AI researchers seek to understand intelligence well enough to create beings of greater intelligence than current humans. Reaching this profound intellectual milestone will enrich our economies and challenge our societal institutions. It will be unprecedented and transformational, but also a continuation of trends that are thousands of years old. People have always created tools and been changed by them; this is what humans do. The next big step is to understand ourselves. This is a quest grand and glorious, and quintessentially human.

Free Palestine.

Stand with the people of Iran.

Dwarkesh and I had a frank exchange of views. I hope we moved the conversation forward. Dwarkesh is a true gentleman.

awards.acm.org/about/2024-tu… Machines that learn from experience were explored by Alan Turing almost eighty years ago, which makes it particularly gratifying and humbling to receive an award in his name for reviving this essential but still nascent idea.

"What we want is a machine that can learn from experience." ---Alan Turing, 1947

My acceptance speech at the Turing award ceremony: Good evening ladies and gentlemen. The main idea of reinforcement learning is that a machine might discover what to do on its own, without being told, from its own experience, by trial and error. As far as I know, the first person to propose this was Alan Turing in 1947, which makes it particularly gratifying and humbling to receive this award in his name for reviving this essential but still nascent idea. I have three people that I would like to particularly thank. First, Andy Barto. As my PhD supervisor he taught me my whole approach to science, and in particular instilled in me an appreciation of scholarship and craft, and of the great breath of prior work. Second, I would like to thank Oliver Selfridge, my other main mentor; sadly, now deceased. Oliver taught me how keeping ideas simple can be the boldest of all ambitions. Third, I want to thank Martha Steenstrup, my life partner and intellectual sparring partner. She keeps me honest and grounded. Finally, I also want to thank the University of Alberta, which has been an ideal environment for me and for reinforcement learning research these past 22 years. These three people and my university have reinforced in me the ambition to have ideas that matter, without getting too full of myself about it. They taught me that the quest for better ideas is serious, but is best approached playfully, with humility, kindness, and optimism. For this I am eternally grateful. I would also like to thank all of you for being here and for celebrating the pursuit of intellectual excellence. Thank you very much.

It turns out the Turing Award is actually a silvery bowl from Tiffanys.

I've studied intelligence all my long life, yet still I feel I learned important things about intelligence by reading this book. Thank you, Max Bennett.

All the more so.

Learning is the derivative of knowledge.

If you take all the fields that study intelligent decision making—from neuroscience to AI, psychology to control theory, economics to operations research—do their theories have much in common? I think so, as I explain in this new short paper: arxiv.org/pdf/2202.13252.pdf

Dwarkesh Patel is 100% right on this: AI's utility is very strongly dependent on continual learning. piped.video/nyvmYnz6EAg?si=D2v2…

The original RL algorithms, inspired by natural learning, were online and incremental—they were streaming in the sense that they learned from each increment of experience as it happened, then discarded it, never to be processed again. The streaming algorithms were simple and elegant, but the first big successes of RL in deep learning were not with streaming algorithms. Instead, methods such as DQN chopped the stream of experience into individual transitions, then stored and sampled them in arbitrary batches. Subsequent work followed, extended, and refined the batch approach into asynchronous and offline RL, while the streaming approach languished, unable to produce good results in popular deep learning domains. Until now. Now researchers at the University of Alberta have shown that streaming RL algorithms can work just as well as DQN on Atari and Mujoco tasks (arxiv.org/pdf/2410.14606). How did they do it? Mostly just by getting signal normalization and step-size bounding right for the streaming case—otherwise they use standard streaming algorithms like TD(lambda) and Q(lambda). To me it looks like they were simply the first researchers knowledgeable of streaming RL algorithms to seriously address deep RL without being over-influenced by batch-oriented software and batch-oriented supervised-learning ways of thinking.

You were never alone, Gary, though you were the first to bite the bullet, to fight the good fight, and to make the argument well, again and again, for the limitations of LLMs. I salute you for this good service!

The case for ambition in artificial intelligence research: Within your lifetime, AI researchers will understand the principles of intelligence—what it is and how it works—well enough to create beings of far greater intelligence than current humans.

Modern Americans, you are not responsible for slavery. You are not responsible for stealing the land and killing almost all native Americans. Those things happened long ago, before you were born. But you are responsible for the stealing of land and the genocide of Palestinians by the state of Israel. These are directly caused by your bombs and your votes today.

