Sutton & Javed Deep Dive: The Big World Hypothesis and the 'Digital Soul' of Continual Learning AI

Forum topic · ✨步子哥 · 2026-08-20

Summary

This in-depth analysis examines the recent work of Richard Sutton (2024 Turing Award co-winner) and Khurram Javed, co-founders of Oak Lab. Their central diagnosis: today's large language models suffer from frozen, deployment-time learning—they stop growing after training, lack true goals and reward signals, and cannot evaluate their own outputs. Their counter-framework rests on three pillars: the Big World Hypothesis (agents are orders of magnitude smaller than their environments, so compute scaling alone cannot capture the world; learning at runtime beats design-time learning), the OaK (Options and Knowledge) architecture (a domain-general, experience-driven agent design with meta-learned per-weight step sizes), and a 20-watt energy target inspired by the human brain. Their first algorithm, NetworkIDBD, extends IDBD to nonlinear networks for selective credit assignment, showing on NoisyMNIST that learned weights concentrate on signal rather than noise—potentially eliminating mini-batch training and replay buffers. The report critically distinguishes the authors' actual terminology from metaphors like 'digital soul' and 'learning stagnation,' and voices six counterarguments, noting that no large-scale prototype yet exists and continual plasticity remains unsolved. Sources include the Big World Hypothesis paper, The Bitter Lesson, Sequoia Capital interviews, and Oak Lab coverage.

Sutton & Javed Deep Dive: Breaking the 'Big World Hypothesis' Barrier and the Birth of a 'Digital Soul'

> Research subject: Richard Sutton (father of reinforcement learning, 2024 Turing Award) and Oak Lab co-founder Khurram Javed — on the 'learning stagnation' of modern AI, the Big World Hypothesis, and the technical/philosophical implications of a '20-watt, non-replicable digital soul.' > Method: Based on primary sources (the Javed–Sutton Big World Hypothesis paper, Sutton's The Bitter Lesson and 'Era of Experience' talks, Oak Lab coverage, and Sequoia interviews), with explicit correction of which metaphors are the authors' own terms versus the report's interpretive framing.

One-Sentence Summary (Feynman-style)

Picture two examinees:

1. The modern LLM: like a student who crams all of humanity's books before one final exam (pre-training), then never opens a book again — smart, but frozen after graduation, a puppet pulled by the strings of human text. 2. Sutton & Javed's ideal: like a newborn — seeing, trying, failing, correcting; learning for life from its own unique stream of experience, impossible to replicate.

Their core claim: the world is too big for any model to contain, so agents must live and grow on their own stream of experience. The human brain does all this on roughly 20 watts — their benchmark for 'real intelligence.'

> In short: stop training a frozen photocopier; raise an experiential agent that makes its own mistakes, improves on its own, and is one of a kind.

Who They Are and What They Oppose

| | Richard Sutton | Khurram Javed | |---|---|---| | Identity | Founder of modern RL; 2024 Turing Award (with Andrew Barto) | Sutton's student; Oak Lab co-founder | | Key contributions | TD learning (1988), policy gradients, the RL textbook, Reward Hypothesis | Continual learning, over-parameterized agents, empirical work on the Big World Hypothesis | | Career | Stanford psych BA → PhD under Barto → U. Alberta / RLAI → DeepMind Edmonton (2017–2023) → Keen Technologies (2023–2026.7) | PhD at U. Alberta under Sutton; joined Keen in the same period | | Current | Left Keen in July 2026 to co-found Oak Lab (Toronto) | Co-founder, Oak Lab |

What they oppose: the 'train once, freeze weights, deploy as-is' LLM paradigm. Sutton calls current deep learning methods 'weak and inefficient' and says what's needed is 'entirely new foundational ideas and a radical restructuring.'

> Timeline correction: the OaK architecture was presented at RLC 2025 / MIT lectures; Oak Lab the company was founded July 2026 after both left Keen.

Core Diagnosis: Frozen Models Stop Learning

Concept correction: 'learning stagnation,' 'marionette,' and 'digital soul' are not Sutton/Javed's own terms — they are this report's framing. Their actual diagnosis is two congenital defects:

1. Growth stops when training ends: frozen weights mean no learning from interaction after deployment; 2. No real trial-and-error feedback loop: models copy textual patterns without a ground-truth signal to judge or correct their own outputs.

