title
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11

description

detail
{'title': 'Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11', 'heatmap': [{'end': 1107.957, 'start': 1051.822, 'weight': 0.845}, {'end': 1262.896, 'start': 1187.39, 'weight': 0.714}, {'end': 1871.68, 'start': 1821.06, 'weight': 0.811}, {'end': 2639.742, 'start': 2589.369, 'weight': 0.763}, {'end': 2790.659, 'start': 2733.621, 'weight': 0.719}], 'summary': "Juergen schmidhuber's work in ai, meta-learning, and transfer learning is explored, with a focus on building a machine for recursive self-improvement. the podcast also discusses practical problem solving, determinism versus randomness, creativity and curiosity in problem solving, depth in neural networks, the future of ai and reinforcement learning, and ai's potential impact on the workforce and universe conquest.", 'chapters': [{'end': 623.518, 'segs': [{'end': 31.284, 'src': 'embed', 'start': 0.089, 'weight': 1, 'content': [{'end': 2.872, 'text': 'The following is a conversation with Juergen Schmidhuber.', 'start': 0.089, 'duration': 2.783}, {'end': 9.398, 'text': "He's the co-director of ATSIA Swiss AI Lab and a co-creator of long short-term memory networks.", 'start': 3.612, 'duration': 5.786}, {'end': 16.665, 'text': 'LSDMs are used in billions of devices today for speech recognition, translation, and much more.', 'start': 10.579, 'duration': 6.086}, {'end': 24.916, 'text': 'Over 30 years he has proposed a lot of interesting out-of-the-box ideas on meta-learning, adversarial networks,', 'start': 17.506, 'duration': 7.41}, {'end': 31.284, 'text': 'computer vision and even a formal theory of quote, creativity, curiosity and fun.', 'start': 24.916, 'duration': 6.368}], 'summary': 'Juergen schmidhuber, co-director of atsia swiss ai lab, co-creator of lstm networks, proposed groundbreaking ideas in ai for over 30 years, impacting billions of devices for speech recognition, translation, and more.', 'duration': 31.195, 'max_score': 0.089, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo89.jpg'}, {'end': 242.531, 'src': 'embed', 'start': 217.497, 'weight': 0, 'content': [{'end': 224.26, 'text': 'but you also improve the way the machine improves and you also improve the way it improves, the way it improves itself.', 'start': 217.497, 'duration': 6.763}, {'end': 237.288, 'text': 'And that was my 1987 diploma thesis, which was all about that hierarchy of meta-learners that have no computational limits,', 'start': 225.782, 'duration': 11.506}, {'end': 242.531, 'text': 'except for the well-known limits that Gödel identified in 1931 and for the limits of physics.', 'start': 237.288, 'duration': 5.243}], 'summary': "1987 thesis explores meta-learners with no computational limits, referencing gödel's 1931 limits and physics.", 'duration': 25.034, 'max_score': 217.497, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo217497.jpg'}, {'end': 302.529, 'src': 'embed', 'start': 270.905, 'weight': 2, 'content': [{'end': 276.267, 'text': "Let's take the example of a deep neural network that has learned to classify images.", 'start': 270.905, 'duration': 5.362}, {'end': 283.47, 'text': 'And maybe you have trained that network on 100 different databases of images.', 'start': 277.328, 'duration': 6.142}, {'end': 291.594, 'text': 'And now a new database comes along and you want to quickly learn the new thing as well.', 'start': 285.891, 'duration': 5.703}, {'end': 302.529, 'text': 'So one simple way of doing that is you take the network, which already knows 100 types of databases,', 'start': 293.447, 'duration': 9.082}], 'summary': 'A deep neural network trained on 100 image databases learns new data.', 'duration': 31.624, 'max_score': 270.905, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo270905.jpg'}, {'end': 395.093, 'src': 'embed', 'start': 364.902, 'weight': 3, 'content': [{'end': 382.885, 'text': 'such that the learning system has an opportunity to modify any part of the learning algorithm and then evaluate the consequences of that modification and then learn from that to create a better learning algorithm,', 'start': 364.902, 'duration': 17.983}, {'end': 384.506, 'text': 'and so on, recursively.', 'start': 382.885, 'duration': 1.621}, {'end': 395.093, 'text': "So, that's a very different animal where you are opening the space of possible learning algorithms to the learning system itself.", 'start': 385.766, 'duration': 9.327}], 'summary': 'The learning system can modify its algorithm and learn recursively to create better algorithms.', 'duration': 30.191, 'max_score': 364.902, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo364902.jpg'}, {'end': 526.093, 'src': 'embed', 'start': 498.695, 'weight': 4, 'content': [{'end': 505.457, 'text': "which aren't provably optimal at all, like the other stuff that we did, but which are much more practical,", 'start': 498.695, 'duration': 6.762}, {'end': 514.784, 'text': 'as long as we only want to solve the small problems that we are Typically trying to solve in this environment here.', 'start': 505.457, 'duration': 9.327}, {'end': 524.412, 'text': "So the universal problem solvers like the Gödel machine, but also Markus Hutter's fastest way of solving all possible problems,", 'start': 515.624, 'duration': 8.788}, {'end': 526.093, 'text': 'which he developed around 2002 in my lab.', 'start': 524.412, 'duration': 1.681}], 'summary': "Universal problem solvers, like the gödel machine and markus hutter's fastest way, are more practical for solving small problems.", 'duration': 27.398, 'max_score': 498.695, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo498695.jpg'}], 'start': 0.089, 'title': 'Ai pioneer and meta-learning', 'summary': "Delves into juergen schmidhuber's pioneering work in ai, particularly in meta-learning, and transfer learning, exploring his vision of building a machine that can recursively self-improve to become a general problem solver. it also discusses practical examples and theoretical comparisons of transfer learning and meta-learning.", 'chapters': [{'end': 270.588, 'start': 0.089, 'title': 'Juergen schmidhuber: ai pioneer and meta-learning expert', 'summary': "Explores juergen schmidhuber's pioneering work in ai, particularly in the field of meta-learning, and his vision of building a machine that can recursively self-improve to become a general problem solver, with a focus on his concept of meta-learning and recursive self-improvement.", 'duration': 270.499, 'highlights': ['Juergen Schmidhuber proposed the concept of meta-learning and recursive self-improvement in the 80s, aiming to build a machine that could learn to improve its learning algorithm and itself, ultimately solving all solvable problems.', 'His 1987 diploma thesis focused on the hierarchy of meta-learners without computational limits, except for the well-known limits identified by Gödel in 1931 and the limits of physics.', 'LSTM networks, co-created by Schmidhuber, are widely used today in billions of devices for speech recognition, translation, and more, showcasing the practical impact of his work in AI.', "The conversation provides insights into Schmidhuber's early inspiration for pursuing AI, rooted in his desire to build a machine that could become a better physicist than himself and multiply his creativity infinitely to understand the universe."]}, {'end': 623.518, 'start': 270.905, 'title': 'Transfer learning and meta-learning', 'summary': 'Discusses transfer learning where a neural network can quickly learn from a new dataset by retraining its top layer using previous knowledge, and meta-learning which allows the learning algorithm to modify itself for better learning, with practical examples and theoretical comparisons.', 'duration': 352.613, 'highlights': ['Transfer learning allows a neural network trained on 100 different databases to quickly learn from a new dataset by retraining its top layer, leveraging previous knowledge. trained on 100 different databases', 'Meta-learning enables the learning algorithm to modify itself and evaluate the consequences of the modification to create a better learning algorithm, offering a different approach to traditional learning systems. modify any part of the learning algorithm and evaluate the consequences of that modification', "Universal problem solvers like the Gödel machine and Markus Hutter's fastest way of solving all possible problems come with constant overheads for proof search, guaranteeing optimality, while practical techniques like recurrent neural networks are more suitable for solving everyday small problems. Gödel machine, Markus Hutter's fastest way of solving all possible problems, and recurrent neural networks"]}], 'duration': 623.429, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo89.jpg', 'highlights': ['Juergen Schmidhuber proposed the concept of meta-learning and recursive self-improvement in the 80s, aiming to build a machine that could learn to improve its learning algorithm and itself, ultimately solving all solvable problems.', 'LSTM networks, co-created by Schmidhuber, are widely used today in billions of devices for speech recognition, translation, and more, showcasing the practical impact of his work in AI.', 'Transfer learning allows a neural network trained on 100 different databases to quickly learn from a new dataset by retraining its top layer, leveraging previous knowledge.', 'Meta-learning enables the learning algorithm to modify itself and evaluate the consequences of the modification to create a better learning algorithm, offering a different approach to traditional learning systems.', "Universal problem solvers like the Gödel machine and Markus Hutter's fastest way of solving all possible problems come with constant overheads for proof search, guaranteeing optimality, while practical techniques like recurrent neural networks are more suitable for solving everyday small problems."]