Frage an Gemini:
hi Gemini, I just saw a video of Rich Sutton and oak. Are you familiar with that?
Yes, absolutely. You're likely referring to the buzz around Richard Sutton and his newly announced Oak Lab (co-founded with Khurram Javed in July 2026), or his keynotes on the OaK Architecture (which stands for Options and Knowledge).
Sutton—often called the father of reinforcement learning (RL)—has made a massive, contrarian pivot by leaving John Carmack’s Keen Technologies to build this new lab. It is essentially a direct challenge to the dominant Large Language Model (LLM) scaling paradigm.
FourWeekMBA+ 1
What is the OaK Architecture?
Instead of training massive static models on curated, human-generated internet datasets (which Sutton argues is a dead end for true general intelligence), the OaK architecture focuses on continual learning from real-time, first-person experience.
At its core, the architecture relies on three distinct pillars:
Continual Learning: Every single component of the agent learns and updates continuously at runtime. There is no separate "training phase" and "inference phase."
MIT CSAIL
Meta-Learned Step Sizes: Each learned weight has its own dedicated step-size parameter, which is meta-learned using online cross-validation to optimize how fast the agent adapts.
MIT CSAIL
The FC-STOMP Progression: This is how the agent builds abstract state and temporal representations of the world from scratch without human intervention:
Feature Construction: Creating new features from raw input.
SubTask: Posing a localized goal based on those features.
Option: Learning a policy (a set of actions) to solve that subtask.
YouTube
Model: Creating a world-model transition chunk of how that option behaves.
Planning: Using that option's model to plan future actions.
The "Holy Grail" of Oak Lab
Sutton's target with Oak Lab is incredibly ambitious: building a brain-like, trillion-parameter agent that plans in real time while drawing only 20 watts of power (roughly the power consumption of a human brain or a dim lightbulb).
BigGo Finance
It represents a beautiful return to the core principles of his famous Bitter Lesson essay—leveraging pure computation and general-learning methods—but focused entirely on local, real-time interaction rather than brute-force pre-training.
Frage Gemini:
Did you watch his keynote from the AGI conference, or was it a more recent discussion of his split from Keen Technologies to launch the lab?
explainx.ai
Want to look closer at how the FC-STOMP loop actually builds abstract options?
Meine Antwort:
Yes
I saw a video of the reinforded learning group lecture. Do you think, what he says is a valid method to build an AI?
Geminis Antwort:
To answer whether Richard Sutton’s OaK (Options and Knowledge) approach is a "valid" way to build artificial general intelligence (AGI), we have to look at one of the biggest philosophical divides in AI today.
Sutton is essentially waging a war against frozen models. His approach is not just valid; it is theoretically one of the most elegant and biologically plausible blueprints we have. However, executing it in the real world faces massive practical engineering hurdles.
TNW
Here is a breakdown of why his method is incredibly compelling—and where the critics think it might hit a wall.
Why Sutton's Method is Highly Valid (The Case for OaK)
Sutton’s core thesis is grounded in what he and his co-founder, Khurram Javed, call the "Big World Hypothesis". This is the idea that the universe is infinitely complex, and no agent—no matter how many GPU clusters you throw at it—can pre-learn everything in a static training phase.
BetaKit+ 1
If you accept this hypothesis, then the current LLM paradigm of "train once on the internet, freeze the weights, and run inference" is a dead end for true autonomy.
OaK introduces three highly valid, necessary fixes to current AI:
True Continual Learning: Today's neural networks suffer from catastrophic forgetting—if you try to teach an LLM new facts on the fly, it degrades its old knowledge. OaK uses meta-learned step sizes (via algorithms like their new NetworkIDBD) to dynamically adjust how fast individual weights update. This allows the agent to filter out random background noise while locking onto stable, predictive signals in real time.
TNW
Efficiency (The 20-Watt Goal): Frontier models require megawatts of power to train and run. A human brain runs on about 20 watts. By learning incrementally from a direct stream of experience without needing to store or endlessly replay massive datasets, OaK aims to match biological energy efficiency.
TNW+ 1
Building Abstractions from Scratch: Instead of humans hardcoding what a "door" or a "step" is, the FC-STOMP progression lets the agent build its own state and temporal abstractions based strictly on what helps it predict and control its environment.
The Skeptics' View: Where OaK Might Stumble
While the theory is beautiful, critics point out three massive challenges that Sutton’s lab has to overcome:
1. The "Cold Start" and Sample Inefficiency Problem
Pure reinforcement learning is notoriously sample-inefficient. It takes millions of trial-and-error steps for an RL agent to learn to play a simple Atari game, let alone navigate a messy physical kitchen. LLMs bypass this by starting with a massive, pre-digested map of human culture, language, and physics. Starting from absolute zero experience (as OaK proposes) might take a prohibitively long time, even if the learning algorithm is perfect.
2. Hardware Mismatch
Our current hardware ecosystem (Nvidia GPUs and TPUs) is highly optimized for massive parallel matrix multiplication—the exact math that powers transformers. Sutton's approach requires highly sequential, real-time, online weight updates. Running OaK efficiently might require fundamentally different hardware architectures, such as neuromorphic chips or analog compute-in-memory systems.
3. The "Good Enough" Hybrid Threat
Many AI researchers believe we don't need to throw away LLMs to get what Sutton wants. Instead, the industry is moving toward hybrid architectures—wrapping LLMs in outer loops of reinforcement learning, search, and planning (similar to what we see in reasoning models). Critics argue that patching existing models is a much faster path to AGI than building a pure RL agent from scratch.
The Verdict
Sutton is betting on the long game. If you want an AI that can pilot a spacecraft to a distant planet, adapt to entirely alien physics, and learn continuously on a tiny solar battery, the OaK architecture is arguably the only valid path we have. But if the goal is to build highly capable virtual assistants and software tools in the next two to three years, the current scaling of transformer models will likely continue to dominate.
TNW
If you want to watch the full conceptual breakdown of how he intends to map out this architecture, you can watch Richard Sutton's OaK Architecture keynote at AGI-25. This video is highly relevant because it features Sutton himself explaining the theoretical pillars of the OaK architecture, including the FC-STOMP progression and the physics of temporal abstraction.
Meine Antwort:
thank you, that here is for mz blog and should suffice
Gemini:
You're very welcome! That breakdown should give your "Keylmann Report" readers a great, balanced overview of where Sutton is trying to take the field. It’s the ultimate clash of AI philosophies—the brute-force scaling of frozen data versus the elegant, biological adaptation of continuous experience.
Nessun commento:
Posta un commento