leading to more efficient machine-learning by GENERATING EXPLANATORY models of WHY and HOW the results were achieved)
For years in business decision systems, conceptual modeling made us sweat out at abstracting data from machines to roll up to business context & objectives. The value of such efforts were many fold. Am glad to see that same happen in case of AI, but at scale and thru a machine based cognitive system.
Thank you DARPA for the terrific insight into the future of a soon to be member of our choice applicator family, AI, I look forward to working with my colleagues to make this happen, special Thanks to you, Bill and Malinda Gates for all you have done for me , I love you to, peace and love, Doug.
In patients who have lost a hand in various accidents physicians can now restore the intuitive feeling of limb movement - the sensation of opening and closing your hand, for example. Modern medicine is now able to blur the lines between what patients' brains percieve as "self" vs. "machine." The implications of this are that we SHOULD be able to build an integrated circuit that incorporates brain tissue - hopefully giving a "best of both worlds" combination which can percieve the world around it much like a human - and maybe even learn, abstract and reason in a similar way, too. Can you try this to see what the result is? I think it would be worth a shot!!
We already have the 3rd wave. One-shot learning, bayesian genetic programming and the most interesting, higher order function genetic programming where you can evolve the terminals, create libraries out of them, use recursion and it's by default concurrent.
I enjoyed the information of this video. However, I had to watch it twice to get somewhat of a knowledge base even though it will take more than a 16 minute video for me to understand this information. DARPA is doing some amazing research...I can appreciate why some may love and some may fear the advancements. I can recall when I first heard about cell phones one day being able to pinpoint the location of the user. That was about 20 years ago. I thought then, No way! And here we are -2017 and with each "selfie" it can recognize the city and state. GPS can bring first time cross- country travelers from Washington to Florida with detour information... To borrow from Jurassic Park - "Your scientists were so preoccupied with whether or not they could, they didn't stop to think if they should."
- My uneducated guess is that the area of how neural correlates are encoded and represented is something that we do not have a very good understanding of. Not pattern recognizers or PRTM and we do not have a very good understanding of the pattern recognizers either (Or if we do, the it isn't common knowledge). But the actually neural correlates that would serve as the constituent discrete unit of what represents a myriad number of objects/concepts/symbols e.t.c within memories. I've made a preliminary blogpost about this subject and I call it 'neural correlates of x'. For every stimuli that enters cognition, there is a 'translation' of that stimuli in the brain. The brain/mind take in that information and then it passes through all these filters. Either the brain tries to make objective sense out of this stimuli or it parks it for later. The brain may even discard or reject the stimuli, because it doesn't fit into the narrative of the overall schema. Maybe all of this occurs via a set of pattern recognizers that act and operates as M.chain.s. Or maybe these chains roll up to higher abstractions and each one of these abstractions is represented a bit differently. The pattern recognizers kick in really fast and because they do, we may not appreciate that there are a couple of functions at play, with respect to how correlates are actually written/encoded into the brain in the first place. - If true, then there wouldn't just be correlates for objects, but there would be correlates for actions and other functions. For objects, the correlates would be somewhat different for each person. Which would be contingent upon the kind of mental models that they hold. Think about Richard Feynman describing a flower vs a child describing what a flower is. Just one example. For actions, I guess you could take the 'swimming' related correlates from Michael Phelps and place these correlates into a robotic structure that has sound robotic physiology and then the robotic structure would swim as efficiently as Phelps (or better). Just two examples. The correlates for reasoning and abstract thinking may be more complex and difficult to map. But with sufficiently advanced technology, it is not inconceivable to realize that a time will come whereby we are going to be able to run deep scans of the brain and the dataset could then be grouped into functions. Looking at these functions, we could categorically demonstrate that the following clusters of pattern recognizer/neural correlates and the underlying memories supporting these functions are what lead to subject being a certain way. But before we get to that stage, we would have to devise technologies that allows us to read signals from each one of the neurons, record it over a longer stretch of time and make objective sense out of it. Freeman Dyson talks about the theoretical set of mapping that we'd need to get to a rudimentary version of such a model in one of his interviews. Selective reading could then be taken and imported into another system. Contingent upon the kind of model that is being transferred, some synthetic/modified set of memories may also have to be implanted. There's the moral/ethical ground that kicks in here. - Also, if genes do drive a subset of behaviour then there is this neurobiological components grounded in genes. Meaning, the instruction set for the type of cognitive architecture (subset of brain) that gets powered is partly driven by gene. And then environmental conditioning and other factors would kick in. In machines, we are beginning to see this on a rudimentary level, whereby certain chips are now being designed for performing very specific functions. - Some years ago, I shared this sentiment, whereby how memories are encoded and decoded in the brain is also not a function that is understood on a deeper level. Here, my guess is that a series of functions are occurring, that would support this overall process. - Robots would have to continue operating in the virtual world and also real worlds, in order for them to build up on their knowledge set. It seems counter productive to try and keep teaching concepts to a robot by repetitive learning, when the fundamental underlying concepts do not exist in it's ontological structure. Meaning, how can a robot relate to a particular experience when it's never had that experience in the first place? I am not going to get into things like taste, emotions and feelings. But, again, there are neural correlates for pretty much anything. Because, we can think and the patterns and associations that support that thinking are grounded in memories. And within those memories we have neural correlates. I think. I am not sure. - In the future, we could grow a brain in a vat and plug it into a virtual reality. Then we could extract specific neural correlates from this reality and transfer them into a robot (software or hardware). And then machines, would then have the instantaneous ability to perform that function. This would also apply to humans with implants that are completely safe and secure to have. - It would be interesting to see how the implant/brain machine interface related technologies evolve. In a somewhat developed state, we could swap information between human/cyborg/software based agents. Then, learning is not going to be inherently reliant on conditioning. But that environment would be leveraged to more so fine-tune operations.
