Skip to content

By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Sequoia CapitalSequoia Capital

Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again

Rich Sutton, who helped pioneer reinforcement learning and wrote the seminal AI essay The Bitter Lesson, has now cofounded Oak Lab with his former student Khurram Javed.

Rich SuttonguestSonya HuanghostAlfred LinhostKhurram JavedguestKhurram Javedguest
Aug 18, 202653mWatch on YouTube ↗

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

At a glance

WHAT IT’S REALLY ABOUT

Sutton and Javed’s case for continual learning beyond frozen LLMs

  1. Rich Sutton frames intelligence as continual learning from ongoing interaction, and argues the field became “weird” by treating post-deployment learning as optional rather than fundamental.
  2. He summarizes The Bitter Lesson as a call to avoid relying on human knowledge injection and instead build methods that improve as computation scales, while still allowing prior knowledge and learning to be “friends.”
  3. Rich Sutton and Khurram Javed criticize synthetic data and heavy simulation pipelines as human-bottlenecked and inevitably incomplete relative to the real world, advocating learning directly from experience.
  4. They argue today’s mainstream assistants largely stop learning because their weights do not change during use, and that context/memory features are not a substitute for continual weight updates.
  5. Oak Lab’s agenda centers on the Alberta Plan, especially continual deep learning and learned abstractions for model-based planning, with proposed algorithmic tools to prevent catastrophic forgetting and sustain open-ended learning.

IDEAS WORTH REMEMBERING

5 ideas

“Continual learning” should just be “learning.”

Rich Sutton argues that the normal way to think about intelligence is that an agent acts and learns continuously; the need to label it “continual learning” reflects how unusual it is that many deployed AI systems don’t keep learning once released.

The Bitter Lesson: prioritize methods that scale with computation, not human input.

Sutton’s compressed version emphasizes avoiding methods that depend on hand-injected human knowledge and instead building learning/search methods whose effectiveness increases with more compute.

LLMs are both a positive and negative example of the Bitter Lesson.

Sutton says LLMs exemplify scaling with compute via ingesting huge amounts of text, but he also claims they become limited because the internet is finite and the world is far larger than stored text.

Synthetic data doesn’t remove the human bottleneck or capture the “big world.”

Sutton calls synthetic data a “big mistake,” and Javed argues it remains bottlenecked by human expertise because humans must decide what to generate and evaluate; both emphasize that simulators are small compared to the complexity of the real world.

Most assistants don’t truly learn at deployment time because the weights don’t change.

Both Sutton and Javed point to a core limitation: most LLM assistants don’t keep learning during use because their weights are frozen; they distinguish this from context-based adaptation, which changes state but not the underlying learned structure.

WORDS WORTH SAVING

5 quotes

You wouldn't have to say continual learning 'cause it wouldn't make any sense to talk about learning that wasn't continual. All learning is continual.

— Rich Sutton

Don't be distracted by human knowledge, as AI traditionally has been many times. Instead, focus on learning methods that will scale with computation.

— Rich Sutton

No, that's- ... that's just a big mistake.

— Rich Sutton

Their weights never change.

— Rich Sutton

But it's totally curable.

— Rich Sutton

Read the available transcript
Reinforcement learning as a route to understanding mindsThe Bitter Lesson and scaling with computationLLMs as frozen systems vs experiential learnersSynthetic data critique and the Big World HypothesisContinual learning and catastrophic forgettingStep-size meta-learning and generate-and-test (continual backprop)Alberta Plan: abstractions, models, and planning

SPEAKERS

  • Rich Sutton

    guest

    Reinforcement learning pioneer and co-author of the seminal RL textbook, known for the Bitter Lesson and the Alberta Plan.

  • Sonya Huang

    host

    Sequoia Capital partner and podcast host/interviewer focused on AI and technology.

  • Alfred Lin

    host

    Sequoia Capital partner who co-hosts interviews on technology and AI.

  • Khurram Javed

    guest

    AI researcher focused on continual learning and agent learning from experience; collaborator of Rich Sutton.

  • Khurram Javed

    guest

    AI researcher in continual learning and agent learning from experience; discusses the Big World Hypothesis and building a new company (Oak).

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.

iOSAndroidClaudeChatGPTChrome