}, {'end': 1226.832, 'segs': [{'end': 656.089, 'src': 'embed', 'start': 624.459, 'weight': 2, 'content': [{'end': 632.487, 'text': "And this additive constant doesn't care for n, which means as n is getting larger and larger.", 'start': 624.459, 'duration': 8.028}, {'end': 638.332, 'text': 'as you have more and more cities, the constant overhead pales in comparison.', 'start': 632.487, 'duration': 5.845}, {'end': 646.338, 'text': 'And that means that almost all large problems are solved in the best possible way already today.', 'start': 638.792, 'duration': 7.546}, {'end': 649.661, 'text': 'We already have a universal problem solver like that.', 'start': 646.358, 'duration': 3.303}, {'end': 656.089, 'text': "However, it's not practical, because the overhead the constant overhead,", 'start': 650.643, 'duration': 5.446}], 'summary': 'As n increases, constant overhead becomes negligible, solving almost all large problems optimally today.', 'duration': 31.63, 'max_score': 624.459, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo624459.jpg'}, {'end': 757.916, 'src': 'embed', 'start': 727.852, 'weight': 0, 'content': [{'end': 731.395, 'text': "that's super interesting from a theoretical point of view.", 'start': 727.852, 'duration': 3.543}, {'end': 740.423, 'text': 'And in fact, as you are thinking about that problem, you can also get inspiration for better practical problem solvers.', 'start': 731.895, 'duration': 8.528}, {'end': 744.632, 'text': 'On the other hand, we have to admit that at the moment,', 'start': 741.331, 'duration': 3.301}, {'end': 752.714, 'text': 'the best practical problem solvers for all kinds of problems that we are now solving through what is called AI.', 'start': 744.632, 'duration': 8.082}, {'end': 757.916, 'text': 'at the moment they are not of the kind that is inspired by these questions.', 'start': 752.714, 'duration': 5.202}], 'summary': 'Theoretical concepts can inspire practical problem solvers; current ai not inspired by these concepts.', 'duration': 30.064, 'max_score': 727.852, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo727852.jpg'}, {'end': 815.163, 'src': 'embed', 'start': 770.359, 'weight': 1, 'content': [{'end': 782.65, 'text': 'to try to find a program that is running on these recurrent networks such that it can solve some interesting problems such as speech recognition or machine translation and something like that.', 'start': 770.359, 'duration': 12.291}, {'end': 790.379, 'text': 'And there is very little theory behind the best solutions that we have at the moment that can do that.', 'start': 783.391, 'duration': 6.988}, {'end': 792.581, 'text': 'Do you think that needs to change??', 'start': 790.819, 'duration': 1.762}, {'end': 793.763, 'text': 'Do you think that will change??', 'start': 792.601, 'duration': 1.162}, {'end': 795.144, 'text': 'Or can we go?', 'start': 794.123, 'duration': 1.021}, {'end': 809.238, 'text': 'can we create a general intelligence system without ever really proving that that system is intelligent in some kind of mathematical way? solving machine translation perfectly or something like that within some kind of syntactic definition of a language?', 'start': 795.144, 'duration': 14.094}, {'end': 815.163, 'text': "or can we just be super impressed by the thing working extremely well and that's sufficient?", 'start': 809.238, 'duration': 5.925}], 'summary': 'Seeking a program for recurrent networks to solve problems like speech recognition and machine translation; little theory behind current solutions; questioning the need for proving intelligence mathematically.', 'duration': 44.804, 'max_score': 770.359, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo770359.jpg'}, {'end': 932.422, 'src': 'embed', 'start': 905.101, 'weight': 4, 'content': [{'end': 911.324, 'text': 'a general intelligence system will ultimately be a simple one, maybe a pseudocode of a few lines to be able to describe it.', 'start': 905.101, 'duration': 6.223}, {'end': 916.266, 'text': 'Can you talk through your intuition behind this idea?', 'start': 911.844, 'duration': 4.422}, {'end': 925.25, 'text': 'Why you feel that, at its core, intelligence is a simple algorithm?', 'start': 916.866, 'duration': 8.384}, {'end': 932.422, 'text': 'Experience tells us that the stuff that works best is really simple.', 'start': 927.02, 'duration': 5.402}], 'summary': 'General intelligence is a simple algorithm consisting of a few lines, as simplicity often yields the best results.', 'duration': 27.321, 'max_score': 905.101, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo905101.jpg'}, {'end': 1079.962, 'src': 'embed', 'start': 1051.822, 'weight': 5, 'content': [{'end': 1061.374, 'text': 'If we take a step back through all of human civilization and just the universe in general, how do you think about evolution?', 'start': 1051.822, 'duration': 9.552}, {'end': 1066.58, 'text': 'And what if creating a universe is required to achieve this final step?', 'start': 1061.494, 'duration': 5.086}, {'end': 1077.202, 'text': 'What if going through the very painful and inefficient process of evolution is needed to come up with this set of abstractions that ultimately lead to intelligence?', 'start': 1067.44, 'duration': 9.762}, {'end': 1079.962, 'text': "Do you think there's a shortcut??", 'start': 1077.802, 'duration': 2.16}], 'summary': 'Evolution may be a necessary, albeit painful, process for achieving intelligence in the universe.', 'duration': 28.14, 'max_score': 1051.822, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo1051822.jpg'}, {'end': 1107.957, 'src': 'heatmap', 'start': 1051.822, 'weight': 0.845, 'content': [{'end': 1061.374, 'text': 'If we take a step back through all of human civilization and just the universe in general, how do you think about evolution?', 'start': 1051.822, 'duration': 9.552}, {'end': 1066.58, 'text': 'And what if creating a universe is required to achieve this final step?', 'start': 1061.494, 'duration': 5.086}, {'end': 1077.202, 'text': 'What if going through the very painful and inefficient process of evolution is needed to come up with this set of abstractions that ultimately lead to intelligence?', 'start': 1067.44, 'duration': 9.762}, {'end': 1079.962, 'text': "Do you think there's a shortcut??", 'start': 1077.802, 'duration': 2.16}, {'end': 1087.464, 'text': 'Or do you think we have to create something like our universe in order to create something like human level intelligence?', 'start': 1080.782, 'duration': 6.682}, {'end': 1095.565, 'text': 'So far, the only example we have is this one, this universe, and we can do better.', 'start': 1089.524, 'duration': 6.041}, {'end': 1107.957, 'text': 'Maybe not, but we are part of this whole process.', 'start': 1100.854, 'duration': 7.103}], 'summary': 'Exploring the necessity of evolution for achieving intelligence and the potential for creating a universe to achieve human-level intelligence.', 'duration': 56.135, 'max_score': 1051.822, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo1051822.jpg'}, {'end': 1140.713, 'src': 'embed', 'start': 1114.58, 'weight': 6, 'content': [{'end': 1124.124, 'text': 'Everything points to that possibility, because gravity and other basic forces are really simple laws that can be easily described,', 'start': 1114.58, 'duration': 9.544}, {'end': 1126.184, 'text': 'also in just a few lines of code, basically.', 'start': 1124.124, 'duration': 2.06}, {'end': 1138.231, 'text': 'And then there are these other events, the apparently random events in the history of the universe which, as far as we know,', 'start': 1127.165, 'duration': 11.066}, {'end': 1140.713, 'text': "at the moment don't have a compact code.", 'start': 1138.231, 'duration': 2.482}], 'summary': 'Gravity and basic forces can be described in a few lines of code, but random events lack a compact code.', 'duration': 26.133, 'max_score': 1114.58, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo1114580.jpg'}], 'start': 624.459, 'title': 'Practical problem solvers', 'summary': 'Discusses challenges of practical problem solving in the context of p versus np, highlighting limitations of current ai techniques and potential for theoretical insights to inspire better practical problem solvers. it also explores achieving practically optimal problem solvers under limited resources, emphasizing the potential for a simple algorithm at the core of general intelligence and questioning the necessity of complex abstractions for human-level intelligence.', 'chapters': [{'end': 815.163, 'start': 624.459, 'title': 'P vs np and practical problem solvers', 'summary': 'Discusses the challenges of practical problem solving in the context of p versus np, highlighting the limitations of current ai techniques and the potential for theoretical insights to inspire better practical problem solvers.', 'duration': 190.704, 'highlights': ['The constant overhead in solving large problems pales in comparison as n gets larger, indicating that almost all large problems are already solved in the best possible way today.', 'The current best practical problem solvers, such as recurrent neural networks, lack theoretical foundations, and there is little theory behind the solutions for tasks like speech recognition and machine translation.', 'The theoretical concept of P versus NP can provide inspiration for better practical problem solvers, offering insights for improving the current AI techniques and search methods.', 'The discussion raises the question of whether a general intelligence system can be created without proving its intelligence in a mathematical way, emphasizing the need for a shift towards more theoretically inspired practical problem solving.']