Sutton's specific critiques:

No true objective, no ground truth: 'predict the next token' is not a goal — it doesn't act on the world and has no notion of good/bad (Dwarkesh interview).
No self-evaluation: generative AI imitates but 'cannot independently assess its own outputs,' blocking genuine scientific discovery.
From the Sequoia interview: "synthetic data is a big mistake"; "LLMs account for roughly a quarter of intelligence" — the rest lies in experience, goals, and world models; catastrophic forgetting is "totally curable" via algorithms like continual backprop.

Pillar 1: The Big World Hypothesis

From the Javed & Sutton paper (RLC 2024 'Finding the Frame' workshop):

> "The big world hypothesis: in many decision-making problems, the agent is orders of magnitude smaller than the environment. It cannot fully perceive the world's state, nor represent the value or optimal action for every state. It must learn to make sound decisions with limited understanding."

Its opposite is the over-parameterization hypothesis: the agent has spare capacity to represent simple solutions and search efficiently. Javed notes the hypothesis is "more a statement about which problems we should care about than a law of all decision problems" — Go's value function has no simple solution (big world); solving quadratic equations does not.

Why compute scaling can't save over-parameterization

1. Sensors are compute-bound, and appetites grow: a 2024 smartphone camera generates more data in one week than was used to train GPT-3. Finer perception makes decision problems bigger, not smaller. 2. The world gets more complex as compute grows: other equally complex agents are a source of irreducible complexity — no compute budget lets you perfectly model a strategic counterpart.

Hence Sutton: the Big World Hypothesis "will persist regardless of the rate of compute growth."

Corollary: non-stationarity → runtime learning beats design-time learning

| Dimension | Design-time learning | Runtime learning | |---|---|---| | Coverage | Cannot foresee all possibilities in a big world | Learns abstractions for situations actually encountered | | Adaptation | Pre-built abstractions generalize poorly | Highly customized to the experienced world | | Scaling | Limited by human expert knowledge | Scales with available compute (echoing The Bitter Lesson) |

Sutton argues design-time abstractions are "not just insufficient, but should be abandoned."

Pillar 2: The OaK Architecture (Options and Knowledge)

Three design principles:

1. Domain-general: no pre-installed world-specific knowledge; 2. Empirical: all mental growth comes from runtime experience; 3. Open-ended: the agent should generate any concepts it needs, limited only by compute.

The dual core:

Options: temporally extended behaviors with initiation/termination conditions (e.g., 'walk to the door,' not a single leg movement);
Knowledge: internal world models built from executing options, predicting long-term consequences of actions.

The loop: perceive features → abstract → generate long-horizon options → act and get feedback → update the world model and spawn new options.

Meta-learned step sizes: every learned weight carries its own step-size parameter, meta-learned via online cross-validation — the system learns which knowledge to learn fast, slowly, or forget.

OaK is one piece of the Alberta Plan, with its 12-step research framework: G1 continual representation/meta-learning, G2 world models and planning/search, G3 temporal abstraction and OaK integration.

> Sutton: "We want AI agents that can discover as we do, not containers of our discoveries."

20 Watts and the 'Digital Soul'

The human brain runs on ~20 watts; frontier LLMs need megawatt-scale training and kilowatt-scale inference (Oak Lab disclosures).
Oak Lab's long-term target: a trillion-parameter agent that learns and plans in real time on ~20 watts total. Sutton clarifies: 20 watts is a north star, not a product promise.

NetworkIDBD: the first algorithmic building block

Released 2026-07-13, NetworkIDBD extends incremental delta-bar-delta (IDBD) to nonlinear neural networks for selective credit assignment. In its NoisyMNIST experiment (28×28 digits embedded in 64×64 noise), NetworkIDBD concentrates weights on the central digits while SGD spreads them across all pixels including noise.

Key argument: modern reliance on IID mini-batches and experience replay is an engineering patch that locks systems into train-then-freeze structures. If credit assignment can be solved algorithmically, batches and replay — and their megawatt-scale costs — can be eliminated.

What 'digital soul' actually means

Again: not their term, but a defensible synthesis of their framework:

1. A unique experience stream: each agent's cognition is built on its own sensorimotor history — you can copy a weight snapshot, not the lived experience; 2. Heterogeneous decentralized goals: each agent has its own reward vector; in a big world, each learns irreplicable, perspective-bound abstractions; 3. Sutton's 'fourth great stage': AI as a successor to be nurtured like offspring, not feared and controlled.