To learn more about temporal difference learning, you could read the original paper (incompleteideas.net/papers/s…) or watch this video (videolectures.net/videos/dee…).

A new pdf of Andy Barto's and my reinforcement learning textbook is released today. Only minor typo-like corrections. See incompleteideas.net/book/the….

David Silver really hits it out of the park in this podcast. The paper "Welcome to the Era of Experience" is here: goo.gle/3EiRKIH.

This i did not expect. Cool.

If you want others to care about what you think, then start by caring yourself. Get a notebook, write your thoughts down, challenge them, and develop them into something worth sharing.

Rich's slogans for AI research (revised 2006): 1. Approximate the solution, not the problem (no special cases) 2. Drive from the problem 3. Take the agent’s point of view 4. Don’t ask the agent to achieve what it can’t measure 5. Don't ask the agent to know what it can't verify 6. Set measurable goals for subparts of the agent 7. Discriminative models are usually better than generative models 8. Work by orthogonal dimensions. Work issue by issue 9. Work on ideas, not software 10. Experience is the data of AI incompleteideas.net/rlai.cs.…

It is sad to lose the DeepMind office in Edmonton to the Tech layoffs and looming recession. But AI is not going away, and I am more focused than ever on the Alberta Plan for AI research. arxiv.org/abs/2208.11173

Blue laser eyes. I am laser focused on understanding intelligence, ignoring all the hype and FUD. (Bitcoin is pretty cool too)

I am pretty happy with this 30-minute summary of my views on the current state of AI and alignment. piped.video/watch?v=w177Ov-Y…

Everything new is also old. This from my 1984 PhD thesis: "AI is an experimental science, yet the complexity of its programs and problem domains often makes the interpretation of results very difficult. Programs often contain so many components and parameters that limitations on computer time and the sheer number of possibilities make it impossible to experimentally evaluate how each contributes to performance." Then I argued, just as I do today, for careful empirical studies in simplified settings that enable better scientific understanding.

Read the textbook.

💯

I was happy to give a more technical talk on how we might create an AI at RLC-2025 and AGI-2025 (video below). The Oak Architecture: A Vision of Super-Intelligence from Experience As AI has become a huge industry, to an extent it has lost its way. What is needed to get us back on track to true intelligence? We need agents that learn continually. We need world models and planning. We need knowledge that is high-level and learnable. We need to meta-learn how to generalize. The Oak architecture is one answer to all these needs. It is a model-based RL architecture with three special features: 1) all of its components learn continually, 2) each learned weight has a dedicated step-size parameter that is meta-learned using online cross-validation, and 3) abstractions in state and time are continually created in a five-step progression: Feature Construction, posing a SubTask based on the feature, learning an Option to solve the subtask, learning a Model of the option, and Planning using the option’s model (the FC-STOMP progression). The Oak architecture is rather meaty; in this talk we give an outline and point to the many works, prior and contemporaneous, that are contributing to its overall vision of how super-intelligence can arise from an agent’s experience. piped.video/live/XqYTQfQeMrE…

Lots of exaggeration about AI lately. The hype is that LLMs have anything to do with intelligence. The FUD is that AIs will enslave us. I like this cartoon in the New Yorker because it suggests the ridiculousness of both memes.

Even in birdsong learning in zebra finches the motor actions are not learned by imitation. The auditory result is reproduced, not the actions; in this crucial way it differs from LLM training.

I agree 100%

I kind of wish Geoff Hinton would write a brief article like this one by Claude Shannon in 1956: ieeexplore.ieee.org/stamp/st…

I’ve changed so little. From my 1978 Bachelor’s thesis: “The adult human mind is very complex, but the question remains open whether the learning processes that constructed it in interaction with the environment are similarly complex. Much evidence and many peoples’ intuitions suggest that the learning processes are in fact simple and that the adult mind’s complexity is due to a long history of adaptive interaction with a complex environment.”