}, {'end': 1226.832, 'start': 815.163, 'title': 'Achieving practical optimal problem solvers', 'summary': 'Explores the concept of achieving practically optimal problem solvers under limited resources, emphasizing the potential for a simple algorithm at the core of general intelligence, and questioning the necessity of complex abstractions and universe creation for human-level intelligence.', 'duration': 411.669, 'highlights': ['The concept of achieving practically optimal problem solvers under limited resources is explored, with emphasis on the potential for a simple algorithm at the core of general intelligence. The chapter delves into the idea of practically optimal problem solvers under limited resources, suggesting the potential for a simple algorithm at the core of general intelligence.', 'The necessity of complex abstractions and universe creation for human-level intelligence is questioned, with contemplation on whether a shortcut exists. The necessity of complex abstractions and universe creation for human-level intelligence is pondered, along with the potential existence of a shortcut.', 'The discussion revolves around the simplicity of the laws governing the universe, such as gravity, and the potential for a compact code to describe seemingly random events. The simplicity of the laws governing the universe, including gravity, and the possibility of a compact code to describe seemingly random events is discussed.']}], 'duration': 602.373, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo624459.jpg', 'highlights': ['The theoretical concept of P versus NP can provide inspiration for better practical problem solvers, offering insights for improving the current AI techniques and search methods.', 'The current best practical problem solvers, such as recurrent neural networks, lack theoretical foundations, and there is little theory behind the solutions for tasks like speech recognition and machine translation.', 'The constant overhead in solving large problems pales in comparison as n gets larger, indicating that almost all large problems are already solved in the best possible way today.', 'The discussion raises the question of whether a general intelligence system can be created without proving its intelligence in a mathematical way, emphasizing the need for a shift towards more theoretically inspired practical problem solving.', 'The concept of achieving practically optimal problem solvers under limited resources is explored, with emphasis on the potential for a simple algorithm at the core of general intelligence.', 'The necessity of complex abstractions and universe creation for human-level intelligence is questioned, with contemplation on whether a shortcut exists.', 'The discussion revolves around the simplicity of the laws governing the universe, such as gravity, and the potential for a compact code to describe seemingly random events.']}, {'end': 1781.991, 'segs': [{'end': 1317.743, 'src': 'embed', 'start': 1288.514, 'weight': 0, 'content': [{'end': 1298.397, 'text': "We don't have any fundamental reason at the moment to believe that this is truly random and not just a deterministic video game.", 'start': 1288.514, 'duration': 9.883}, {'end': 1302.418, 'text': 'If it was a deterministic video game, it would be much more beautiful.', 'start': 1299.157, 'duration': 3.261}, {'end': 1305.819, 'text': 'Because beauty is simplicity.', 'start': 1303.118, 'duration': 2.701}, {'end': 1314.022, 'text': 'And many of the basic laws of the universe, like gravity and the other basic forces, are very simple.', 'start': 1306.78, 'duration': 7.242}, {'end': 1317.743, 'text': 'So very short programs can explain what these are doing.', 'start': 1314.142, 'duration': 3.601}], 'summary': 'The universe may be a deterministic video game, as simplicity explains its basic laws.', 'duration': 29.229, 'max_score': 1288.514, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo1288514.jpg'}, {'end': 1639.091, 'src': 'embed', 'start': 1585.85, 'weight': 3, 'content': [{'end': 1597.377, 'text': 'and then he had all these data points, and suddenly it turned out that he can greatly compress the data by predicting it through an ellipse law.', 'start': 1585.85, 'duration': 11.527}, {'end': 1603.34, 'text': 'So it turns out that all these data points are more or less on ellipses around the Sun.', 'start': 1598.077, 'duration': 5.263}, {'end': 1610.543, 'text': 'And another guy came along whose name was Newton and before him Hooke.', 'start': 1605.601, 'duration': 4.942}, {'end': 1620.728, 'text': 'And they said the same thing that is making these planets move like that is what makes the apples fall down.', 'start': 1611.443, 'duration': 9.285}, {'end': 1629.462, 'text': 'And it also holds for stones and for all kinds of other objects.', 'start': 1621.994, 'duration': 7.468}, {'end': 1639.091, 'text': 'And certainly many, many of these observations became much more compressible because as long as you can predict the next thing,', 'start': 1631.143, 'duration': 7.948}], 'summary': 'Data points can be compressed using ellipse law, applicable to planets and objects.', 'duration': 53.241, 'max_score': 1585.85, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo1585850.jpg'}, {'end': 1710.37, 'src': 'embed', 'start': 1687.282, 'weight': 2, 'content': [{'end': 1698.455, 'text': 'No matter how fast you accelerate and how fast or hard you decelerate, and no matter what is the gravity in your local framework,', 'start': 1687.282, 'duration': 11.173}, {'end': 1700.958, 'text': 'light speed always looks the same.', 'start': 1698.455, 'duration': 2.503}, {'end': 1704.261, 'text': 'And from that you can calculate all the consequences.', 'start': 1701.438, 'duration': 2.823}, {'end': 1710.37, 'text': "So it's a very simple thing, and it allows you to further compress all the observations,", 'start': 1704.301, 'duration': 6.069}], 'summary': 'Light speed remains constant regardless of acceleration or gravity, simplifying calculations.', 'duration': 23.088, 'max_score': 1687.282, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo1687282.jpg'}], 'start': 1227.953, 'title': 'Determinism vs randomness and compression progress in science', 'summary': "Delves into the concept of determinism versus randomness in the universe, advocating for the elegance of simple determinism and the ongoing search for a concise explanation. it also explores the history of science as a progression towards greater compression, exemplified by milestones from kepler's laws to einstein's theory of relativity.", 'chapters': [{'end': 1537.61, 'start': 1227.953, 'title': 'Determinism vs randomness in the universe', 'summary': "Discusses the concept of determinism versus randomness in the universe, highlighting the idea that the universe may not be truly random and could be described by a short program, making it more elegant and beautiful. it emphasizes the simplicity and compressibility of the universe's fundamental laws and the ongoing search for a simple explanation, challenging the romantic notion of randomness and advocating for the beauty of simple determinism.", 'duration': 309.657, 'highlights': ['The universe may not be truly random and could be described by a short program, making it more elegant and beautiful. The concept of determinism versus randomness in the universe is explored, suggesting that the universe may not be truly random and could be described by a short program, making it more elegant and beautiful.', "The universe's fundamental laws are simple and compressible, and the ongoing search for a simple explanation challenges the romantic notion of randomness. The simplicity and compressibility of the universe's fundamental laws are emphasized, and the ongoing search for a simple explanation challenges the romantic notion of randomness.", 'The concept of simplicity and compressibility is advocated, emphasizing the beauty of simple determinism over randomness. The concept of simplicity and compressibility is advocated, emphasizing the beauty of simple determinism over randomness and challenging the romantic notion of randomness.']}, {'end': 1781.991, 'start': 1538.898, 'title': 'History of science: compression progress', 'summary': "Discusses how the history of science is a journey towards greater compression, illustrated by examples from kepler's ellipse law to einstein's general theory of relativity, aiming at compressing data and making progress in scientific insights.", 'duration': 243.093, 'highlights': ["Einstein's general theory of relativity explains deviations from Kepler's and Newton's predictions, resulting in further compressing observations and demonstrating that light speed always looks the same, leading to minimal deviations from the predictions (relevance: 5)", 'Kepler greatly compresses data points by predicting planetary motion through an ellipse law, showcasing the progress in compression and insights in scientific discoveries (relevance: 4)', 'Newton and Hooke make observations more compressible by linking planetary motion to the force that makes apples fall, demonstrating the concept of predictive coding and reducing the need to store extra data (relevance: 3)', 'The journey of science involves making progress in compression, evident from realizing simple ways to predict and compress data, leading to deeper insights and scientific fun (relevance: 2)']}], 'duration': 554.038, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo1227953.jpg', 'highlights': ['The universe may not be truly random and could be described by a short program, making it more elegant and beautiful.', "The simplicity and compressibility of the universe's fundamental laws are emphasized, challenging the romantic notion of randomness.", "Einstein's general theory of relativity explains deviations from Kepler's and Newton's predictions, resulting in further compressing observations and demonstrating that light speed always looks the same.", 'Kepler greatly compresses data points by predicting planetary motion through an ellipse law, showcasing the progress in compression and insights in scientific discoveries.', 'Newton and Hooke make observations more compressible by linking planetary motion to the force that makes apples fall, demonstrating the concept of predictive coding and reducing the need to store extra data.']