The Missing Gradients: Why Sutton and SynDE Agree the Oracle is Dead

The Fundamental Diagnosis

Richard Sutton's critique of Large Language Models isn't just another contrarian take. It's a mathematical diagnosis of why these systems can't learn, can't improve, and can't achieve intelligence. His recent "Era of Experience" essay with David Silver, combined with his declaration that "LLMs are a dead end," reveals the core problem:

There are no gradients to follow.

Not weak gradients. Not conflicting gradients. Missing gradients.

The Mathematics of Learning

Every learning system needs three things:

  1. A prediction about what will happen

  2. An observation of what actually happened

  3. A gradient - the difference that drives learning

LLMs have none of these. As Sutton devastatingly observes: "They won't be surprised by what happens. If something unexpected happens, they will not change because an unexpected thing has happened."

A system that can't be surprised can't learn. A system without surprise has no gradients. A system without gradients is frozen in time, forever performing the same statistical mimicry of its training distribution.

The Oracle's Missing Feedback Loop

The SynDE critique identified this from an engineering perspective: the oracle model fails because it lacks decomposition and ground truth. Sutton identifies it from a learning perspective: without environmental feedback, there's nothing to optimize against.

"There's no ground truth in large language models," Sutton states. "You don't have a prediction about what will happen next... There's no goal. If there's no goal, then there's one thing to say, another thing to say. There's no right thing to say."

Without "right" or "wrong," there's no loss function. Without a loss function, there are no gradients. Without gradients, there's no learning. It's that simple and that fatal.

The Mimicry Trap

Sutton and Silver's "Era of Experience" essay makes this explicit: "While imitating humans is enough to reproduce many human capabilities to a competent level, this approach in isolation has not and likely cannot achieve superhuman intelligence."

Why? Because mimicry provides no gradients for improvement. You're not learning to solve problems; you're learning to sound like someone who solves problems. The difference is everything.

Consider their devastating example: An agent trained on human thoughts from different eras would reason about physics using:

  • Animism (5,000 years ago)

  • Theistic explanations (1,000 years ago)

  • Newtonian mechanics (300 years ago)

  • Quantum mechanics (today)

Each paradigm shift required interaction with reality - making predictions, observing failures, following gradients toward better models. Without environmental grounding, the agent becomes "an echo chamber of existing human knowledge."

Where SynDE Creates the Missing Gradients

The two-phase SynDE architecture isn't just organizationally cleaner - it's mathematically necessary:

Phase 1 (Intent Capture): Transform vague human intent into precise, executable specifications

  • Creates testable predictions about outcomes

  • Defines success and failure conditions

  • Establishes the ground truth that LLMs lack

Phase 2 (Execution): Generate actual outcomes in the environment

  • Every execution produces a prediction: "This workflow will achieve this result"

  • Every outcome provides feedback: success or failure

  • Every feedback creates a gradient: adjust toward what worked

This isn't an optimization. It's the minimum viable architecture for learning.

The Industry's Desperate Gradient Hunt

Look at what everyone is frantically building:

GitHub Spec Kit: Forces specification → planning → execution phases to create feedback loops

Claude Code: Separates planning from implementation from review to generate correctness signals

OpenAI o1: Adds "thinking" steps trying to create intermediate checkpoints for self-correction

DeepSeek-R1: "Rather than explicitly teaching the model how to solve a problem, we simply provide it with the right incentives"

They're all trying to retroactively create the gradients their foundation models lack. But you can't bolt gradients onto a system trained without them. The architecture must be designed for learning from the ground up.

The AlphaProof Validation

The essay's example of AlphaProof is perfect: It started with 100,000 human proofs, then generated 100 million more through environmental interaction. The human data provided initial direction, but the real learning came from trying proofs and seeing what worked.

That's a 1,000:1 ratio of experience to human data. And it achieved what no LLM could: medal-level performance at the International Mathematical Olympiad.

Why? Because each proof attempt created a gradient. Each failure taught something. Each success reinforced what worked. The system could be surprised, could be wrong, could learn.

The Bitter Truth About the Bitter Lesson

Sutton's original Bitter Lesson said compute beats human knowledge. But LLMs seemed to validate this - they scaled with compute!

Now Sutton reveals the deeper truth: LLMs violate the Bitter Lesson because they're "a way of putting in lots of human knowledge." They don't learn from compute; they memorize from data. When the data runs out, the learning stops.

True scaling requires experience, not data. Gradients, not gigabytes.

The Revolutionary Implication

This isn't about improving LLMs. It's about replacing them.

As Sutton says: "Once we have [continual learning architecture], we won't need a special training phase—the agent will just learn on-the-fly, like all humans, and indeed, like all animals. This new paradigm will render our current approach with LLMs obsolete."

SynDE doesn't need to compete with LLMs on their terms. It's playing a different game entirely - one where every interaction generates gradients, every execution provides feedback, every workflow creates ground truth.

The Missing Gradients Create Everything Else Missing

No gradients means:

  • No learning from experience

  • No adaptation to new situations

  • No discovery beyond training distribution

  • No genuine problem-solving

  • No path to intelligence

With gradients from environmental interaction:

  • Continuous improvement

  • Adaptation to specific contexts

  • Discovery of novel solutions

  • Real problem-solving capability

  • Actual intelligence emergence

Conclusion: The Gradients Were Never There

The industry spent $100 billion looking for intelligence in statistical pattern matching. They found impressive mimicry instead.

Sutton's diagnosis is final: Without environmental gradients, without surprise, without ground truth, these systems are "an echo chamber of existing human knowledge." They're not learning. They never were.

SynDE works because it creates what LLMs fundamentally lack: a closed loop between prediction and reality, intent and execution, action and consequence. Every workflow generates gradients. Every gradient enables learning. Every learning step moves toward genuine intelligence.

The oracle is dead because it was never alive. It was just replaying compressed human knowledge without the gradients needed to build its own understanding.

The future belongs to systems that can be surprised.