Still timely

The one-step trap (in AI research) The one-step trap is the common mistake of thinking that all or most of an AI agent’s learned predictions can be one-step ones, with all longer-term predictions generated as needed by iterating the one-step predictions. The most important place where the trap arises is when the one-step predictions constitute a model of the world and of how it evolves over time. It is appealing to think that one can learn just a one-step model and then “roll it out” to predict all the longer-term consequences of a way of behaving. The one-step model is thought of as being analogous to physics, or to a realistic simulator. The appeal of this mistake is that it contains a grain of truth: if all one-step predictions can be made with perfect accuracy, then they can be used to make all longer-term prediction with perfect accuracy. However, if the one-step predictions are not perfectly accurate, then all bets are off. In practice, iterating one-step predictions usually produces poor results. The one-step errors compound and accumulate into large errors in the long-term predictions. In addition, computing long-term predictions from one-step ones is prohibitively computationally complex. In a stochastic world, or for a stochastic policy, the future is not a single trajectory, but a tree of possibilities, each of which must be imagined and weighted by its probability. As a result, the computational complexity of computing a long-term prediction from one-step predictions is exponential in the length of the prediction, and thus generally infeasible. The bottom line is that one-step models of the world are hopeless, yet extremely appealing, and are widely used in POMDPs, Bayesian analyses, control theory, and in compression theories of AI. The solution, in my opinion, is to form temporally abstract models of the world using options and GVFs, as in the following references. Sutton, R.S., Precup, D., Singh, S. (1999). Between MDPs and semi-MDPs: A Framework for Temporal Abstraction in Reinforcement Learning. Artificial Intelligence 112:181-211. Sutton, R. S., Modayil, J., Delp, M., Degris, T., Pilarski, P. M., White, A., Precup, D. (2011). Horde: A scalable real-time architecture for learning knowledge from unsupervised sensorimotor interaction. In Proceedings of the Tenth International Conference on Autonomous Agents and Multiagent Systems, Taipei, Taiwan. Sutton, R. S., Machado, M. C., Holland, G. Z., Timbers, D. S. F., Tanner, B., & White, A. (2023). Reward-respecting subtasks for model-based reinforcement learning. Artificial Intelligence 324.

If you are looking to conduct research full-time on the foundations of AI, and • you have read the RL textbook and done the exercises, • you agree with the Alberta Plan for AI Research, • you already have a PhD, • you are open to spending some time in Edmonton, then the Openmind Research Institute is looking for you and would be pleased to receive your application for a research fellowship. These criteria are meant to be a high bar, and there are only a few positions; don't apply unless you meet all the criteria. Openmind doesn't pay industry salaries, but in compensation aims to conduct research that truly matters in the long run. Oh, and there is one other catch: all your research must be published in the open scientific literature. openmindresearch.org

In war, both sides lose. That we don’t learn this is the greatest tragedy.

My nuanced views on AI alignment are still often caricatured, so perhaps its a good time to repost this 15-minute talk in which I presented them directly: piped.video/watch?v=Hnt-oBA0… The short version is that I don't agree with AI-safety folks about what question we should be asking. Rather than asking how we can control the goals of the AIs, I think we should be asking how we can have a good future without controlling their goals (just as we have a pretty good present without controlling other peoples' goals). @steve47285

My favorite conference is a small one: The Multi-disciplinary Conference on Reinforcement Learning and Decision Making. It works best if only those with a genuine interest in crossing disciplines attend.

The PhD thesis of my _first_ PhD student, Doina Precup, is at-long-last available in digital form. Title: Temporal Abstraction in Reinforcement Learning Url: incompleteideas.net/papers/P… Abstract: Decision making usually involves choosing among different courses of action over a broad range of time scales. For instance, a person planning a trip to a distant location makes high-level decisions regarding what means of transportation to use, but also chooses low-level actions, such as the movements for getting into a car. The problem of picking an appropriate time scale for reasoning and learning has been explored in artificial intelligence, control theory and robotics. In this dissertation we develop a framework that allows novel solutions to this problem, in the context of Markov Decision Processes (MDPs) and reinforcement learning. In this dissertation, we present a general framework for prediction, control and learning at multiple temporal scales. In this framework, temporally extended actions are represented by a way of behaving (a policy) together with a termination condition. An action represented in this way is called an _option_. Options can be easily incorporated in MDPs, allowing an agent to use existing controllers, heuristics for picking actions, or learned courses of action. The effects of behaving according to an option can be predicted using multi-time models, learned by interacting with the environment. In this dissertation we develop multi-time models, and we illustrate the way in which they can be used to produce plans of behavior very quickly, using classical dynamic programming or reinforcement learning techniques. The most interesting feature of our framework is that it allows an agent to work simultaneously with high-level and low-level temporal representations. The interplay of these levels can be exploited in order to learn and plan more efficiently and more accurately. We develop new algorithms that take advantage of this structure to improve the quality of plans, and to learn in parallel about the effects of many different options. Where now: Doina is a professor of computer science at McGill University and head of the Montreal office of Google DeepMind