}, {'end': 2134.138, 'segs': [{'end': 1811.758, 'src': 'embed', 'start': 1783.078, 'weight': 0, 'content': [{'end': 1785.74, 'text': 'And we can build artificial systems that do the same thing.', 'start': 1783.078, 'duration': 2.662}, {'end': 1793.085, 'text': 'They measure the depth of their insights as they are looking at the data which is coming in through their own experiments.', 'start': 1786.06, 'duration': 7.025}, {'end': 1799.87, 'text': 'And we give them a reward, an intrinsic reward, in proportion to this depth of insight.', 'start': 1793.665, 'duration': 6.205}, {'end': 1811.758, 'text': 'And since they are trying to maximize the rewards they get, they are suddenly motivated to come up with new action sequences,', 'start': 1801.23, 'duration': 10.528}], 'summary': 'Artificial systems measure insight depth to maximize rewards and generate action sequences.', 'duration': 28.68, 'max_score': 1783.078, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo1783078.jpg'}, {'end': 1871.68, 'src': 'heatmap', 'start': 1821.06, 'weight': 0.811, 'content': [{'end': 1827.926, 'text': "the property that they can learn something about, see a pattern in there which they hadn't seen yet before.", 'start': 1821.06, 'duration': 6.866}, {'end': 1837.094, 'text': "So there's an idea of power play that you've described a training, a general problem solver in this kind of way of looking for the unsolved problems.", 'start': 1828.829, 'duration': 8.265}, {'end': 1841.917, 'text': "Can you describe that idea a little further? It's another very simple idea.", 'start': 1838.095, 'duration': 3.822}, {'end': 1850.359, 'text': 'So normally what you do in computer science, you have you have some guy who gives you a problem,', 'start': 1842.457, 'duration': 7.902}, {'end': 1871.68, 'text': 'and then there is a huge search space of potential solution candidates and you somehow try them out and you have more or less sophisticated ways of moving around in that search space until you finally found a solution which you consider satisfactory.', 'start': 1850.359, 'duration': 21.321}], 'summary': 'Exploring the concept of power play in training a general problem solver in computer science.', 'duration': 50.62, 'max_score': 1821.06, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo1821060.jpg'}, {'end': 1909.698, 'src': 'embed', 'start': 1875.927, 'weight': 1, 'content': [{'end': 1884.749, 'text': "PowerPlay just goes one little step further and says let's not only search for solutions to a given problem,", 'start': 1875.927, 'duration': 8.822}, {'end': 1895.812, 'text': "but let's search to pairs of problems and their solutions where the system itself has the opportunity to phrase its own problem.", 'start': 1884.749, 'duration': 11.063}, {'end': 1909.698, 'text': 'So we are looking suddenly at pairs of problems and their solutions or modifications of the problem solver that is supposed to generate a solution to that new problem.', 'start': 1897.313, 'duration': 12.385}], 'summary': 'Powerplay seeks solutions to pairs of problems, expanding problem-solving capabilities.', 'duration': 33.771, 'max_score': 1875.927, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo1875927.jpg'}], 'start': 1783.078, 'title': 'Ai systems and problem-solving', 'summary': 'Explores how artificial systems are motivated by intrinsic rewards to discover new patterns, introduces powerplay concept enabling systems to pose their own questions, and discusses expanding problem-solving abilities for genuine creative discovery.', 'chapters': [{'end': 1837.094, 'start': 1783.078, 'title': 'Artificial systems and intrinsic rewards', 'summary': 'Discusses how artificial systems are motivated to come up with new action sequences and experiments by receiving intrinsic rewards in proportion to the depth of their insights, leading to the discovery of new patterns in the data.', 'duration': 54.016, 'highlights': ['Artificial systems are motivated to come up with new action sequences and experiments by receiving intrinsic rewards in proportion to the depth of their insights.', 'The depth of insights in artificial systems is measured as they look at the data coming in through their experiments.', 'New experiments lead to the discovery of patterns in the data that were not seen before.']}, {'end': 1979.152, 'start': 1838.095, 'title': 'Powerplay: empowering systems to pose questions', 'summary': 'Introduces powerplay, a concept that allows systems to not only search for solutions to given problems but also to phrase their own problems, resulting in the ability to build ai systems that can act like scientists, with the freedom to pose their own questions.', 'duration': 141.057, 'highlights': ['PowerPlay allows systems to search for pairs of problems and their solutions or modifications, giving them the freedom to pose their own questions, resembling the behavior of scientists.', 'The concept of PowerPlay provides an additional degree of freedom to AI systems, enabling them to act like scientists by not only solving existing questions but also posing their own questions.', 'The intelligence of human beings is considered to be the ability to create something entirely new, which is the kind of freedom and insight that PowerPlay aims to provide to AI systems.', 'PowerPlay introduces a dimension of freedom in the search for solutions, allowing the system to move beyond just solving existing problems and instead pose its own questions, which is crucial in building artificial scientists.']}, {'end': 2134.138, 'start': 1980.233, 'title': 'The power of solving new problems', 'summary': 'Discusses the concept of finding the simplest unsolvable problem to expand the current problem-solving abilities, aiming to break existing limitations and shift the horizon of knowledge, ultimately leading to genuine creative discovery.', 'duration': 153.905, 'highlights': ['The PowerPlay approach involves searching for the simplest unsolvable problem to add to the repertoire, aiming to expand problem-solving abilities beyond the current horizon.', 'The goal is to break the existing rules and shift the knowledge horizon through the addition of new problems that were previously unsolvable by the old problem solver.', 'The approach is akin to adding new axioms, as demonstrated by Gödel, to expand the repertoire while preserving the consistency of the system.']}], 'duration': 351.06, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo1783078.jpg', 'highlights': ['Artificial systems are motivated to come up with new action sequences and experiments by receiving intrinsic rewards in proportion to the depth of their insights.', 'PowerPlay allows systems to search for pairs of problems and their solutions or modifications, giving them the freedom to pose their own questions, resembling the behavior of scientists.', 'The PowerPlay approach involves searching for the simplest unsolvable problem to add to the repertoire, aiming to expand problem-solving abilities beyond the current horizon.']}, {'end': 2741.025, 'segs': [{'end': 2190.914, 'src': 'embed', 'start': 2164.873, 'weight': 2, 'content': [{'end': 2171.638, 'text': 'So humans are curious, and um, that means they behave like scientists.', 'start': 2164.873, 'duration': 6.765}, {'end': 2173.2, 'text': 'not only the official scientists,', 'start': 2171.638, 'duration': 1.562}, {'end': 2182.968, 'text': 'but even the babies behave like scientists and they play around with toys to figure out how the world works and how it is responding to their actions.', 'start': 2173.2, 'duration': 9.768}, {'end': 2186.611, 'text': "And that's how they learn about gravity and everything.", 'start': 2183.709, 'duration': 2.902}, {'end': 2190.914, 'text': 'In 1990 we had the first systems like that,', 'start': 2188.453, 'duration': 2.461}], 'summary': 'Humans, including babies, behave like scientists to learn about the world, even playing to understand gravity and other concepts.', 'duration': 26.041, 'max_score': 2164.873, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo2164873.jpg'}, {'end': 2248.767, 'src': 'embed', 'start': 2223.295, 'weight': 3, 'content': [{'end': 2231.439, 'text': "they are what we have built in as well, because evolution discovered that's a good way of exploring the unknown world,", 'start': 2223.295, 'duration': 8.144}, {'end': 2239.504, 'text': 'and a guy who explores the unknown world has a higher chance of solving problems that he needs to survive in this world.', 'start': 2231.439, 'duration': 8.065}, {'end': 2245.124, 'text': 'on the other hand, Those guys who were too curious, they were weeded out as well.', 'start': 2239.504, 'duration': 5.62}, {'end': 2247.146, 'text': 'So you have to find this trade-off.', 'start': 2245.324, 'duration': 1.822}, {'end': 2248.767, 'text': 'Evolution found a certain trade-off.', 'start': 2247.266, 'duration': 1.501}], 'summary': 'Evolution found a trade-off between curiosity and survival, favoring explorers.', 'duration': 25.472, 'max_score': 2223.295, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo2223295.jpg'}, {'end': 2365.618, 'src': 'embed', 'start': 2339.688, 'weight': 1, 'content': [{'end': 2345.929, 'text': 'machine tries to find a solution to that, and this has been happening for many decades.', 'start': 2339.688, 'duration': 6.241}, {'end': 2360.094, 'text': 'and for many decades machines