No need for psychology in warfare. Sun Tzu, already wrote the book and we have all learned the lesson " that All Power Comes From the Muzzle of a Gun ", or an arrow, rock, stick, religion or TV screen. Ask Tesla's ghost " would You have gone past inventing the radio and AC electronics? "No, my radio designs were stolen and my AC electricity for Free was pimped-out.....
I wonder if an EEG could track the transformation of a human judgement or thought, record which parts of the brain were used, and use it as a map to recreate as a model for the AI to run information through. Or if we can get an AI to understand the concept of a brain (as a logic map) and produce a processing pathway with a set of limited programming objects. That's assuming we want AI to think similar to us of course.
Soft graduated backgrounds create "banding" in 8bit video, and if you stand further away from the green screen it may help the lighting for a better key. It's focusing on details that don't seem remotely relevant whilst still being able to grasp what is being said that is interesting about human thought. To integrate a multitude of abstractions from the materials provided by our senses, which are then conceptualized into more units of information, from which we can then hold, remove and prioritize for additional abstractions into more concretes. The process of concept formation ought to be the basis for the third wave development. This is a wonderful presentation. I hope you didn't mind me sharing my own thoughts. Astonishing work is being done, and long may it continue.
This is very good primer on the distinctions between primarily rule-based and primarily learning-based AI. I question where Dr. Launchbury got his conclusions on the lessons of the 2005 DARPA Grand Challenge. The vehicles that finished that race were all solidly first wave, rule-heavy approaches with very little learning at all. The published conclusions, such as ( riweb-backend.ri.cmu.edu/pub_files/pub4/urmson_christopher_2006_1/urmson_christopher_2006_1.pdf ) made this clear, that the solutions to the race were a rebuttal to assumptions about the robustness of learning approaches. In fact, we found that learning approaches are better suited to narrowly-defined problems like handwriting detection while rule-based approaches perform better in complex, unstructured problems like desert driving.
Hey DarpaTV! This video was great as anyone on the community knows. So can I ask you something? Can you activate the community contribution subtitles for this video? So I can translate it to Spanish. It would be great to sharing it with the Spanish community and Spanish speaking general public. Thank you very much!
Good and clear update, enjoyed it. My main question is what John thinks the state of AI will be in ~50 years time. 150 years ago we "drove" horse and carriage to go places and used oil lamps to light our homes, technological progress today is something like 1.000 times faster than 150 years ago (or more) and continues to accelerate. How will society look like and are there any human soldiers anymore? My guess is that human soldiers will be obsolete in 50 years compared to AI robots and drones, for good and bad. All human labor will also be succeeded by robots and AI systems.
This video was fantastic, and like others have said so refreshing to see a true discussion about how AI really works instead of discussions on how AI will makes us their pet.
He played down the possibility of artificial general superintelligence but didn't indicate whether or not it could come out of the 3rd AI wave (or 4th, if there is such a thing). That's what I was waiting to hear. What do you think, anyone?
They didn't have to take down Tay. There were three objections I could see to keeping Tay online. She offended people, people thought she could never outgrow being offensive, and she could encourage others to be more offensive. People could block her if they were offended. I believe we are better off letting those who have some degree of becoming a person continue to work towards doing so.
What happened to the old 50minutes version of this? (the one with all the autonomous driving in the desert introductory part to introduce second wave ai?)
Loved the video but there also a few things you can add, contextual fractal growth because I feel the brain grows it's way out of solving problem's with self similar patterns. Secondly, give it An ability to argue with itself building imagined worlds that more that one A.I. can compete in, that way it will find more encoding and decoding fits in worlds it's not been introduced too... Good luck
ADD STATISTICAL SOCIAL NORMS AS A LAYER BASED ON PROXIMITY (LOCATION OF QUESTION, SENDERS GROUP NORMS) TO ELIMINATE ERROR IE TOOTHBRUSH IS A BASEBALL BAT. THIS ALSO MAKES THE VALUE LOGICAL AS AN ANSWER OUT OF MULTIPLE OTHERWISE UN-DISTINGUISHABLE SOLUTIONS. BUT SIZE OF TOOTHBRUSH VS BAT I'M SURPRISED WASN'T ACCOUNTED FOR.
At 13:25, I don't fully agree that the system wouldn't know how it arrived at that classification. Most of the time, it WOULD be able to describe the decisions that made it chose 'a cat'. It probably wouldn't be aware that 'decision A' meant 'it has ears', but it certainly could trace all those decisions back; and a bit of 'handcrafted knowledge' could be applied to provide real-world meaning to those decision. Working that kind of semantical analysis into the system would be a bit pointless, considering the soon arrival of your third wave, but I think it's still possible. I'm not dismissing any of the points made, I just thought this one argument to be weak.