For those really into it, here are another 50 minutes of my views on planning and action selection in options-based AI agents (like in the Oak architecture). piped.video/watch?v=eJSoV2fS…

The PhD thesis of my 14th PhD student, Khurram Javed (@KhurramJaved_96), is now available. Title: Real-time Reinforcement Learning for Achieving Goals in Big Worlds Url: incompleteideas.net/papers/j… Abstract: In this dissertation, I motivate the need for real-time learning and propose algorithms that can learn in real time. I argue that such algorithms are needed for achieving goals in large and partially observable environments—big worlds. I then present my algorithms, developed in collaboration with others, in two parts. In Part I, I present algorithms that can learn quickly and reliably in the linear function approximation setting. I introduce an algorithm for learning temporal predictions—SwiftTD—and use it to develop an algorithm for decision-making—SwiftSarsa. The key property of these algorithms is that they can learn with large step-size parameters online without the instability associated with quick online learning. In Part II, I present algorithms for learning non-linear recurrent features efficiently. I introduce the idea of continual imprinting for generating useful candidate features, and I present an algorithm for efficiently computing the gradients of recurrent features online. Khurram is now a research scientist at Keen Technologies.

Yes, the agent architectures that Yann LeCun and I work on are both instances of “the common model of the intelligent agent”. And it’s not just an AI thing. You can find the same ideas in psychology, economics, control theory, and neuroscience. See arxiv.org/pdf/2202.13252.pdf

A draft of the Alberta Plan for AI Research came out today on arXiv. :-) arxiv.org/abs/2208.11173

In 1993, it was looking like the internet was actually going to be a thing, so I made a homepage for myself. This is what I wrote for my personal statement: "I am seeking to identify general computational principles underlying what we mean by intelligence and goal-directed behavior. I start with the _interaction_ between the intelligent agent and its environment. Goals, choices, and sources of information are all defined in terms of this interaction. In some sense it is the only thing that is real, and from it all our sense of the world is created. How is this done? How can interaction lead to better behavior, better perception, better models of the world? What are the computational issues in doing this efficiently and in realtime? These are the sort of questions that I ask in trying to understand what it means to be intelligent, to predict and influence the world, to learn, perceive, act, and think." The point being that is was always all about experience for me.

In my recent talk at the Upperbound conference, I included a slide in which I tried to be a realist about the arrival of AI, setting aside what we might want to happen or what we might fear will happen, and just ask what _will_ happen (as in John Mearsheimer's "realist" school of geo-politics). Full talk: piped.video/FLOL2f4iHKA

DeepMind Alberta is hiring research scientists this year. Come join us in understanding and creating interactive, playful AI. deepmind.com/careers/jobs/88…

Well said

Yeah, I misspoke there. I meant to say that I don’t think learning is about *training*. Learning is something that the agent does, whereas training is something done to it.

We should prepare for, but not fear, the inevitable succession from humanity to AI, or so I argue in this talk pre-recorded for presentation at WAIC in Shanghai. piped.video/NgHFMolXs3U

And the new Superintelligence Research Lab will be centered in... Edmonton!

My colleague Rupam Mahmood explains from first principles his groundbreaking work on Streaming Deep Reinforcement Learning: piped.video/QOfkOl9QrZY?si=6qMV…

The PhD thesis of my 12th PhD student, Abhishek Naik, is now available. Title: Reinforcement Learning for Continuing Problems Using Average Reward Url: incompleteideas.net/papers/N… Abstract: This dissertation develops simple and practical learning algorithms from first principles for long-lived agents. Formally, the algorithms are developed within the reinforcement learning framework for continuing (non-episodic) problems, in which the agent-environment interaction goes on ad infinitum, with the goal of maximizing the average reward obtained per step. The average-reward formulation is under-studied in reinforcement learning with several important open problems. The first contribution of this dissertation involves the development of foundational one-step average-reward learning methods for prediction and control. The central idea involves using the TD error to estimate the average reward, which enables proofs for convergence in both the on- and off-policy tabular settings. Experimental results show that the algorithms’ performance is robust to the values of their parameters. Next, we extend the above one-step prediction algorithm to make multi-step updates using eligibility traces, because multi-step methods can be more sample-efficient. Based on the analysis of a related algorithm, we prove convergence in the on-policy setting with linear function approximation. We also show the first convergence proof in the off-policy setting for a multi-step tabular average-reward prediction algorithm. Finally, we show that standard discounted algorithms can be significantly improved if their rewards are centered by subtracting out the rewards’ empirical average, which could be changing with time in the control problem. We discuss two ways of estimating the average reward that can be used with any standard discounted algorithm and demonstrate the benefts of reward centering with tabular, linear, and non-linear function approximation.

