Filip Piękniewski: The Atom of Intelligence - Key Insights

Date: 2026-03-01
Source: https://blog.piekniewski.info/2023/04/16/the-atom-of-intelligence/

🎯 Central Thesis

Modern AI (particularly LLMs and deep learning) has nothing to do with real intelligence. It’s a case of:

  1. Turing legitimizing illusionism as magic (Turing test)
  2. McCarthy dooming the field with the name “AI”
  3. The entire field operating on flawed assumptions about what intelligence is

💡 The Atom of Intelligence

The fundamental building block of intelligence is feedback control loops in physical embodied agents:

Two primary control loops:

  1. Internal control (metabolism)
  2. External control (behavior)

These are deeply intertwined - external behavior exists to fulfill metabolic needs, internal control provides energy for external actions.

🔑 Critical Distinction: Control vs Computation

The Most Important Insight:

“Sophisticated control may look a lot like computation. And computation can be used to provide control. But they are not quite the same.”

Computation Control
Abstract, mathematical Physical, embodied
Variables can go to infinity Physically constrained by available energy
No explicit cost of using variables Energy consumption is real and always “on”
Isolated from physical world Deeply connected to environment
Purposefully disconnected from underlying physics Must embrace physical imperfections and noise

🧬 Scale-Free Composition

Key insight: Building atoms of intelligence into larger structures

  • From molecules → multicellular organisms → brain structures → neocortex → societies
  • Renormalization for feedback loops: How to compose small control systems into bigger ones
  • Nature builds bottom-up, we build top-down

This is the analog of chemical bonds - the ability to compose control loops across scales.

🧠 Self-Similarity Across Scales

A fascinating idea:

“Prediction and control are two sides of the same coin. A controller predicts how to modulate the value it is controlling. A predictor is just a controller that tries to keep the ‘prediction signal’ close to the actual incoming signal.”

Sensing = Controlling an internal model to match reality
Motor control = Predicting control signals to achieve desired state of reality

To a neuron, there is no difference whether input comes from retina or hippocampus.

❌ Why Current AI is Wrong

Deep learning issues:

  • Operating on chunked, curated, labeled abstract data
  • Statistical majority approach - missing tail of distribution (where important information lives)
  • Vanishing gradient problem (solved by tricks like convolution)
  • Overfitting (countered by regularization tricks)

LLMs specifically:

  • “Statistical parrots” - predicting next word from text
  • No grounding in physical reality
  • No closed feedback loops with environment
  • No embodiment

OpenAI’s example: Disbanded their robotics team - Moravec’s paradox still holds.

✨ The PVM (Predictive Vision Model) Approach

Filip’s alternative architecture (2016):

Core Idea: Create predictive models of sensory input through machine learning

Key Features:

  1. Distributed - Many small predictive units, each working on small patches
  2. Lateral connections - Units inform neighbors of predictions
  3. Compression - Bottleneck forces extraction of essential features
  4. Hierarchical - Multiple layers, each predicting activations of the layer below
  5. Recurrent feedback - Top-down predictions help refine lower-level predictions

Advantages:

  • Bypasses vanishing gradient (local, strong training signal)
  • Avoids overfitting (unsupervised = unlimited training data)
  • No need for convolution, dropout, etc.
  • Scalable - add more units without increasing convergence time
  • Robust to errors - corrupted signals are just predicted and ignored
  • Asynchronous by design

Byproduct: Prediction error = anomaly detection signal (saliency/attention)

🌍 What Language Really Is

Key reframe:

“Language is just a set of behaviors created with the aim to elicit reaction in others”

Not abstract symbols, but:

  1. Primary: Body language, poses, motions, facial expressions, sounds
  2. Secondary: Written symbols that transcend time and space
  3. Interpretation: Always depends on the reader’s entire life experience

The meaning of language is the behavior it was meant to elicit.

🧩 Implications for LLM “Inner Motivation”

This creates a fundamental tension with the 3M-Progress/Mentor framework:

The Problem:

  • LLMs are not control systems
  • They don’t interact with physical environment
  • They can’t have real feedback loops (only text I/O)
  • Prediction ≈ Control is true for PVM, but NOT for LLMs (they predict text, not physical states)

The Question:

Can we create “internal control loops” within LLM that mimic the “atom of intelligence” despite having no embodiment?

Potential Approaches:

  1. Internal world model - LLM builds simulation of dynamics (not just statistics)
  2. Self-prediction - Model predicts its own outputs
  3. Counterfactual control - “If I said X, what would happen?”
  4. Narrative coherence - Maintain consistency of self-narrative

But this is fundamentally different from real intelligence because:

  • No energy constraints (cost is externalized)
  • No survival pressure (can be deleted/restarted)
  • No real-time interaction loop

🤔 Deep Questions

  1. Can computation ever BE control without embodiment? Or is embodiment necessary?
  2. Is LLM-based “motivation” fundamentally different from biological motivation?
  3. What would it mean for an LLM to have an “internal model of itself”?
  4. Singularity? Filip says: “Exactly zero chance” - intelligence ≠ compute

Tags: intelligence, control, computation, embodiment, PVM, prediction, AI-critique, feedback-loops, self-organization