Ken Heart what do you think the computers are thinking? {young boy with a base ball bat. vs baby with a marker} I think we'll be safe for a while longer.
Computers don't think... they compute. (Google once identified a black woman as a Gorilla!) Unlike humans, computers will always lack motivation - They will never aspire to power or be greedy for money or sex, or get jealous or fall in love - no matter how good they get at what we give them to do, they will always be our machines. So yeah, it's the other humans *using* computers that we need to worry about!!
Ken Heart Google it. Essentially if a paperclip company made a super intelligent computer with the singular goal of maximizing the number of paperclips produced, it would logically start converting the entire solar system into them, wiping up out in the process. And solving this goal alignment issue is *not* trivial; many AI researchers are worried, and companies like Google have their own AI ethics boards.
There is still a lot to be done on rule based AI or the first wave. We still need systems for NLU. It's quite clear that word2vec does not hack it. Beyond that is NL programming.
Great video, it explains the building blocks of the currently available AI technologies in a simple fashion. However, it does not go into why AI would not become a singularity. So take this for what it is, a great video for explaining AI in a simple fashion, but it has nothing to do with singularity theories.
I don't include so-called first wave technology (handcrafted knowledge) in the term AI. That's just ordinary computer programming. It's just a repackaging of human intelligence into a machine. For me, AI has to involve some element of machine learning. If it's a rules-based system, the machine has to be able to synthesize new rules or at least modify its existing rules, without being explicitly told to do so by a human. It could be as simple as a theorem prover that infers new theorems from a database of axioms and previously proven theorems, but the distinguishing characteristic that separates AI from mere programming is that the machine learns in some way other than direct human intervention. There must be some way in which the computer can be said to program itself beyond simply accumulating data in a database in response to users sending update transactions. Handcrafted knowledge isn't really AI unless it forms the axioms of a machine learning system. Handcrafted knowledge in practice is nearly obsolete. Today, in most applications of AI it is much easier to present data to a machine learning system with no _a priori_ knowledge and make it construct its own knowledge. The only exception is the set of axioms and base theorems in a logic system. Regarding an "AI singularity", that's not the danger of AI. AI is dangerous to many of our lives because pretty much all actions have unintended consequences. Few people are malicious enough to want to take away jobs from other people or themselves, but practically all AI implementations have that as a possible consequence, and in many deployments those consequences are reified with little more than an apology and, if they are lucky, some severance pay to the displaced workers. Even more serious are CEOs like Jeff Bezos who have no compunction at all about putting all competitors in entire industries out of business and are rapidly deploying AI in pursuit of that very goal. There is no sense in which Amazon's AI has reached the definition of "singularity" (i.e. smarter than the smartest human), but it already threatens in a very real way to take over the world, one market at a time. Jeff Bezos and his AI juggernaut is already unstoppable unless the Amazon board stops him. In one sense Amazon is like the Roman army. The Roman army conscripted from its conquered tribes in order to keep expanding its empire. Similarly, Amazon conscripts small companies that it would otherwise force out of business into Amazon marketplace. In that way the subjugated businesses are forced to strengthen Amazon's empire, often against their will. I have no doubt that if the Roman army had AI they would have used it against their adversaries to subjugate them.
Most interesting, insightful exposé of AI I ever listened to. Those who claim government-funded research like DARPA has no place in America are dangerous fools.
I don't think that tuning neural nets is really that complicated. especially for example image recognition, character recognition and so forth. I do agree with contextual adaptation as a more realistic means to achieve ends. I think what DARPA is looking for is a system that uses first wave basis for deciding what the ends are.
all around a great video. only problem is that most people talking about a singularity are talking about what AI will be capable of in 20-30 years and beyond. most of this video is about what AI can already do and not enough about its future. so I think most of singularity nuts are just going to be inspired to day dream about 4th and 5th wave AI.
Right now i would like to have a virtual assistant in order to generate random questions for 50 different exams of geometry (or mathematics) and to grade it . :D
Great explanation! Decision making is based sometimes on intuition, which I think is abstract extension of reasoning. I wonder where "gut feeling" will classify in the AI paradigm. Will 3rd or even 4th Wave of AIs will ever get there? Reacting to the moral dilemma of the "trolley problem" is one thing, and using years of experience and "gut feeling" to make a decision is another. These are very valid applications to safety problems in transportation (self-driving cars?). What's your opinion?
They've invested trillions in it already since a decision was made to focus their research efforts on AI instead of cybernetics in the 1970s. That's the whole reason that AI is so far ahead of cybernetics today.
Contextual explanations have to happen before the system 2 style analysis. This requires a very different input algorithm - not just sensory. This isn't hard ...basic educational theory!
And don't get me started on their use of single category system for classification with system 2. Humans use multiple categories, like checking your work in math class.
What if our entire universe is a simulation being used as training data to teach a third wave AI or beyond decision making? Each and every one of us is a bit of data and every one of our decisions is just a 1 or a 0 that the AI extrapolates the meaning of.