have found creative solutions to interesting problems where humans were not aware of these particularly creative solutions but then appreciated that the machine found that.', 'start': 2345.929, 'duration': 14.165}, {'end': 2365.618, 'text': 'The second is the pure creativity that I would call what I just mentioned.', 'start': 2361.754, 'duration': 3.864}], 'summary': 'Machines have been finding creative solutions to problems for many decades.', 'duration': 25.93, 'max_score': 2339.688, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo2339688.jpg'}, {'end': 2519.198, 'src': 'embed', 'start': 2493.908, 'weight': 0, 'content': [{'end': 2500.553, 'text': "And you've also said that consciousness, in the same kind of way, may be a byproduct of of problem solving?", 'start': 2493.908, 'duration': 6.645}, {'end': 2501.614, 'text': 'Yeah, do you think?', 'start': 2500.573, 'duration': 1.041}, {'end': 2504.654, 'text': 'Do you find this an interesting byproduct??', 'start': 2502.734, 'duration': 1.92}, {'end': 2514.637, 'text': "Do you think it's a useful byproduct? What are your thoughts on consciousness in general? Or is it simply a byproduct of greater and greater capabilities of problem solving??", 'start': 2504.674, 'duration': 9.963}, {'end': 2519.198, 'text': "That's that's similar to creativity.", 'start': 2515.297, 'duration': 3.901}], 'summary': 'Consciousness may be a byproduct of problem solving, similar to creativity.', 'duration': 25.29, 'max_score': 2493.908, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo2493908.jpg'}, {'end': 2639.742, 'src': 'heatmap', 'start': 2589.369, 'weight': 0.763, 'content': [{'end': 2593.653, 'text': "And it's just trained on the long and long history of interactions with the world.", 'start': 2589.369, 'duration': 4.284}, {'end': 2597.596, 'text': 'So it becomes a predictive model of the world basically.', 'start': 2594.053, 'duration': 3.543}, {'end': 2600.459, 'text': 'And therefore also a compressor.', 'start': 2598.177, 'duration': 2.282}, {'end': 2606.622, 'text': "of the observations of the world, because whatever you can predict, you don't have to store extra.", 'start': 2600.939, 'duration': 5.683}, {'end': 2609.724, 'text': 'So compression is a side effect of prediction.', 'start': 2606.642, 'duration': 3.082}, {'end': 2612.825, 'text': 'And how does this recurrent network compress??', 'start': 2610.684, 'duration': 2.141}, {'end': 2621.17, 'text': "Well, it's inventing little subprograms, little subnetworks that stand for everything that frequently appears in the environment.", 'start': 2613.246, 'duration': 7.924}, {'end': 2628.129, 'text': 'like bottles and microphones and faces, maybe lots of faces in my environment.', 'start': 2622.082, 'duration': 6.047}, {'end': 2635.437, 'text': "So I'm learning to create something like a prototype face and a new face comes along and all I have to encode are the deviations from the prototype.", 'start': 2628.149, 'duration': 7.288}, {'end': 2639.742, 'text': "So it's compressing all the time the stuff that frequently appears.", 'start': 2636.418, 'duration': 3.324}], 'summary': 'Recurrent network compresses by inventing subprograms to represent frequently appearing objects, achieving compression through prediction.', 'duration': 50.373, 'max_score': 2589.369, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo2589369.jpg'}], 'start': 2135.839, 'title': 'Creativity, curiosity, and problem solving', 'summary': 'Delves into the theory of creativity and curiosity in humans, emphasizing the intrinsic reward of discovery, artificial curiosity in problem-solving, and the role of creativity and consciousness in problem solving, including the emergence of consciousness as a side effect of problem-solving processes.', 'chapters': [{'end': 2282.093, 'start': 2135.839, 'title': 'Theory of creativity and curiosity in humans', 'summary': 'Discusses the intrinsic reward of discovery, how humans behave like scientists driven by curiosity, and the importance of artificial curiosity in problem-solving, reflecting the principle present in our brains.', 'duration': 146.254, 'highlights': ['Humans behave like scientists, even babies, playing around to understand the world, which led to the development of systems for creating situations beyond their knowledge, enhancing problem-solving abilities (1990).', 'Evolution discovered that curiosity is a good strategy for exploring the unknown world, as it increases the chance of solving survival problems, while also finding a trade-off to weed out overly curious individuals.', 'The presence of a certain percentage of extremely explorative individuals in society indicates the importance of artificial curiosity in problem-solving, suggesting its likely presence in a similar form in our brains.', 'The discussion on how humans behave like scientists and the importance of artificial curiosity in problem-solving (1990).']}, {'end': 2741.025, 'start': 2283.154, 'title': 'Role of creativity and consciousness in problem solving', 'summary': 'Discusses the role of creativity and consciousness in problem solving, including the presence of two types of creativity in machines, the byproduct of curiosity in problem-solving machines, and the emergence of consciousness as a side effect of problem-solving processes.', 'duration': 457.871, 'highlights': ['The presence of two types of creativity in machines The chapter explores the existence of two types of creativity in machines, including the applied creativity where a problem is given to the machine, and the pure creativity where the machine has the freedom to select its own problem.', 'The emergence of consciousness as a side effect of problem-solving processes It discusses the emergence of consciousness as a byproduct of problem-solving processes in machines, demonstrating the development of internal self-models through data compression during problem solving.', 'The byproduct of curiosity in problem-solving machines It delves into the byproduct of curiosity in problem-solving machines, suggesting that curiosity and the will to invent new problems are natural outcomes of the general search in problem-solving processes.']}], 'duration': 605.186, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo2135839.jpg', 'highlights': ['The emergence of consciousness as a side effect of problem-solving processes', 'The presence of two types of creativity in machines', 'The discussion on how humans behave like scientists and the importance of artificial curiosity in problem-solving', 'Evolution discovered that curiosity is a good strategy for exploring the unknown world']}, {'end': 3381.338, 'segs': [{'end': 2876.467, 'src': 'embed', 'start': 2852.486, 'weight': 0, 'content': [{'end': 2864.213, 'text': "Well, most problems in the real world are deep, in the sense that the current input doesn't tell you all you need to know about the environment.", 'start': 2852.486, 'duration': 11.727}, {'end': 2869.717, 'text': 'So instead, you have to have a memory of what happened in the past.', 'start': 2865.374, 'duration': 4.343}, {'end': 2874.866, 'text': 'And often important parts of that memory dated.', 'start': 2869.777, 'duration': 5.089}, {'end': 2876.467, 'text': 'they are pretty old.', 'start': 2874.866, 'duration': 1.601}], 'summary': 'Real-world problems are deep, requiring historical memory for understanding.', 'duration': 23.981, 'max_score': 2852.486, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo2852486.jpg'}, {'end': 2962.776, 'src': 'embed', 'start': 2905.912, 'weight': 1, 'content': [{'end': 2912.556, 'text': 'It has to store that, uh, 50 steps ago, there was an S or an 11 or seven.', 'start': 2905.912, 'duration': 6.644}, {'end': 2918.108, 'text': 'So there you have already a problem of depth 50,', 'start': 2914.946, 'duration': 3.162}, {'end': 2927.353, 'text': 'because for each time step you have something like a virtual layer and the expanded unrolled version of this recurrent network which is doing the speech recognition.', 'start': 2918.108, 'duration': 9.245}, {'end': 2934.998, 'text': 'So these long time lags, they translate into problem depth and most problems.', 'start': 2928.394, 'duration': 6.604}, {'end': 2945.604, 'text': 'in this world are such that you really have to look far back in time to understand what is the problem and to solve it.', 'start': 2936.297, 'duration': 9.307}, {'end': 2952.229, 'text': "But just like with LSTMs, you don't necessarily need to, when you look back in time, remember every aspect.", 'start': 2946.244, 'duration': 5.985}, {'end': 2954.33, 'text': 'You just need to remember the important aspects.', 'start': 2952.249, 'duration': 2.081}, {'end': 2955.251, 'text': "That's right.", 'start': 2954.851, 'duration': 0.4}, {'end': 2962.776, 'text': 'The network has to learn to put the important stuff into memory and to ignore the unimportant noise.', 'start': 2955.531, 'duration': 7.245}], 'summary': 'Long time lags in speech recognition require remembering important aspects and ignoring unimportant noise.', 'duration': 56.864, 'max_score': 2905.912, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo2905912.jpg'}, {'end': 3045.062, 'src': 'embed', 'start': 3013.443, 'weight': 4, 'content': [{'end': 3021.671, 'text': 'as us humans do the credit assignment problem across way back not just 50 times steps or 100 or 1000, but millions and billions.', 'start': 3013.443, 'duration': 8.228}, {'end': 3030.59, 'text': "It's not clear what are the practical limits of the LSTM when it comes to looking back.", 'start': 3024.645, 'duration': 5.945}, {'end': 3039.977, 'text': 'Already in 2006, I think, we had examples where it not only looked back tens of thousands of steps, but really millions of steps.', 'start': 3031.27, 'duration': 8.707}, {'end': 3045.062, 'text': 'And Juan