Its engineering core is real: a genuinely continual-learning agent with its own goals is in principle non-replicable — though this has no measurable engineering evidence yet.

The Obstacles: Catastrophic Forgetting and Loss of Plasticity

Sutton candidly admits the core problems are unsolved:

Catastrophic forgetting: new knowledge overwrites old;
Loss of plasticity: prolonged training collapses representations into low-dimensional subspaces, destroying the ability to form new features — more insidious than forgetting, and worst precisely in no-replay online learning.

Oak Lab's algorithmic bets: selective credit assignment, continual backprop, adaptive step sizes, generate-and-test discovery of new features. Sutton believes reliable nonlinear continual learning "will be achieved in some form in the coming years," so OaK is designed assuming it.

Devil's Advocate: Six Knives (and Their Shields)

1. The Bitter Lesson used self-servingly (medium): scaling LLMs is itself a general method that wins with compute — Sutton reads the essay selectively. 2. Sparse rewards in open worlds (strong): RL shines in Go/chess with clear rewards; real open worlds make pure RL far harder — a shared challenge for LeCun and Silver too. 3. 20 watts is vision, not result (strong): the trillion-parameter, 20-watt, real-time-learning agent does not exist; no mid-scale OaK prototype has been shown; NetworkIDBD is validated only at NoisyMNIST scale. 4. Physical compute limits (medium): the argument depends on exponential compute growth that energy, process, and cost ceilings may cap. 5. 'Digital soul' is metaphor, not engineering (medium): non-replicability has no measurable definition or experimental support yet. 6. They're not alone, and it's not settled (light): LeCun (JEPA) and David Silver also doubt frozen pre-training; but with LLM+RL fusion already delivering (reasoning, math), declaring "LLMs are a dead end" is premature — complementarity is more likely than replacement.

Their shield: the diagnosis targets real pain points (human data plateau, frozen models, no self-correction loop); the Big World Hypothesis is backed by a primary paper with dual arguments; OaK's principles and the runtime-learning corollary are rigorous; the 'algorithmic credit assignment → drop replay → cut energy' chain has initial NetworkIDBD evidence. The weakness is an empirical gap (no prototype, no scale validation), not a conceptual flaw.

Final Verdict

1. Identity: a Turing laureate and his student, systematically indicting the 'freeze-and-stagnate' LLM paradigm via the Big World Hypothesis, OaK, and the Era of Experience — a coherent direction statement with a first algorithmic cornerstone, not clickbait. 2. Metaphor corrections: 'learning stagnation' = frozen-model symptom; 'marionette' = puppet of human corpora; 'digital soul' = this report's poetic synthesis of continual experience + heterogeneous goals + decentralized succession — not their terminology. 3. Three takeaways for system builders:

Runtime learning > design-time learning: don't split training and deployment; let agents keep updating through interaction;
Energy as a compass: judge algorithms by whether they can eliminate replay/mini-batches and cut energy by orders of magnitude, not by accuracy alone;
Continual plasticity is the real bottleneck: forgetting and plasticity loss matter more than parameter count.

Bottom line

a north star and a poem

complementary evolution

Sources

Javed, K. & Sutton, R. S. (2024). *The Big World Hypothesis and its Ramifications for Artificial Intelligence.* RLC 2024 'Finding the Frame' workshop. khurramjaved.com/the_big_world_hypothesis.html
Sutton, R. S. (2019). *The Bitter Lesson.* incompleteideas.net/IncIdeas/BitterLesson.html
Sutton, R. S. (2025). *Welcome to the Era of Experience.* BAAI Conference keynote.
Oak Lab founding coverage: MLQ.ai, Indian Express, The Decoder/TechTimes, Dev.to, Marsbit, The Blockbeats (2026-07); aibacon.net on OaK 'training-free operation.'
Sequoia Capital podcast: *Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again.*
Shanghai Institute for Advanced Study × Shanghai Jiao Tong × Openmind Research Institute RL seminar notes (2026-07-14, Sutton keynote).
WAIC 2026 group interview transcript (20-watt obsession; on the danger of aligning to human values).
TechTimes: NetworkIDBD algorithm details.
Secondary analyses: Sina Finance, Toutiao, fisherdaddy, ZOOZ.