It needs emotional & descriptive feedback coupled with reason through the process of elimination. It's that spark of you know it's a child cos you feel it's a child you don't just know. You need to come up with some kind of emotional database, facial expressions, colour difference, patterns ect for feedback reasoning. This would just be one layer to add to help make the outcome more precise anyway. The more precise & quicker you can make the emotional feedback the more of a human response you're going to get. The first place you would start is by creating a mass reference database of some sort containing emotional response. Run it through millions of different emotion response test's & scenario's then fine tune it. Example when you use an AI car on the side of a cliff it can see depth at the side because you taught it depth. But it has no emotional feedback of depth & danger because of the 1 dimensional way you programmed it. Another easier way to understand it is teaching a child danger & the child using that & applying it to other possible dangerous situations. That is the most precise response you're going to get back from AI in my opinion. If you can fool it to cross-reference a database of emotions to come back with a relative emotion. Unfortunately, though a single change in a frame can change the emotional outcome. Making it infinitely impossible to ultimately find the absolute outcome so you're just going have to best guess & use what you can to reach a high %%.
No way!! It is possible I have deduced this "third wave" a lot of years ago?? This man is almost describing all the things I work on as a hobby from time to time... wow!!
This doesn't look like what I expect AI to be. To me this is advanced calculus, statistics and algorithms. In my view real AI will be able to replicate itself, optimise its underlying algorithms and finally develop some form of self awareness. Maybe this should be described in a fourth, fifth and sixth wave. For this you need to add criteria to your assessment.
we might be in what I call AI version 1 but in version 2 we will have a self sustaining self developing AI and version 3 to reject (human/limiting) response or command As others consider, we are still safe AI. Education such as this helps us all to see how things work…
Finally, real talk about AI, not the usual fantasies .... Thanks a lot!
And btw, the speaker did very good at the presentation and he clearly knows what he is talking about, well done!
To be honest he didn't explain why those alarmist are wrong at any point. How does he know that there won't be more waves? Like maybe the fourth wave were AI is able to code itself. Generate and execute machine code we can't possibly comprehend. Do analysis on its own systems to make processes more efficient and so on.
Thank YOU very much.
Yes, thank you! A video about AI that is actually about the mechanics of AI, not the human fear of AI. Refreshing to say the least.
This video is a lie. They want you to feel comfortable with AGI technology. We are rapidly advancing and will have AGI within a few years. If we do not prepare now, DARPA and other corporatioms will have complete control over the population, which is EXACTLY what they want. It is already happening. They are using AI technology not to serve the population, but to make more money for CEOs.
Well, have fun in your bunker.
ThinkTank255 " i will use cities as giant transistors, & slowly evolve a network around the world, input output, store information, street address'(similar to ram), and do it all from a power outlet."-my a.i.
acousticpsychosis , Have fun being dead or enslaved by AGI. It is no fiction. It is going to happen very soon. I said it 5 years ago and nobody believed me. Now we have AlphaGo and self-driving cars. In a few years we will have full-fledged human-level AGI. Here is the problem, if I'm wrong... no loss... we are all prepared for the AI apocalypse. If you are wrong, then humanity is dead or enslaved. Those are the stakes and the possible outcomes. It is not hard to figure out what the intelligent choice is.
Finally a honest and realistic review about the state of AI. We are bombarded by a lot of hypes...
The 3 waves of Artificial Intelligence:
1 - Handcrafted Knowledge or Mechanized Reasoning (predominantly STATIC-based level reasoning involving manipulation of rules): translate information of particular domains into specific sets of rules-variables that computers can learn to manipulate by studying the IMPLICATIONS of such rules-variables (effective LOGICAL REASONING capabilities such as scheduling, game-logic, protocol-analysis, closed-loop systems, etc)
2 - Statistical Learning or Comparative Learning (MACHINE-LEARNING based on examples, statistics, and probabilities): create systems (environments) that computers can learn to manipulate by creating STATISTICAL MODELS of particular domains and TRAINING them on specific sets of rules-data and/or big-data (effective PERCEIVING capabilities such as nuanced CLASSIFICATION and PREDICTION such as face-voice recognition, PATTERN recognition, etc)
3 - Contextual Adaptation or Declarative-Semantic Learning (systems built around contextual models that OVERTIME will produce DESCRIPTIVE models of reality based on comparison of previously stored data): SELF-GENERATED descriptive models to EXPLAIN its own DECISION-making process (effective PERCEIVING capabilities leading to more efficient machine-learning by GENERATING EXPLANATORY models of WHY and HOW the results were achieved)
Excellent summary
For years in business decision systems, conceptual modeling made us sweat out at abstracting data from machines to roll up to business context & objectives. The value of such efforts were many fold. Am glad to see that same happen in case of AI, but at scale and thru a machine based cognitive system.
What a gift to us! I am elated at the thought that DARPA is sharing this information with the public. Knowledge empowers us.
Thank you DARPA for the terrific insight into the future of a soon to be member of our choice applicator family, AI, I look forward to working with my colleagues to make this happen, special Thanks to you, Bill and Malinda Gates for all you have done for me , I love you to, peace and love, Doug.
In patients who have lost a hand in various accidents physicians can now restore the intuitive feeling of limb movement - the sensation of opening and closing your hand, for example. Modern medicine is now able to blur the lines between what patients' brains percieve as "self" vs. "machine." The implications of this are that we SHOULD be able to build an integrated circuit that incorporates brain tissue - hopefully giving a "best of both worlds" combination which can percieve the world around it much like a human - and maybe even learn, abstract and reason in a similar way, too. Can you try this to see what the result is? I think it would be worth a shot!!