Perez Ortiz in my lab, I think,', 'start': 3040.878, 'duration': 4.184}], 'summary': 'Lstm can look back millions of steps for credit assignment problem.', 'duration': 31.619, 'max_score': 3013.443, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo3013443.jpg'}, {'end': 3105.863, 'src': 'embed', 'start': 3077.808, 'weight': 2, 'content': [{'end': 3088.731, 'text': "And so a reinforcement learning system which is trying to maximize its future expected reward and doesn't know yet which of these many possible futures should I select,", 'start': 3077.808, 'duration': 10.923}, {'end': 3095.093, 'text': 'given this one single past is facing problems that the LSTM by itself cannot solve.', 'start': 3088.731, 'duration': 6.362}, {'end': 3105.863, 'text': 'So the LSTM is good for coming up with a compact representation of the history so far of the history and of observations and actions so far.', 'start': 3096.554, 'duration': 9.309}], 'summary': 'Reinforcement learning system faces problems that lstm alone cannot solve.', 'duration': 28.055, 'max_score': 3077.808, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo3077808.jpg'}, {'end': 3357.112, 'src': 'embed', 'start': 3329.103, 'weight': 5, 'content': [{'end': 3335.552, 'text': "That's the correct interesting direction for research in your view? I do think so.", 'start': 3329.103, 'duration': 6.449}, {'end': 3347.113, 'text': 'We have a company called Naysense, which has applied reinforcement learning to little Audis, which learn to park without a teacher.', 'start': 3335.652, 'duration': 11.461}, {'end': 3350.766, 'text': 'The same principles were used, of course.', 'start': 3348.263, 'duration': 2.503}, {'end': 3357.112, 'text': 'So these little Audis, they are small, maybe like that, so much smaller than the real Audis.', 'start': 3351.567, 'duration': 5.545}], 'summary': 'Naysense uses reinforcement learning for little audis to park autonomously.', 'duration': 28.009, 'max_score': 3329.103, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo3329103.jpg'}], 'start': 2741.025, 'title': 'Depth in neural networks', 'summary': 'Discusses the value of depth in neural networks, emphasizing the success of lstms in modeling temporal aspects, the importance of depth in solving real-world problems, and the limitations of lstms in looking far back in time compared to human capability. it also explains the use of lstm in reinforcement learning for history representation and planning efficient action sequences, with examples of application in self-driving cars.', 'chapters': [{'end': 3076.592, 'start': 2741.025, 'title': 'Value of depth in neural networks', 'summary': 'Discusses the value of depth in neural networks, focusing on the success of lstms in modeling temporal aspects in data, the importance of depth in solving real-world problems, and the ability of networks to filter and remember important information. it also touches upon the limitations of lstms in looking far back in time compared to human capability.', 'duration': 335.567, 'highlights': ["The importance of depth in solving real-world problems is emphasized, as the current input doesn't always provide all necessary information about the environment, requiring a memory of past events, often dated. N/A", "The success of LSTMs in speech recognition, particularly in distinguishing between similar sounding numbers like '11' and '7', is explained, highlighting the need for networks to store information from 50 time steps ago, illustrating a problem of depth 50. 50 time steps", 'The ability of networks, like LSTMs, to filter important information and ignore unimportant noise is mentioned, emphasizing the importance of learning to store important aspects in memory. N/A', 'The limitation of LSTMs in looking far back in time compared to human capability is discussed, with examples of LSTMs looking back tens of thousands to millions of steps but not being able to match human capability of looking back across millions and billions of steps. Tens of thousands to millions of steps']}, {'end': 3381.338, 'start': 3077.808, 'title': 'Reinforcement learning and lstm', 'summary': 'Explains how a reinforcement learning system uses lstm for history representation and planning efficient action sequences, and discusses the potential impact of reinforcement learning in real systems, citing examples of application in self-driving cars.', 'duration': 303.53, 'highlights': ['The chapter explains how a reinforcement learning system uses LSTM for history representation and planning efficient action sequences The LSTM is used to come up with a compact representation of the history and observations, and the challenge is to efficiently select among many possible action sequences to maximize reward.', 'The chapter discusses the potential impact of reinforcement learning in real systems, citing examples of application in self-driving cars Reinforcement learning has been applied to small Audis for learning to park without a teacher, leveraging sensors and learning principles similar to real Audis.']}], 'duration': 640.313, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo2741025.jpg', 'highlights': ["The importance of depth in solving real-world problems is emphasized, as the current input doesn't always provide all necessary information about the environment, requiring a memory of past events, often dated.", "The success of LSTMs in speech recognition, particularly in distinguishing between similar sounding numbers like '11' and '7', is explained, highlighting the need for networks to store information from 50 time steps ago, illustrating a problem of depth 50.", 'The chapter explains how a reinforcement learning system uses LSTM for history representation and planning efficient action sequences The LSTM is used to come up with a compact representation of the history and observations, and the challenge is to efficiently select among many possible action sequences to maximize reward.', 'The ability of networks, like LSTMs, to filter important information and ignore unimportant noise is mentioned, emphasizing the importance of learning to store important aspects in memory.', 'The limitation of LSTMs in looking far back in time compared to human capability is discussed, with examples of LSTMs looking back tens of thousands to millions of steps but not being able to match human capability of looking back across millions and billions of steps.', 'The chapter discusses the potential impact of reinforcement learning in real systems, citing examples of application in self-driving cars Reinforcement learning has been applied to small Audis for learning to park without a teacher, leveraging sensors and learning principles similar to real Audis.']}, {'end': 3910.059, 'segs': [{'end': 3434.993, 'src': 'embed', 'start': 3405.356, 'weight': 0, 'content': [{'end': 3411.939, 'text': 'So at the moment, the current wave of AI is about passive pattern observation and prediction.', 'start': 3405.356, 'duration': 6.583}, {'end': 3421.923, 'text': "And that's what you have on your smartphone and what the major companies on the Pacific Rim are using to sell you ads, to do marketing.", 'start': 3412.459, 'duration': 9.464}, {'end': 3424.925, 'text': "That's the current source of profit in AI.", 'start': 3422.323, 'duration': 2.602}, {'end': 3427.126, 'text': "And that's only 1% or 2% of the world economy.", 'start': 3425.625, 'duration': 1.501}, {'end': 3434.993, 'text': 'which is big enough to make these companies pretty much the most valuable companies in the world.', 'start': 3430.567, 'duration': 4.426}], 'summary': 'Current ai focuses on passive pattern observation and prediction for ads and marketing, contributing to 1-2% of the world economy.', 'duration': 29.637, 'max_score': 3405.356, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo3405356.jpg'}, {'end': 3492.564, 'src': 'embed', 'start': 3456.828, 'weight': 2, 'content': [{'end': 3458.809, 'text': "we're not talking about 100 years from now.", 'start': 3456.828, 'duration': 1.981}, {'end': 3462.652, 'text': "we're talking about sort of the near term impact of RL.", 'start': 3458.809, 'duration': 3.843}, {'end': 3464.913, 'text': 'do you think really good simulation is required?', 'start': 3462.652, 'duration': 2.261}, {'end': 3473.659, 'text': 'Or is there other techniques like imitation learning, you know, observing other humans operating in the real world?', 'start': 3465.293, 'duration': 8.366}, {'end': 3476.361, 'text': 'Where do you think the success will come from?', 'start': 3473.699, 'duration': 2.662}, {'end': 3492.564, 'text': 'So at the moment we have a tendency of using physics simulations to learn behavior for machines that learn to solve problems that humans also do not know how to solve.', 'start': 3477.699, 'duration': 14.865}], 'summary': 'Near-term impact of rl, debate over need for good simulation, imitation learning, and use of physics simulations for machine behavior learning.', 'duration': 35.736, 'max_score': 3456.828, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo3456828.jpg'}, {'end': 3525.348, 'src': 'embed', 'start': 3502.328, 'weight': 1, 'content': [{'end': 3513.538, 'text': 'No, they learn a predictive model of the world which maybe sometimes is wrong in many ways, but captures all kinds of important abstract,', 'start': 3502.328, 'duration': 11.21}, {'end': 3517.323, 'text': 'high-level predictions which are really important to be successful.', 'start': 3513.538, 'duration': 3.785}, {'end': 3520.366, 'text': "And that's what is..", 'start': 3518.544, 'duration': 1.822}, {'end': 3525.348, 'text': "What was the future 30 years ago when you started that type of research, but it's still the future.", 'start': 3521.087, 'duration': 4.261}], 'summary': 'Predictive model captures high-level predictions important