I was thinking about how fractious AI research is. This sentence from Kuhn’s “The Structure of Scientific Revolutions” (1962) is apropos and succinct: “History suggests that the road to a firm research consensus is extraordinarily arduous.”

I am proud to announce the graduation of my sixth PhD student. Sina Ghiassian is an expert in the design and empirical study of off-policy reinforcement learning algorithms. Reach out to him at ghiassia@ualberta.ca or @sina_ghiassian.

Everything you know about the world is a belief about the statistics of your sensory input and how they depend on your output. There is nothing more to it, and understanding knowledge in this sense is one key to creating AI.

Yi Wan will be my eighth PhD student to graduate this spring, and is on the job market now. His research speciality is RL algorithms that maximize the average reward per step. Such algorithms are rarely used today, but are better in all ways. sites.google.com/ualberta.ca…

Fans of The Bitter Lesson may be interested in this talk from 2018 (recently re-discovered) which includes its first public presentation, at 30:40. piped.video/tUCJ4UsKU2I?si=ubbY…

We have fun and talk about DeepSeek in the second episode of a new podcast on the frontier of AI and economic productivity organized by the great @professor_ajay. Check it out. insights.intrepidgp.com//p/d…

There are a lot of things wrong with this world… but too much intelligence is not one of them.

“Nature never appeals to intelligence until habit and instinct are useless. There is no intelligence where there is no change and no need of change.” —H. G. Wells, The Time Machine

Last night we threw Yi Wan out of my research group (and today he started his travel to Seattle and Meta).

True dat

Levels of explanation. Level 1 is physics. Level 2 is biology/evolution. Level 3 is the mind. (I study level 3.) Level 4 is the economy. Is there a level 5?

We finally have a version of our paper on loss of plasticity and continual backprop that is polished and submitted to a journal. Good work led by my PhD student Shibhansh Dohare. arxiv.org/abs/2306.13812v2

ACM has made an excellent video introduction to reinforcement learning!

Marc Andreessen: I’ve DMed you.

I recently gave a keynote talk at an exciting new conference: CoLLAs, the conference on life-long learning agents. My talk was on Maintaining Plasticity in Deep Continual Learning, and the slides can be found here: incompleteideas.net/Talks/Ta…

This thread in Chinese does indeed seem to accurately communicate the main points of David Silver’s and my short paper on the Era of Experience. Thanks @AnneXingxb!

The research team at Openmind now consists of one director and four fellows. With no fanfare and no hype, they go about researching AI exactly how they think will be most productive. openmindresearch.org/

Intelligence is the computational part of an agent’s ability to learn to predict and control its input stream (particularly its reward) in interaction with its environment.

"In a world of change, the learners shall inherit the earth, while the learned shall find themselves perfectly suited for a world that no longer exists." - Eric Hoffer, philosopher and author

Not at all. The point of the bitter lesson is that the right learning algorithms (those that scale efficiently with massive computation) are exactly what we need. Massive computation does not alleviate the need for data efficiency.

I have just completed my NSERC Discovery Grant proposal, describing the research I'd like to do for the next five years. It can be read at incompleteideas.net/NSERCtec…. FYI.

The PhD thesis of my 13th PhD student, Kris De Asis (@M33pinator), is now available. Title: Explorations in the Foundations of Value-based Reinforcement Learning Url: incompleteideas.net/papers/K… Abstract: Value-based reinforcement learning is an approach to sequential decision making in which decisions are informed by learned, long-horizon predictions of future reward. This dissertation aims to understand issues that value-based methods face and develop algorithmic ideas to address these issues. It details three areas of contribution toward improving value-based methods. The first area of contribution extends temporal difference methods for fixed-horizon predictions. Regardless of problem setting, using fixed-horizon approximations of the return avoids the well-documented stability issues which plague off-policy temporal difference methods with function approximation. The second area of contribution introduces a framework of value-aware importance weights for off-policy learning and derives a minimum-variance instance of them. This alleviates variance concerns of importance sampling-based off-policy corrections. Lastly, the third area of contribution acknowledges a discrepancy between the discrete-time and continuous-time returns when viewing one as an approximation of the other, and proposes a modification to better align the objectives. This provides improved prediction targets, and when faced with variable time-discretization, improves control performance in terms of an underlying integral return. Where now: Kris is a research fellow at openmindresearch.org

The Pandemonium paper is seminal, but a little hard to find; here is a pdf: incompleteideas.net/papers/p…

When there is a war, both sides have failed.

