GPT summary of comments: Rich Sutton, a leading researcher in reinforcement learning, advocates for a shift in AI research toward **continual learning**, criticizing the field's reliance on transient, task-specific models. He highlights the limitations of static approaches and the need for AI systems to adapt dynamically, like humans. Sutton defines intelligence as trial-and-error learning, emphasizing adaptive systems capable of updating and retaining knowledge over time. He critiques current methods such as replay buffers and normalization, which hinder continual learning. Sutton’s vision focuses on creating AI systems that model real-world transitions, adapt to evolving tasks, and redefine intelligence, blending human and machine capabilities.

You guys are genius. Please bring it market as soon as possibble. I cant wait for this treatment.

Most young adults' floaters are near-retina (less than 1mm) and are rather small. They are most devastating for the patient, because they cast the sharpest and brightest shadows onto the retina. How would the approach deal with these?

??

Please answer ... But what if it hits to the retina and its cells got ionised

You guys are absolute heroes. This is real trailblazing that is going to help so many people. Godspeed

Ojalá pronto haya una cura para estás moscas que generan estrés y ansiedad

Thanks for the video, I have high hopes that one day I will be floterless without victrctomy or yag.

love you dr I hope You will help❤❤❤❤ us suffering with floaters

Is Mr. Sutton aware of the Thousand Brains Project by Numenta? They focus heavily on using cortical models of mammalian brains to process inputs (i.e. sample environment states and rewards) specifically to do continual learning as he emphasizes heavily in this discussion. I'd love to sit in the room for a conversation with Jeff Hawkins and Richard Sutton.

Non biological entities do not have the biological drives at the FLESH CELLULAR LEVEL:AIR, CALORIES, PLEASURE, FEAR of DEATH,..etc. Machines dont efing learn dummies, they just make humans happy that they "appear" (did what humans want them to do to)/ convince humans they are "LEARNING". Take a great detailed photo of a brain down to the cellular level and press CONTINUE on the scan. Thats about it, but you wont be programming it with human hands in scripts. Its insane you are wasiting time on this.

One of my research ideas is to combine reinforcement learning with auto-regressive learning (eg. the training of LLMs and auto-encoders). But I have not gotten very far in this direction... wonder if others have done similar research?

His view that there is one unified goal, driven by a single “reward” or motivational signal, is a deep assumption that people should actively push back on with new research. How does an agent know why a reward signal is valuable. How does an agent adjudicate between many possible values. These things won't be resolved from his perspective because reward doesn't have semantics, and it has already been decided to be relevant to the agent, thus the agent plays no role in the development of its values. Keep in mind that sensory data are also scalars, but we don't necessarily interpret it as reward. So how do we know what matters? This isn't answered.

It is definitely better than first time . 😄

I think the AI field is so high on current methods that they just can't grasp Sutton's position. Proof of that is how often they get wrong the bitter lesson. Doesn't matter if you call it a pipeline, where all that enter are "piped" in one direction from the start, or an industrial complex, where all tech companies fight to be nearest that one chair they are circling, or as a superhighway to nowhere. Even AGI, once meant to be the return to the basics of AI research, is just a smudge on the marketing blob seeking unlimited funding. It doesn't matter that people talk about how soon human-level thinking gets here, because no one in industry is on the right path. It will be a black swan moment that big tech is going to demand Gov take control of for them, else, they get zero return on half a trillion in bad bets.

Enjoyed every slide and every minute of your talk. I am from the Building Digital Twin domain, but every insight you shared and every point you made was sheer AI brilliance. Well done, Afzal! Looking forward to hearing more of your webinars and seminars in person.

Notice the correlation between uncomfortable concepts outside of our current realm of feasibility being discussed and Sutton's eyes lighting up. This is someone who is truly invested & passionate about advancing our very way of being.

Curb your Ai-thusiasm.

Beautiful words, 💯

Great talk❤

I’m submitting research proposals now, so the last 10 minutes were incredibly helpful. Thanks!

thanks

Do more. Talk less.

finally someone who tells the truth unlike the many who almost claim that AGI is just around the corner.

One thing interesting about the mind is that it seems to figure things out without conscious effort. Maybe the main defining principle that could explain what’s happening is the principal of least action. Seems like that would make sense from the stand point of conservation of energy and the primary objective of the mind to help solve problems, find patterns and find optimizations... Just an observation.

The volume is so low. After trying to increase it for the 4th time unsuccessfully on my speaker, I gave up and switched to another podcast. Which then or course more my ears away. Maybe don't put your volume crazy low?

great!