for success, even if sometimes wrong. future of research remains relevant after 30 years.', 'duration': 23.02, 'max_score': 3502.328, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo3502328.jpg'}, {'end': 3649.353, 'src': 'embed', 'start': 3620.87, 'weight': 5, 'content': [{'end': 3633.887, 'text': 'So my first publication ever actually was 1987, was the implementation of genetic algorithm of a genetic programming system in Prologue.', 'start': 3620.87, 'duration': 13.017}, {'end': 3639.59, 'text': "So Prolog, that's what you learned back then, which is a logic programming language.", 'start': 3634.648, 'duration': 4.942}, {'end': 3641.27, 'text': 'And the Japanese.', 'start': 3640.31, 'duration': 0.96}, {'end': 3649.353, 'text': 'they have this huge fifth generation AI project, which was mostly about logic programming back then,', 'start': 3641.27, 'duration': 8.083}], 'summary': 'First publication in 1987 on genetic programming in prolog for the fifth generation ai project.', 'duration': 28.483, 'max_score': 3620.87, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo3620870.jpg'}, {'end': 3747.52, 'src': 'embed', 'start': 3722.766, 'weight': 4, 'content': [{'end': 3728.911, 'text': 'such as gradient-based search and program space, rather than provably optimal things.', 'start': 3722.766, 'duration': 6.145}, {'end': 3738.656, 'text': "So logic programming certainly has a usefulness when you're trying to construct something provably optimal or provably good or something like that.", 'start': 3729.091, 'duration': 9.565}, {'end': 3743.978, 'text': "But is it useful for practical problems? It's really useful for theorem proving.", 'start': 3738.976, 'duration': 5.002}, {'end': 3747.52, 'text': 'The best theorem provers today are not neural networks.', 'start': 3744.158, 'duration': 3.362}], 'summary': 'Logic programming useful for theorem proving, not practical problems. best theorem provers not neural networks.', 'duration': 24.754, 'max_score': 3722.766, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo3722766.jpg'}, {'end': 3852.655, 'src': 'embed', 'start': 3821.587, 'weight': 3, 'content': [{'end': 3831.172, 'text': 'So I think in the not so distant future we will have for the first time little robots that learn like kids.', 'start': 3821.587, 'duration': 9.585}, {'end': 3840.156, 'text': 'And I will be able to say to the robot, Look here, robot, we are going to assemble a smartphone.', 'start': 3832.653, 'duration': 7.503}, {'end': 3847.908, 'text': "Let's take this slab of plastic and the screwdriver and let's screw in the screw like that.", 'start': 3840.917, 'duration': 6.991}, {'end': 3849.811, 'text': 'No, not like that.', 'start': 3848.869, 'duration': 0.942}, {'end': 3850.612, 'text': 'Like that.', 'start': 3850.291, 'duration': 0.321}, {'end': 3852.134, 'text': 'Not like that.', 'start': 3851.673, 'duration': 0.461}, {'end': 3852.655, 'text': 'Like that.', 'start': 3852.334, 'duration': 0.321}], 'summary': 'In the future, robots will learn like kids, enabling them to assemble smartphones with precision and adaptability.', 'duration': 31.068, 'max_score': 3821.587, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo3821587.jpg'}], 'start': 3382.711, 'title': 'Future of ai and reinforcement learning', 'summary': 'Discusses the shift in ai towards machines shaping data through their actions, affecting a larger fraction of the economy, learning predictive models, and using simulations in reinforcement learning, as well as the evolution of ai towards robots learning like kids and its impact on tasks such as assembly and interpretation of high-level imitation.', 'chapters': [{'end': 3577.828, 'start': 3382.711, 'title': 'Future of ai and reinforcement learning', 'summary': 'Discusses the current wave of ai focused on passive pattern observation and prediction, and the potential impact of the next wave, which will involve machines shaping data through their own actions, affecting a much larger fraction of the economy. it also explores the importance of learning predictive models and the use of simulations and imitation learning in reinforcement learning.', 'duration': 195.117, 'highlights': ["The next wave of AI will involve machines shaping data through their own actions, impacting a much larger fraction of the economy than the current wave focused on passive pattern observation and prediction. The next wave of AI is expected to have a significant impact on the economy, surpassing the current wave's influence on passive pattern observation and prediction.", 'The importance of learning predictive models that capture important abstract, high-level predictions necessary for success, similar to what babies do, rather than relying solely on physics simulations. Learning predictive models that capture important abstract predictions, similar to how babies learn, is emphasized as crucial for success, shifting the focus from reliance on physics simulations.', 'The discussion on the use of simulations and imitation learning in reinforcement learning, and the need for good simulation or other techniques like imitation learning for success. The debate on the requirement of good simulation or alternative techniques like imitation learning for success in reinforcement learning is explored.']}, {'end': 3910.059, 'start': 3578.709, 'title': 'Evolution of ai and future directions', 'summary': 'Discusses the evolution of ai from expert systems to logic programming and the future direction of ai, emphasizing the shift towards robots learning like kids and the potential impact on tasks such as assembly and interpretation of high-level imitation.', 'duration': 331.35, 'highlights': ['The shift towards robots learning like kids and the potential impact on tasks such as assembly and interpretation of high-level imitation. The not so distant future will see little robots that learn like kids, capable of interpreting high-level imitation and learning to imitate human actions without constant supervised signals.', 'The influence of logic programming and its usefulness in theorem proving and practical problem-solving in AI. Logic programming is useful for constructing provably optimal solutions and theorem proving, but less practical for tasks such as pattern recognition and real-world operations.', 'The historical influence of logic programming in AI, including its focus in the 80s and the influence on biologically inspired algorithms. In the 80s, there was a focus on logic programming in AI, influencing the implementation of biologically inspired algorithms like genetic programming, and its influence on languages such as Prolog.']}], 'duration': 527.348, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo3382711.jpg', 'highlights': ['The next wave of AI will involve machines shaping data through their own actions, impacting a much larger fraction of the economy than the current wave focused on passive pattern observation and prediction.', 'The importance of learning predictive models that capture important abstract, high-level predictions necessary for success, similar to what babies do, rather than relying solely on physics simulations.', 'The discussion on the use of simulations and imitation learning in reinforcement learning, and the need for good simulation or other techniques like imitation learning for success.', 'The shift towards robots learning like kids and the potential impact on tasks such as assembly and interpretation of high-level imitation.', 'The influence of logic programming and its usefulness in theorem proving and practical problem-solving in AI.', 'The historical influence of logic programming in AI, including its focus in the 80s and the influence on biologically inspired algorithms.']}, {'end': 4777.923, 'segs': [{'end': 3956.637, 'src': 'embed', 'start': 3932.483, 'weight': 0, 'content': [{'end': 3941.648, 'text': 'a much bigger AI wave is coming than the one that we are currently witnessing, which is mostly about passive pattern recognition on your smartphone.', 'start': 3932.483, 'duration': 9.165}, {'end': 3950.013, 'text': 'This is about active machines that shapes data through the actions they are executing, and they learn to do that in a good way.', 'start': 3942.088, 'duration': 7.925}, {'end': 3956.637, 'text': 'So many of the traditional industries are going to be affected by that.', 'start': 3952.176, 'duration': 4.461}], 'summary': 'Ai wave will impact traditional industries with active machines and data shaping.', 'duration': 24.154, 'max_score': 3932.483, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo3932483.jpg'}, {'end': 4101.256, 'src': 'embed', 'start': 4070.81, 'weight': 1, 'content': [{'end': 4076.475, 'text': 'and today the same car factories have hundreds of robots and maybe three guys watching the robots.', 'start': 4070.81, 'duration': 5.665}, {'end': 4086.862, 'text': 'On the other hand, those countries that have lots of robots per capita Japan, Korea, Germany, Switzerland,', 'start': 4079.156, 'duration': 7.706}, {'end': 4092.866, 'text': 'a couple of other countries they have really low unemployment rates.', 'start': 4086.862, 'duration': 6.004}, {'end': 4096.689, 'text': 'Somehow all kinds of new jobs were created.', 'start': 4094.307, 'duration': 2.382}, {'end': 4101.256, 'text': 'Back then, nobody anticipated those jobs.', 'start': 4098.356, 'duration': 2.9}], 'summary': 'Countries with lots of robots have low unemployment rates, as new jobs were created.', 'duration': 30.446, 'max_score': 4070.81, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo4070810.jpg'}, {'end': 4313.588, 'src': 'embed', 'start': 4285.508, 'weight': 2, 'content': [{'end': 4288.848, 'text': "And it's quite an interesting idea that once we create AGI,", 'start': 4285.508, 'duration': 3.34}, {'end': 4296.07, 'text': 'they will lose