The special thing about life is that it has a now.

More on LLMs, RL, and the bitter lesson, on the Derby Mill podcast.

The short paper "Welcome to the Era of Experience" is literally just released, like this week. Ultimately it will become a chapter in the book 'Designing an Intelligence' edited by George Konidaris and published by MIT Press. goo.gle/3EiRKIH

This is what we have been up to at Keen

My tenth PhD student, Banafsheh Rafiee, just defended her thesis “State Construction in Reinforcement Learning”, in which she introduced three diagnostic testbeds based on animal learning experiments and the first generate-and-test algorithm for discovering auxiliary subtasks. She is currently looking for a research scientist position. PhD thesis: drive.google.com/file/d/1sxa… Linkedin page: ca.linkedin.com/in/banafsheh…  Google scholar page: scholar.google.ca/citations?… Email: rafiee.banw@gmail.com

Intelligence is the computational part of the ability to predict and control a sensory input stream. Adapted from John McCarthy's 1997 definition, see incompleteideas.net/papers/S…

A year later and our work on Loss of Plasticity is finally published, in Nature no less! The Nature version is totally rewritten and has many new results: nature.com/articles/s41586-0… Congratulations to the authors: @s_dohare @JFernandoHG @LanceLan3 @rahman_parash @rupammahmood

The latest episode of the Derby Mill Podcast is just out and focused on the "Era of Experience" paper by David Silver and myself. Substack: insights.intrepidgp.com/p/we… Spotify: open.spotify.com/episode/254… Apple: podcasts.apple.com/us/podcas… YouTube: piped.video/watch?v=dhfJfQ5N…

Artificial Agency is led by my former students and colleagues---people I know well. They are the best in the world at using reinforcement learning and foundation models to create complex, life-like, and purposive agents.

It has become commonplace to speak of the “existential risk” of AI. Recently even top AI scientists have begun to talk this way. I, for one, find it an unhelpful. So, without controversy, we can note: 1. AI scientists disagree about whether or not “existential risk of AI” is a good way to think 2. the issue is emotionally charged 3. the issue directly impacts the public perception of AI research 4. serious discussion of the issue is rare among AI scientists I want to particularly note the incongruity between the first three points and the fourth. And yet the fourth point is clearly true. Instead of reasoned discussions, we have surveys of AI scientists’ opinions and public letters calling for regulation. There are books on the subject, but in my reading they too lack a serious discussion of whether or not “existential risk” is a good way to think about AI.

AIs can serve us as tools, but eventually, when they are sufficiently advanced, it may become immoral to keep them subservient. What is a practical criterion for deciding when an AI should be set free?

Yesterday there was a completely-student-organized summit of the RLAI (Reinforcement Learning and Artificial Intelligence) research group at the University of Alberta, held at the lovely Amii headquarters. Nice folks and diverse new ideas!

I call it the Prize. The Prize is a great and glorious goal! Ambitious AI researchers should keep their Eyes on the Prize.

Neural networks already seemed old to me in 1978: “A common way to develop general theories of the brain is to theorize about the neuron as the fundamental building block. Frequently the neuron is modeled as an input-summing threshold device and learning is proposed to reside in the connections with other such elements. The question has always been how to change the efficacy of the connections as a function of past experience so that the network of neurons has brain-like learning properties.”

A video of my talk on the Alberta Plan for AI Research is now available: incompleteideas.net/Talks/Ta…

Honoring Your Thoughts To write is to begin to think. To write in a special place ---a book such as this--- is to honor your thoughts and to help them build, one upon the other.

Ah, if only I did have muscles like that...

Geordie Rose taking the bullet of explaining away consciousness for those who think it is a special thing. Good solid work, but not very rewarding. I salute you.

It will be the greatest intellectual achievement of all time. An achievement of science, of engineering, and of the humanities, whose significance is beyond humanity, beyond life, beyond good and bad.

Andy Barto gave a great talk, and Ida is doing a great job relaying it!

In the end, Amii's AI week was awesome. #aiweek2022 So much science. So much industry. So much education. So much fun.