I agree with him. But solving this problem is also really difficult

**Title**: Rich Sutton’s New Path for AI | Approximately Correct Podcast **Chapter 1: Foundations and Early Inspirations** *Timestamp: **00:00** - **03:05* - Rich Sutton discusses his initial interest in **systems that interact with the world** and learn from it. - Early AI systems lacked **goal-oriented behaviors**, focusing instead on **pattern recognition**. - Sutton describes his drive to create a **goal-based learning framework** and how it led him to contribute to **reinforcement learning (RL)**. - Mentions "Bandits" as an early attempt but points out its limitations due to its **stateless approach**. **Chapter 2: Shift from Goal-Orientation to Pattern Recognition** *Timestamp: **03:06** - **05:53* - The evolution of AI research led to a preference for **supervised learning**, which is simpler and more predictable. - Sutton observes that this shift sidelined **interactive, goal-seeking AI approaches** in favor of **pattern-based models**. - He explains the early trade-offs in AI: **linear mappings** allow adaptability, but **nonlinear mappings** (achieved via backpropagation) restrict continuous learning. - This trade-off, while powerful, has hindered **continual learning capabilities**. **Chapter 3: Present Challenges in Reinforcement Learning** *Timestamp: **05:54** - **08:09* - Sutton critiques the field’s **narrow focus on transient learning**, with a trend of building frozen models for specific tasks. - He compares current AI development to **looking under the streetlamp**, where researchers focus on **what AI can currently do**, neglecting its limitations. - Highlights that **continual learning** remains underexplored, especially as **deep learning** has dominated the field. **Chapter 4: Importance of Continual Learning and Its Definition** *Timestamp: **08:10** - **10:13* - Defines **continual learning** as systems that continuously learn and adapt in real-world environments. - Contrasts it with **transient learning**, where models are trained once and remain static. - Discusses the need for **adaptive, flexible AI** to parallel human learning, retaining and updating knowledge over time. **Chapter 5: Technical Obstacles to Continual Learning** *Timestamp: **10:14** - **12:34* - Sutton explains that AI research has developed techniques (e.g., **replay buffers, normalization, early stopping**) optimized for transient learning. - These methods create **barriers to adopting continual learning** in standard AI benchmarks (ImageNet, Atari). - Sutton calls for a **paradigm shift** to tackle real-world problems that require adaptability. **Chapter 6: A Contrarian Path and Personal Motivation** *Timestamp: **12:35** - **14:53* - Sutton criticizes the field’s focus on transient learning and feels a need to address **the limitations in current AI**. - Embraces a **contrarian approach** by choosing to focus on continual learning, even if it’s less popular or supported. - Expresses disappointment that nonlinear, continual learning methods remain undeveloped, requiring him to take the initiative. **Chapter 7: Defining Intelligence and Future Directions** *Timestamp: **14:54** - **19:26* - Sutton views intelligence as **trial-and-error learning**, creating **models of the world** to achieve specific goals. - Defines the mind as a system capable of **modeling transitions, planning, and operating at multiple abstraction levels**. - Suggests that by 2030, there’s a **25% chance we will understand intelligence enough** to create truly adaptive systems. **Chapter 8: Broader Implications and Human Connection** *Timestamp: **19:27** - **24:50* - Explores the idea of **one unified goal**, driven by a single “reward” or motivational signal, which drives complex behaviors and abstractions. - Predicts AI’s understanding of goals will challenge human perceptions of **motivation and identity**. - Discusses how insights into intelligence might **blur distinctions between human and animal intelligence**. **Chapter 9: Long-Term Impact and Future of AI in Society** *Timestamp: **24:51** - **28:30* - Predicts profound changes in **sociology, technology, and education** as we understand intelligence better. - Believes understanding how minds work could lead to **augmentation**, enhancing human cognition. - Emphasizes that understanding minds will have consequences beyond AI, affecting **self-perception** and **society at large**. **Chapter 10: Advice for Researchers and Students** *Timestamp: **28:31** - **34:36* - Sutton advises keeping a **daily research notebook** to document ideas and work through confusing thoughts. - Stresses the importance of **choosing research topics** based on personal interest and foundational value rather than popularity. - Suggests balancing multiple research ideas to adapt as some ideas may fail, fostering **resilience and flexibility** in research pursuits.

Next-word prediction is a goal.

Where I can find some of his publications? I want cite some of this in my research project.

He knows what he is talking about.

👌👌

Enjoyed it!

i would have loved to hear something about his collaboration with John Carmack

Thank you for your insights! Time to overtake transient learning with continual learning

Those research tips at the end are golden.

Loved this quote from Rich: "A good way to lose is to convince other people they should do what you think is important"

Sutton is one of the most forward-thinking researchers I've ever seen in AI.

I could only dream of a company that would be willing to help me and gave a life. I have bone death in 4 joints 2 in each leg, no surgeons yet are willing to do joint replacements and I'm facing double disarticulation of my legs. I can't make money, due to several severe condtions. If anyone can connect ne with someone to help me, would be a dream so I can have a life. I am beyond hopeless. I'm scared to death of what my life will become, and I need help. I have gotten screwed over, and I really need someone to give me hope to have a life again. I have a huge story, I would be a power guy for any company at this point. I just need help and any hope. If anyone can help connect me with a company to figure out somthing as a "prototype" that I could possibly keep as I can't make any money due to severe illnesses, would be a dream. I need help, I can't afford things, I don't know what to do. I'm 25, and the last 8 years have been hell. I have no clue when I will loose my legs. But If I can find a company willing to use me like a guinie pig, and eventually have my own robotic legs would be a dream. I figure if I can connect with a company before loosing my legs, I feel like I would have the best chance of a ai robotic limb to work while I have the most connections still connected to my current legs would be effective. So if anyone can help me with a Connection, would be a dream come true.

You are the best prof

😍👌🏻✨️

did not know that sutton is at John Carmack's startup

But how are we going to prevent interference to old data when we change the backbone? IMO, the major issue of using a dynamic architecture is, when the fringe joins the backbone, it not only provide you the capacity to deal with new data, but also change the function mapping for all the old data, this change can be catastrophic especially when we're doing this with a very long temporal sequence, and some early data may take a long time to appear again. I guess some form of information of the old data is still needed, e.g. replay buffer or a bayesian posterior of weights.

Great thinking. But, doesn't this reverse regularization and tend to overfit?

Exellent

Very good, ❤

Why he laughing like that

Cool!

Will you discuss self-play RL in the recent lecture?