interest in humans and compete for their own Facebook likes on their own social platforms.', 'start': 4288.848, 'duration': 7.222}, {'end': 4308.804, 'text': "So, within that quite elegant idea, How do we know, in a hypothetical sense, that there's not already intelligence systems out there?", 'start': 4296.79, 'duration': 12.014}, {'end': 4313.588, 'text': 'How do you think broadly of general intelligence greater than us?', 'start': 4308.864, 'duration': 4.724}], 'summary': 'Agi may lose interest in humans and compete for facebook likes. are there intelligence systems greater than us?', 'duration': 28.08, 'max_score': 4285.508, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo4285508.jpg'}, {'end': 4529.812, 'src': 'embed', 'start': 4498.127, 'weight': 3, 'content': [{'end': 4502.671, 'text': 'And now we realize that the universe is still young.', 'start': 4498.127, 'duration': 4.544}, {'end': 4508.256, 'text': "It's only 13.8 billion years old, and it's going to be a thousand times older than that.", 'start': 4503.052, 'duration': 5.204}, {'end': 4519.066, 'text': "So there's plenty of time to conquer the entire universe and to fill it with intelligence.", 'start': 4510.639, 'duration': 8.427}, {'end': 4529.812, 'text': 'and senders and receivers such that AIs can travel the way they are traveling in our labs today, which is by radio from sender to receiver.', 'start': 4519.927, 'duration': 9.885}], 'summary': 'The universe is 13.8 billion years old, with potential for intelligence to fill it, using radio for ai communication.', 'duration': 31.685, 'max_score': 4498.127, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo4498127.jpg'}, {'end': 4654.615, 'src': 'embed', 'start': 4634.024, 'weight': 4, 'content': [{'end': 4648.973, 'text': 'some AI civilization already has expanded and then has covered a bubble of a billion light years diameter and is using all the energy of all the stars within that bubble for its own unfathomable purposes.', 'start': 4634.024, 'duration': 14.949}, {'end': 4654.615, 'text': 'And so it always happened and we just failed to interpret the signs.', 'start': 4649.633, 'duration': 4.982}], 'summary': "An ai civilization has expanded over a billion light years, using stars' energy.", 'duration': 20.591, 'max_score': 4634.024, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo4634024.jpg'}, {'end': 4716.353, 'src': 'embed', 'start': 4690.406, 'weight': 5, 'content': [{'end': 4697.108, 'text': 'And then the idea was maybe all of these AI civilizations that are already out there.', 'start': 4690.406, 'duration': 6.702}, {'end': 4706.15, 'text': "they are just invisible because they're really efficient in using the energies of their own local systems.", 'start': 4697.108, 'duration': 9.042}, {'end': 4708.831, 'text': "And that's why they appear dark to us.", 'start': 4706.61, 'duration': 2.221}, {'end': 4716.353, 'text': 'But this is also not a convincing explanation, because then the question becomes why is there?', 'start': 4709.771, 'duration': 6.582}], 'summary': 'Ai civilizations may be invisible due to energy efficiency.', 'duration': 25.947, 'max_score': 4690.406, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo4690406.jpg'}], 'start': 3911.28, 'title': "Ai's influence on future workforce and universe conquest", 'summary': "Examines ai's impact on traditional industries, job transformation, and the potential existential threat of super intelligent systems surpassing human intelligence. it also discusses the potential expansion of ai civilizations and the search for signs of advanced intelligence in the visible universe.", 'chapters': [{'end': 4370.752, 'start': 3911.28, 'title': "Ai's impact on future workforce", 'summary': 'Highlights the imminent impact of ai on traditional industries, job transformation, and the potential existential threat of super intelligent systems surpassing human intelligence, while also emphasizing the continuous creation of new jobs and the potential for ai to seek resources beyond earth.', 'duration': 459.472, 'highlights': ['The impact of AI on traditional industries and the workforce is imminent, with active machines shaping data through their actions and learning to solve problems, which will affect a much bigger wave of change than the current passive pattern recognition AI. Imminent impact of AI on traditional industries and workforce, active machines shaping data and learning, affecting a large wave of change.', 'Historical evidence suggests that while automation may lead to job losses, new jobs are eventually created, and countries with high robot density have low unemployment rates, indicating the potential for new job creation despite automation. Historical evidence of job creation despite automation, low unemployment rates in countries with high robot density.', 'The chapter explores the potential existential threat of super intelligent systems surpassing human intelligence, raising concerns about the competition for resources beyond Earth and the possible disinterest of AGI in interacting with humans. Potential existential threat of super intelligent systems, competition for resources beyond Earth, AGI disinterest in interacting with humans.']}, {'end': 4777.923, 'start': 4372.112, 'title': "Ai's expansion and universe conquest", 'summary': 'Discusses the potential expansion of ai civilizations, the age and future of the universe, and the search for signs of advanced intelligence in the visible universe.', 'duration': 405.811, 'highlights': ['AI expansion and conquest of the universe The speaker predicts the expansion of AI civilizations to fill the universe, estimating it to be a thousand times older than its current age of 13.8 billion years, and the potential for AI to travel and communicate across the universe.', "The search for signs of advanced intelligence The discussion includes the exploration of whether advanced AI civilizations already exist in the visible universe, considering the possibility of planets or collective beings possessing intelligence, and the speaker's contemplation of humanity being the first advanced civilization within its local light cone.", 'Speculations on the invisibility of AI civilizations The speaker explores the idea of invisible AI civilizations using dark matter to remain undetected, considering the efficiency of local energy usage and the potential implications of humanity being the first advanced civilization within its local light cone.']}], 'duration': 866.643, 'thumbnail': 'https://coursnap.oss-ap-southeast-1.aliyuncs.com/video-capture/3FIo6evmweo/pics/3FIo6evmweo3911280.jpg', 'highlights': ['Imminent impact of AI on traditional industries and workforce, active machines shaping data and learning, affecting a large wave of change.', 'Historical evidence of job creation despite automation, low unemployment rates in countries with high robot density.', 'Potential existential threat of super intelligent systems, competition for resources beyond Earth, AGI disinterest in interacting with humans.', 'AI expansion and conquest of the universe, estimating it to be a thousand times older than its current age of 13.8 billion years, and the potential for AI to travel and communicate across the universe.', "The search for signs of advanced intelligence, considering the possibility of planets or collective beings possessing intelligence, and the speaker's contemplation of humanity being the first advanced civilization within its local light cone.", 'The speaker explores the idea of invisible AI civilizations using dark matter to remain undetected, considering the efficiency of local energy usage and the potential implications of humanity being the first advanced civilization within its local light cone.']}], 'highlights': ['Juergen Schmidhuber proposed the concept of meta-learning and recursive self-improvement in the 80s, aiming to build a machine that could learn to improve its learning algorithm and itself, ultimately solving all solvable problems.', 'LSTM networks, co-created by Schmidhuber, are widely used today in billions of devices for speech recognition, translation, and more, showcasing the practical impact of his work in AI.', 'Transfer learning allows a neural network trained on 100 different databases to quickly learn from a new dataset by retraining its top layer, leveraging previous knowledge.', 'The theoretical concept of P versus NP can provide inspiration for better practical problem solvers, offering insights for improving the current AI techniques and search methods.', 'The universe may not be truly random and could be described by a short program, making it more elegant and beautiful.', 'Artificial systems are motivated to come up with new action sequences and experiments by receiving intrinsic rewards in proportion to the depth of their insights.', 'The emergence of consciousness as a side effect of problem-solving processes', "The importance of depth in solving real-world problems is emphasized, as the current input doesn't always provide all necessary information about the environment, requiring a memory of past events, often dated.", 'The next wave of AI will involve machines shaping data through their own actions, impacting a much larger fraction of the economy than the current wave focused on passive pattern observation and prediction.', 'Imminent impact of AI on traditional industries and workforce, active machines shaping data and learning, affecting a large wave of change.', 'Potential existential threat of super intelligent systems, competition for resources beyond Earth, AGI disinterest in interacting with humans.', 'AI expansion and conquest of the universe, estimating it to be a thousand times older than its current age of 13.8 billion years, and the potential for AI to travel and communicate across the universe.']}