From neurons to Hilbert spaces to the edge of spacetime, a map of every framework that claims to explain why you do what you do.
Begin the DescentClick the field below. Three competing possibilities exist in superposition, until observation collapses them into a single outcome.
Before quantum cognition, there was the ABC model, the bedrock of behavioral psychology. Every action is a chain: Antecedent triggers Behavior, which produces a Consequence that feeds back into the loop.
Skinner's framework: reinforcement strengthens behavior, punishment weakens it. The environment sculpts the agent. No internal states needed, just stimulus and response frequencies.
Simon's insight: humans don't maximize utility, they satisfice. The agent uses heuristics, truncates search trees, and settles for "good enough." Rational, but within limits.
Kahneman and Tversky's Nobel contribution: losses loom larger than gains. The value function is concave above a reference point and convex below it. Decisions are relative, not absolute.
These frameworks capture enormous predictive power. But they all share an assumption: that the decision state is definite before it is measured. Quantum cognition challenges exactly this.
In the 2000s, researchers noticed something odd: human judgment violates classical probability. Conjunction fallacies, order effects, disjunction effects, patterns that cannot arise from any classical probability distribution. But they arise naturally from quantum probability.
Instead of placing beliefs on a classical probability simplex, map them onto a Hilbert space. A person's cognitive state is a state vector |ψ⟩. A question is a projector. The probability of a "yes" answer is:
Crucially, projectors don't commute: P_A P_B ≠ P_B P_A. The order in which you ask questions changes the answer: exactly as observed in human surveys.
Positive Operator-Valued Measures generalize projectors. They allow "fuzzy" measurements, partial observations that don't fully collapse the state. Perfect for modeling ambiguous social judgments.
The Linda Problem's conjunction fallacy: P(feminist & bank teller) > P(bank teller). Classically impossible. But in quantum probability, the cross-term interference between "feminist" and "bank teller" amplitudes adds probability mass to the conjunction.
See the same decision, "Should I take this job offer?", through five lenses.
The decision is a function of reinforcement history. Past job switches that led to positive outcomes (higher salary, praise) increased the probability of "accept." Past negative outcomes (social disruption, regret) decreased it. The agent's internal states are a black box.
The agent builds a mental model: a decision tree with nodes for salary, culture, commute, growth. Each branch has a subjective utility weighted by estimated probability. Bounded rationality means the tree is pruned. System 1 (fast heuristics) pre-filters; System 2 (slow deliberation) evaluates the shortlist.
Before deliberation, the agent is in a superposition of "accept" and "reject." Thinking about salary projects the state onto one basis; thinking about family projects it onto another. These bases are incompatible: they don't share eigenstates. The order of consideration creates interference that biases the final measurement (the decision).
The optimal agent compresses all past observations into the shortest program that predicts them (Kolmogorov complexity), then acts to maximize expected future reward over all computable environments, weighted by their algorithmic probability. The job decision is a single step in an infinite-horizon optimization.
There is no "agent" and no "job." Both are patterns in the Ruliad, the entangled limit of all possible computations. What appears as a "decision" is a particular foliation of rulial space, a cross-section where observer and environment momentarily crystallize. Constructor theory asks: which transformations are possible, regardless of dynamics?
What if we could build the mathematically optimal agent? Marcus Hutter's AIXI is exactly that, a formal definition of perfect intelligence that is also perfectly uncomputable.
AIXI combines Solomonoff induction (the optimal predictor) with sequential decision theory (the optimal actor). At each time step, it:
1. Enumerates all computable hypotheses about the environment.
2. Weights each by 2-K(h): shorter programs are more probable.
3. Selects the action maximizing expected lifetime reward across all hypotheses.
AIXI is the ceiling of intelligence: no agent in any computable environment can do better. Its uncomputability is the price of universality.
The length of the shortest program that produces a string. It's the mathematical formalization of "simplicity", and it's uncomputable. We can only approximate it, which is what compression algorithms (and human intuition) do.
The optimal way to predict the next bit of a sequence: weight every consistent program by its algorithmic probability. It converges on the true distribution faster than any other predictor. It's also uncomputable, a recurring theme.
AIXItl truncates the horizon and limits program length. MC-AIXI uses Monte Carlo tree search. These lose optimality guarantees but become runnable. Modern LLMs can be seen as massively parallel, approximate Solomonoff predictors.
The deepest frameworks dissolve the boundary between agent and environment entirely. Here, "choice" is not something an agent does: it's a pattern that the universe exhibits.
Click each layer to descend from familiar neuroscience into increasingly alien territory.
The boundary between self and world. Sensory states flow in; active states flow out. Internal states are shielded from the environment, they can only "see" the world through the blanket.
David Deutsch's framework shifts physics from dynamics to possibility. Instead of asking "what happens next?", it asks "which transformations are possible and which are impossible?" An agent is a constructor: a system that can cause a transformation and retain the ability to cause it again.
Stephen Wolfram's ultimate structure: the entangled limit of all possible computations applied to all possible initial conditions. Physics, mathematics, and consciousness are different samplings of this single object. Observer-reality boundaries are foliations of rulial space.
All information in a 3D volume can be described by data on its 2D boundary. The Markov blanket of an agent is the agent, a holographic screen encoding everything the internal states need. Consciousness may be a boundary phenomenon.
Five questions. Five agent types. Where do you fall on the spectrum from classical reinforcement learner to non-dual pattern in the Ruliad?
Share your result, the architecture of choice starts with understanding your own.