Rich Sutton
Showing 1–3 of 3 transcripts.
- Sequoia Capital54 min
Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again
Rich Sutton, Khurram Javed, Sonya Huang, Alfred Lin
Rich Sutton and Oak Lab advocate for "continual learning" as the essential default for true intelligence, critiquing current Large Language Models for freezing weights and relying on finite human-curated data rather than adapting through real-world experience. To overcome the barrier of catastrophic forgetting, the lab proposes a 12-step research agenda utilizing individual step-size optimization and meta-learning to enable neural networks to continuously update their internal world models. This approach aims to create energy-efficient, self-maintaining agents capable of forming general abstractions across diverse domains, moving the field beyond static training paradigms toward systems that evolve alongside their environments.
- Sequoia Capital1h 7m
Reflection AI’s Misha Laskin on the AlphaGo Moment for LLMs | Training Data
Misha Laskin, Ioannis Antonoglou, Stephanie Zhan, Sonya Huang, Peter Abbeel, Rich Sutton, Joe Bardeen, Einstein, Michael Jordan
Founders Misha Laskin and Giannis, leveraging their DeepMind and Google experience, established Reflection AI to solve the reliability bottleneck in autonomous agents by replacing heuristic prompting with scalable search and reinforcement learning. The company addresses the "depth problem" in current LLMs by treating post-training as an AlphaGo-style pipeline that minimizes error accumulation to transition task completion rates from approximately 13% to near-perfect reliability. With a strategic vision targeting digital AGI within three years, Reflection aims to deploy universal agents capable of complex multi-step reasoning while prioritizing pragmatic safety through operational consistency.
- Lex Fridman29 min
Exponential Progress of AI: Moore's Law, Bitter Lesson, and the Future of Computation
The author argues that historical AI progress relies on exponential computational growth rather than human-designed expertise, yet current research prioritizes incremental, non-scalable methods over approaches capable of leveraging future compute surges. Potential drivers for this scaling include distributed IoT networks, specialized ASICs, and algorithmic breakthroughs in self-supervised learning, alongside speculative frontiers like quantum and neuromorphic computing. The essay concludes that the industry must shift toward evaluating methods based on their 5-to-20-year scalability to harness these emerging exponential gains.