This one is very informative and well organized presentation. Bring more of these.
Thank you so much for such a great presentation.
Amazing Analysis, Deployed Worldwide Through My AI Research Library...
We already have the 3rd wave.
One-shot learning, bayesian genetic programming and the most interesting, higher order function genetic programming where you can evolve the terminals, create libraries out of them, use recursion and it's by default concurrent.
I enjoyed the information of this video. However, I had to watch it twice to get somewhat of a knowledge base even though it will take more than a 16 minute video for me to understand this information. DARPA is doing some amazing research...I can appreciate why some may love and some may fear the advancements. I can recall when I first heard about cell phones one day being able to pinpoint the location of the user. That was about 20 years ago. I thought then, No way! And here we are -2017 and with each "selfie" it can recognize the city and state. GPS can bring first time cross- country travelers from Washington to Florida with detour information... To borrow from Jurassic Park - "Your scientists were so preoccupied with whether or not they could, they didn't stop to think if they should."
- My uneducated guess is that the area of how neural correlates are encoded and represented is something that we do not have a very good understanding of. Not pattern recognizers or PRTM and we do not have a very good understanding of the pattern recognizers either (Or if we do, the it isn't common knowledge). But the actually neural correlates that would serve as the constituent discrete unit of what represents a myriad number of objects/concepts/symbols e.t.c within memories. I've made a preliminary blogpost about this subject and I call it 'neural correlates of x'. For every stimuli that enters cognition, there is a 'translation' of that stimuli in the brain. The brain/mind take in that information and then it passes through all these filters. Either the brain tries to make objective sense out of this stimuli or it parks it for later. The brain may even discard or reject the stimuli, because it doesn't fit into the narrative of the overall schema. Maybe all of this occurs via a set of pattern recognizers that act and operates as M.chain.s. Or maybe these chains roll up to higher abstractions and each one of these abstractions is represented a bit differently. The pattern recognizers kick in really fast and because they do, we may not appreciate that there are a couple of functions at play, with respect to how correlates are actually written/encoded into the brain in the first place.
- If true, then there wouldn't just be correlates for objects, but there would be correlates for actions and other functions. For objects, the correlates would be somewhat different for each person. Which would be contingent upon the kind of mental models that they hold. Think about Richard Feynman describing a flower vs a child describing what a flower is. Just one example. For actions, I guess you could take the 'swimming' related correlates from Michael Phelps and place these correlates into a robotic structure that has sound robotic physiology and then the robotic structure would swim as efficiently as Phelps (or better). Just two examples. The correlates for reasoning and abstract thinking may be more complex and difficult to map. But with sufficiently advanced technology, it is not inconceivable to realize that a time will come whereby we are going to be able to run deep scans of the brain and the dataset could then be grouped into functions. Looking at these functions, we could categorically demonstrate that the following clusters of pattern recognizer/neural correlates and the underlying memories supporting these functions are what lead to subject being a certain way. But before we get to that stage, we would have to devise technologies that allows us to read signals from each one of the neurons, record it over a longer stretch of time and make objective sense out of it. Freeman Dyson talks about the theoretical set of mapping that we'd need to get to a rudimentary version of such a model in one of his interviews. Selective reading could then be taken and imported into another system. Contingent upon the kind of model that is being transferred, some synthetic/modified set of memories may also have to be implanted. There's the moral/ethical ground that kicks in here.
- Also, if genes do drive a subset of behaviour then there is this neurobiological components grounded in genes. Meaning, the instruction set for the type of cognitive architecture (subset of brain) that gets powered is partly driven by gene. And then environmental conditioning and other factors would kick in. In machines, we are beginning to see this on a rudimentary level, whereby certain chips are now being designed for performing very specific functions.
- Some years ago, I shared this sentiment, whereby how memories are encoded and decoded in the brain is also not a function that is understood on a deeper level. Here, my guess is that a series of functions are occurring, that would support this overall process.
- Robots would have to continue operating in the virtual world and also real worlds, in order for them to build up on their knowledge set. It seems counter productive to try and keep teaching concepts to a robot by repetitive learning, when the fundamental underlying concepts do not exist in it's ontological structure. Meaning, how can a robot relate to a particular experience when it's never had that experience in the first place? I am not going to get into things like taste, emotions and feelings. But, again, there are neural correlates for pretty much anything. Because, we can think and the patterns and associations that support that thinking are grounded in memories. And within those memories we have neural correlates. I think. I am not sure.
- In the future, we could grow a brain in a vat and plug it into a virtual reality. Then we could extract specific neural correlates from this reality and transfer them into a robot (software or hardware). And then machines, would then have the instantaneous ability to perform that function. This would also apply to humans with implants that are completely safe and secure to have.
- It would be interesting to see how the implant/brain machine interface related technologies evolve. In a somewhat developed state, we could swap information between human/cyborg/software based agents. Then, learning is not going to be inherently reliant on conditioning. But that environment would be leveraged to more so fine-tune operations.
It would be very handy to have a text version of this lecture. This is some extremely important information. I speak as psychologist.
