Taxonomy & Terminology Note
Synthetic Intelligence (SI) refers to the deliberate synthetic recreation of cognitive processes, perception-action loops, active inference, and neuro-symbolic reasoning in non-biological computational substrates. Synthetic Intelligence is not a universally standardized industry category, and should not be confused with SI when used colloquially as an alternate abbreviation for Superintelligence (ASI). For alignment theory and theoretical containment of superintelligent systems, please refer to our Artificial Superintelligence (ASI) Studio Catalogue.
Mastering Synthetic Cognitive Architectures
A deep dive into Global Workspace Theory (GWT), Karl Friston's Free Energy Principle, active inference loops, and hybrid neuro-symbolic arbitration engines.
01. Global Workspace Theory (GWT) in Machine Cognition
Originally formulated by cognitive neuroscientist Bernard Baars and mathematically adapted for neural computation by Stanislas Dehaene, Global Workspace Theory models conscious access as a shared high-bandwidth blackboard. Specialized modules (perception, memory, motor controls, semantic parsers) compute in parallel sub-consciously.
Subconscious Specialists
Independent neural networks, transformer attention heads, or rule-based experts that continuously score inputs for novelty, threat, or informational salience without global broadcast.
Global Broadcast Bottleneck
The winning specialist whose salience exceeds the dynamic attention threshold gains access to the global blackboard, synchronizing state vectors across all cognitive subsystems simultaneously.
// Mathematical Attention Gate in Synthetic Intelligence Workspace
function arbitrateWorkspace(specialists, salienceThreshold) {
// Sort candidates by informational surprisal
const winners = specialists.filter(s => s.salienceScore >= salienceThreshold);
if (winners.length === 0) return { broadcast: null, state: 'quiescent' };
// Winner-take-all with soft competitive lateral inhibition
const dominant = winners.reduce((max, s) => s.salienceScore > max.salienceScore ? s : max);
return {
broadcast: dominant.payload,
suppressionRatio: dominant.salienceScore / (specialists.reduce((acc, s) => acc + s.salienceScore, 0) || 1),
timestamp: Date.now()
};
}
02. Active Inference & The Free Energy Principle
Karl Friston's Free Energy Principle posits that any self-organizing system that resists entropy must minimize the divergence (variational free energy) between its internal generative model of the world and sensory observations.
Perception: Updating internal generative states $\mu$ to accommodate unexpected sensory observations $y$ (learning / belief updating).
Action: Executing motor controls $a$ to alter the physical environment so sensory outcomes conform to internal prior beliefs (active inference).
# Python Active Inference Free Energy Computation (Friston Step)
import numpy as np
def compute_variational_free_energy(sensory_obs, generative_prior, precision):
"""
F = Complexity - Accuracy
F = D_KL[q(s) || p(s)] - E_q[log p(o | s)]
"""
prediction_error = sensory_obs - generative_prior
accuracy_loss = 0.5 * precision * np.sum(prediction_error ** 2)
complexity_penalty = 0.5 * np.log(1.0 / (precision + 1e-8))
free_energy = accuracy_loss + complexity_penalty
return free_energy, prediction_error
03. System 1 (Intuitive) vs System 2 (Deliberative) Arbitration
In synthetic systems, relying exclusively on LLMs or heavy search algorithms incurs massive latency and token expense. Implementing Kahneman-style dual routing creates an efficient energy envelope:
| Attribute | System 1 (Reflex / Embedding Match) | System 2 (Deliberative / Graph Search) |
|---|---|---|
| Latency | < 15ms | 800ms - 8,000ms |
| Mechanism | Vector similarity, associative key-value cache, small distilled SLMs | Monte Carlo Tree Search (MCTS), Formal logic prover, multi-agent debate |
| Surprisal Tolerance | Low (< 0.35 free energy) | High (>= 0.35 free energy trigger) |