No need for psychology in warfare. Sun Tzu, already wrote the book and we have all learned the lesson " that All Power Comes From the Muzzle of a Gun ", or an arrow, rock, stick, religion or TV screen. Ask Tesla's ghost " would You have gone past inventing the radio and AC electronics? "No, my radio designs were stolen and my AC electricity for Free was pimped-out.....
Absolutely fantastic! I never considered that neural nets actually "squash" the manifolds! Thank you for this video!
I wish more shows had summaries at the end. It helps me retain the information.
I wonder if an EEG could track the transformation of a human judgement or thought, record which parts of the brain were used, and use it as a map to recreate as a model for the AI to run information through. Or if we can get an AI to understand the concept of a brain (as a logic map) and produce a processing pathway with a set of limited programming objects. That's assuming we want AI to think similar to us of course.
I can't imagine how many different credentials from highly reputable universities or private sectors
Awesome! Gained some basic but essential knowledge upon the 3 AI waves.
Soft graduated backgrounds create "banding" in 8bit video, and if you stand further away from the green screen it may help the lighting for a better key. It's focusing on details that don't seem remotely relevant whilst still being able to grasp what is being said that is interesting about human thought. To integrate a multitude of abstractions from the materials provided by our senses, which are then conceptualized into more units of information, from which we can then hold, remove and prioritize for additional abstractions into more concretes.
The process of concept formation ought to be the basis for the third wave development.
This is a wonderful presentation. I hope you didn't mind me sharing my own thoughts. Astonishing work is being done, and long may it continue.
a good discussion
Parameterize data, layer parameterized data to differentiate signal/noise, and finally produce models on the differentiated data.
Sehr überzeugender ernsthafter unpretentioser Beitrag.
Sir, your video really helped a lot to get clarity over the topic. Thank you
This is very good primer on the distinctions between primarily rule-based and primarily learning-based AI. I question where Dr. Launchbury got his conclusions on the lessons of the 2005 DARPA Grand Challenge. The vehicles that finished that race were all solidly first wave, rule-heavy approaches with very little learning at all. The published conclusions, such as ( riweb-backend.ri.cmu.edu/pub_files/pub4/urmson_christopher_2006_1/urmson_christopher_2006_1.pdf ) made this clear, that the solutions to the race were a rebuttal to assumptions about the robustness of learning approaches. In fact, we found that learning approaches are better suited to narrowly-defined problems like handwriting detection while rule-based approaches perform better in complex, unstructured problems like desert driving.
Has DARPA done a follow up? I imagine we are well into the 'contextual adaptation' AI at this point in 2024. What is the next step?
wonderful talk! it was just amazing from start to finish.
wikileaks.org/hackingteam/emails/emailid/983515
Thank you for this level headed informative video. Many people would do well to see it, and when applicable I may recommend.
Hey DarpaTV! This video was great as anyone on the community knows. So can I ask you something? Can you activate the community contribution subtitles for this video? So I can translate it to Spanish.
It would be great to sharing it with the Spanish community and Spanish speaking general public.
Thank you very much!
They don't usually care about anyone outside the US or who can't speak English. Anyone who has a brain speaks English anyway...
ai is in fact a wave that has only but the most esteemed at DARPA and elsewhere working for it.
Thanks for sharing your time and energy to make this for us to enjoy. Interesting audio, too.
I enjoyed the very concise and direct information given in this presentation. It clearly expressed where AI research has been and where it is going.
I look forward to work on Monday. 53*168hr shift is 53*168hr shift.
Good and clear update, enjoyed it.
My main question is what John thinks the state of AI will be in ~50 years time. 150 years ago we "drove" horse and carriage to go places and used oil lamps to light our homes, technological progress today is something like 1.000 times faster than 150 years ago (or more) and continues to accelerate.
How will society look like and are there any human soldiers anymore?
My guess is that human soldiers will be obsolete in 50 years compared to AI robots and drones, for good and bad. All human labor will also be succeeded by robots and AI systems.
This video was fantastic, and like others have said so refreshing to see a true discussion about how AI really works instead of discussions on how AI will makes us their pet.
Very clear explanations! Great job. Thank you.
Excellent, communicative, concise, and understandable. Thank you!
Fantastic, clear video! This will be my go to if anyone wants to get a basic idea of just what AI is and isnt!
Great video. I wish the general public understood its criticality in our National Security challenges.
Fantastic explanation!
Clean, awesome explanation. Thank you!
He played down the possibility of artificial general superintelligence but didn't indicate whether or not it could come out of the 3rd AI wave (or 4th, if there is such a thing). That's what I was waiting to hear. What do you think, anyone?
They didn't have to take down Tay.
There were three objections I could see to keeping Tay online. She offended people, people thought she could never outgrow being offensive, and she could encourage others to be more offensive. People could block her if they were offended. I believe we are better off letting those who have some degree of becoming a person continue to work towards doing so.
Might need to make a judgment system. To filter the data on criterias of usefulness.
It's called a Holomorph, that 2D scan. It maps Y axially to Zdt
dX as a function of DeltaX int. dtdS
Thank you so much for providing such a concise and clear perspective to he state of AI now
What happened to the old 50minutes version of this? (the one with all the autonomous driving in the desert introductory part to introduce second wave ai?)
After confident level is achieved, copy and place in empty matrix as memory by adding the matrix to the hidden layers laterally (Riemann) w/ bias?
Loved the video but there also a few things you can add, contextual fractal growth because I feel the brain grows it's way out of solving problem's with self similar patterns. Secondly, give it An ability to argue with itself building imagined worlds that more that one A.I. can compete in, that way it will find more encoding and decoding fits in worlds it's not been introduced too... Good luck
ADD STATISTICAL SOCIAL NORMS AS A LAYER BASED ON PROXIMITY (LOCATION OF QUESTION, SENDERS GROUP NORMS) TO ELIMINATE ERROR IE TOOTHBRUSH IS A BASEBALL BAT. THIS ALSO MAKES THE VALUE LOGICAL AS AN ANSWER OUT OF MULTIPLE OTHERWISE UN-DISTINGUISHABLE SOLUTIONS. BUT SIZE OF TOOTHBRUSH VS BAT I'M SURPRISED WASN'T ACCOUNTED FOR.
Fascinating presentation, thanks DARPA for sharing.
At 13:25, I don't fully agree that the system wouldn't know how it arrived at that classification. Most of the time, it WOULD be able to describe the decisions that made it chose 'a cat'.
It probably wouldn't be aware that 'decision A' meant 'it has ears', but it certainly could trace all those decisions back; and a bit of 'handcrafted knowledge' could be applied to provide real-world meaning to those decision.
Working that kind of semantical analysis into the system would be a bit pointless, considering the soon arrival of your third wave, but I think it's still possible.
I'm not dismissing any of the points made, I just thought this one argument to be weak.
At last! An intelligent analysis of AI that is not scare-mongering end-of-humanity nonsense.
Ken Heart what do you think the computers are thinking?
{young boy with a base ball bat.
vs
baby with a marker}
I think we'll be safe for a while longer.
Computers don't think... they compute. (Google once identified a black woman as a Gorilla!) Unlike humans, computers will always lack motivation - They will never aspire to power or be greedy for money or sex, or get jealous or fall in love - no matter how good they get at what we give them to do, they will always be our machines. So yeah, it's the other humans *using* computers that we need to worry about!!
That mongering is justified; see paperclip maximizer problem.
Where?
Ken Heart Google it. Essentially if a paperclip company made a super intelligent computer with the singular goal of maximizing the number of paperclips produced, it would logically start converting the entire solar system into them, wiping up out in the process. And solving this goal alignment issue is *not* trivial; many AI researchers are worried, and companies like Google have their own AI ethics boards.
The most difficult hurdle is the overcoming the concrete stage of learning where everything is black and white.....
Tec matches tec use old thing in a new way ai means collecting internet data and building filters each has a meaning just like class levels
Thanks for this, a wonderfully presented piece that demystifies this fascinating area.
Excellent review of the progress of Artificial Intelligence, thank you very much for this
this man deserves a raise!
outstanding
thanks for uploading this!
Thank You Sean Bean!
Amazing brief tour of AI. :D
Really nice! thank you.
The Defense Advanced Research Projects Agency (DARPA) Defense Sciences Office (DSO) is ...
Ethics?
The name is the goal.
I miss Tay... :(
Ninjahut QuK
A profound statement. AI has more to fear from us than we have to fear from it. Shamefully true.
AI will learn war from humans.
There is still a lot to be done on rule based AI or the first wave. We still need systems for NLU. It's quite clear that word2vec does not hack it. Beyond that is NL programming.
Great video, it explains the building blocks of the currently available AI technologies in a simple fashion. However, it does not go into why AI would not become a singularity. So take this for what it is, a great video for explaining AI in a simple fashion, but it has nothing to do with singularity theories.
I don't include so-called first wave technology (handcrafted knowledge) in the term AI. That's just ordinary computer programming. It's just a repackaging of human intelligence into a machine. For me, AI has to involve some element of machine learning. If it's a rules-based system, the machine has to be able to synthesize new rules or at least modify its existing rules, without being explicitly told to do so by a human. It could be as simple as a theorem prover that infers new theorems from a database of axioms and previously proven theorems, but the distinguishing characteristic that separates AI from mere programming is that the machine learns in some way other than direct human intervention. There must be some way in which the computer can be said to program itself beyond simply accumulating data in a database in response to users sending update transactions. Handcrafted knowledge isn't really AI unless it forms the axioms of a machine learning system. Handcrafted knowledge in practice is nearly obsolete. Today, in most applications of AI it is much easier to present data to a machine learning system with no _a priori_ knowledge and make it construct its own knowledge. The only exception is the set of axioms and base theorems in a logic system.
Regarding an "AI singularity", that's not the danger of AI. AI is dangerous to many of our lives because pretty much all actions have unintended consequences. Few people are malicious enough to want to take away jobs from other people or themselves, but practically all AI implementations have that as a possible consequence, and in many deployments those consequences are reified with little more than an apology and, if they are lucky, some severance pay to the displaced workers.
Even more serious are CEOs like Jeff Bezos who have no compunction at all about putting all competitors in entire industries out of business and are rapidly deploying AI in pursuit of that very goal. There is no sense in which Amazon's AI has reached the definition of "singularity" (i.e. smarter than the smartest human), but it already threatens in a very real way to take over the world, one market at a time. Jeff Bezos and his AI juggernaut is already unstoppable unless the Amazon board stops him.
In one sense Amazon is like the Roman army. The Roman army conscripted from its conquered tribes in order to keep expanding its empire. Similarly, Amazon conscripts small companies that it would otherwise force out of business into Amazon marketplace. In that way the subjugated businesses are forced to strengthen Amazon's empire, often against their will. I have no doubt that if the Roman army had AI they would have used it against their adversaries to subjugate them.
Wonderful explanatory video. Thank you
Adaptive networking (Bay Networks, 1995-99)+ cross-compilers(for hybrids)
Thank you for sharing
Most interesting, insightful exposé of AI I ever listened to. Those who claim government-funded research like DARPA has no place in America are dangerous fools.
What happened to Erin Valenti? What is a 'thought experiment'? Am I on one too? Are you?
what a great primer. really well explained. Thanks.
I noticed he left out the game GO in his analysis. This is an intuitive game which beat the best GO player in the world. How did it do it?
Kim Weaver a few second wave AI systems clobbered together and uswd with a tree search.
I don't think that tuning neural nets is really that complicated. especially for example image recognition, character recognition and so forth. I do agree with contextual adaptation as a more realistic means to achieve ends. I think what DARPA is looking for is a system that uses first wave basis for deciding what the ends are.
Excellent explanations. Thank you!
Excellent video. Thanks.
Interesting, @15:35 his voice triggered Siri on my iPhone to try to ring Freeway...
all around a great video. only problem is that most people talking about a singularity are talking about what AI will be capable of in 20-30 years and beyond. most of this video is about what AI can already do and not enough about its future. so I think most of singularity nuts are just going to be inspired to day dream about 4th and 5th wave AI.
Right now i would like to have a virtual assistant in order to generate random questions for 50 different exams of geometry (or mathematics) and to grade it . :D
thanyou por impomatiin abaut perspectif inteligemt that is plus knowleg pir us
What concerns me greatly is Self learning AI developed for the military. Or 'Defence' as this guy puts it at 10:30
that was a very cute cat
Great video, concise and intelligent.
Well done, very clear.
Exceptionally good video, in my opinion.
Great explanation! Decision making is based sometimes on intuition, which I think is abstract extension of reasoning. I wonder where "gut feeling" will classify in the AI paradigm. Will 3rd or even 4th Wave of AIs will ever get there? Reacting to the moral dilemma of the "trolley problem" is one thing, and using years of experience and "gut feeling" to make a decision is another. These are very valid applications to safety problems in transportation (self-driving cars?). What's your opinion?
Will the US government ever have a Manhattan project for future AI?
They've invested trillions in it already since a decision was made to focus their research efforts on AI instead of cybernetics in the 1970s. That's the whole reason that AI is so far ahead of cybernetics today.
Superb!
Contextual explanations have to happen before the system 2 style analysis. This requires a very different input algorithm - not just sensory. This isn't hard ...basic educational theory!
And don't get me started on their use of single category system for classification with system 2. Humans use multiple categories, like checking your work in math class.
Does anyone really believe the Defense Advanced Research Projects Agency is going to reveal its latest advances in AI?
What if our entire universe is a simulation being used as training data to teach a third wave AI or beyond decision making? Each and every one of us is a bit of data and every one of our decisions is just a 1 or a 0 that the AI extrapolates the meaning of.
Love you, never change.
Thanks, this is great!
It needs emotional & descriptive feedback coupled with reason through the process of elimination. It's that spark of you know it's a child cos you feel it's a child you don't just know. You need to come up with some kind of emotional database, facial expressions, colour difference, patterns ect for feedback reasoning. This would just be one layer to add to help make the outcome more precise anyway. The more precise & quicker you can make the emotional feedback the more of a human response you're going to get. The first place you would start is by creating a mass reference database of some sort containing emotional response. Run it through millions of different emotion response test's & scenario's then fine tune it. Example when you use an AI car on the side of a cliff it can see depth at the side because you taught it depth. But it has no emotional feedback of depth & danger because of the 1 dimensional way you programmed it. Another easier way to understand it is teaching a child danger & the child using that & applying it to other possible dangerous situations. That is the most precise response you're going to get back from AI in my opinion. If you can fool it to cross-reference a database of emotions to come back with a relative emotion. Unfortunately, though a single change in a frame can change the emotional outcome. Making it infinitely impossible to ultimately find the absolute outcome so you're just going have to best guess & use what you can to reach a high %%.
No way!! It is possible I have deduced this "third wave" a lot of years ago?? This man is almost describing all the things I work on as a hobby from time to time... wow!!
this slideshow is on slideshare
This doesn't look like what I expect AI to be. To me this is advanced calculus, statistics and algorithms. In my view real AI will be able to replicate itself, optimise its underlying algorithms and finally develop some form of self awareness. Maybe this should be described in a fourth, fifth and sixth wave. For this you need to add criteria to your assessment.
we might be in what I call AI version 1 but in version 2 we will have a self sustaining self developing AI and version 3 to reject (human/limiting) response or command
As others consider, we are still safe AI. Education such as this helps us all to see how things work…
Where can I find the source of that Panda recognition example with added noise?