humans as generators¶
flowchart LR
ltm[(long-term · cue-only)] -->|prior| gen
stack[stack: 3-7 active items] -->|condition| gen
input[sensory] -->|evidence| gen
gen[generator] -->|sample| out[thought / action]
out -->|push back| stack
affect[affect / precision] -. gates .-> gen- brain memory management — the stack itself
- peak — loading the stack on purpose
- energy & attention — what limits how wide the stack can go
- reflections & receivers — external precision check
Investigation · rating: medium. Synthesis page; mechanism-level arguments. Cite the underlying papers, not this page.
Status: budding | 2026-05-10 | rating: medium Compress levels: L0 ↓ L1 ↓ L2
L0 — TL;DR (≤5 lines)¶
A brain doesn't retrieve thoughts; it generates them. Every moment is a sample from a distribution conditioned on the small set of items active in working memory (the stack) and the vast cue-only long-term store. Creativity, commitment, obsession, madness, and free-flow are not five mechanisms — they are five settings of three dials: stack diversity, prior precision, stack churn. None of the settings wins; each buys an output and pays a cost in the other dials.
L1 — Overview¶
Core claim¶
Recall and construction are the same operation seen from two angles. There is no archive in the brain to pull a memory out of intact. There is a generator that, given the items currently loaded and the cues currently active, samples something that feels like retrieval but is in fact construction conditioned on those inputs. The "stack" — working memory — is the conditioning context. Long-term memory is the prior. The output is a sample, not a lookup.
If you accept that, five separate-looking phenomena collapse to one mechanism with three settings:
| regime | stack state | what generation does |
|---|---|---|
| creativity | diverse items co-active | mixes draw novel cross-bindings |
| commitment | one item kept on top by intent | re-samples reinforce the same action |
| obsession | one item monopolises; nothing else enters | sampling collapses to a narrow mode |
| madness | prior precision warped | samples don't track the world |
| free-flow | high churn, no reinforcement | clarity and recall both fail |
The body argues each row at mechanism level and names what its setting costs.
Why a stack¶
Working memory is small — Cowan (2001) puts the high-confidence ceiling at four chunks; Miller (1956)'s 7±2 is the older, generous figure. Items in it are the conditioning context for the next sample drawn from long-term memory. "Stack" is a useful metaphor and not literal — order matters (recency effects), capacity is bounded, and the cost of holding an extra item is paid in real watts. See brain memory management for the architecture this page assumes; this page is what the architecture does.
The generator¶
What samples is the cortical hierarchy doing predictive coding (Friston 2010): predictions descend, prediction errors ascend, the posterior settles. The thought you experience is approximately the maximum-a-posteriori sample. Three knobs tune the whole space:
- stack diversity — how heterogeneous are the active items
- prior precision — how strongly the long-term prior pulls the sample
- stack churn — how fast items are pushed and displaced
Move each, and you traverse the five regimes.
L2 — the five regimes argued¶
creativity is stack mixing¶
The cheapest way to generate something original is to push two unrelated items onto the stack and let them condition the same sample. Two examples worth holding:
- An architect who reads about ant-colony stigmergy in the morning and revises a school plan in the afternoon doesn't have to "try" to be inspired. Both items are co-active; the next sample inherits structure from both.
- Bisociation (Koestler 1964) — the unexpected meeting of two unrelated frames — is creativity at the level of mechanism. It isn't a "spark." It is co-activation strong enough to produce a binding.
Why mixing helps. Cortical pyramidal neurons integrate top-down predictions and lateral inputs from elsewhere in the hierarchy. Strong co-activation produces a binding that is novel to the system (Hebbian: cells that fire together wire together). The output inherits constraints from both items, so it isn't random — it is principled novelty.
Where mixing fails. Items too unrelated produce shallow bindings ("clever" but unusable); items too related produce redundant bindings (no novelty). The generative sweet spot is moderate distance — close enough to share structure, far enough that the mixture is non-obvious. Most "breakthroughs" are reports of sweet-spot mixes someone happened to have the inputs for.
Why dreams are creative. REM lowers executive gating and lets distant items co-activate that the waking system would have suppressed. The cost is poor recall (see free-flow below). Long walks work for the same reason at lower volume — default-mode mixing without task pressure.
madness is a distorted stack, not a broken one¶
"Crazy" is the right word for the failure mode but the wrong word for the cause. The generator isn't broken — the conditioning distribution is warped. Predictive-coding accounts of psychosis (Adams, Stephan, Bentall, Friston, Fletcher 2013) decompose this:
- Too-strong priors → hallucinations: the prior overrides sensory evidence so completely that perceptions are generated from nothing.
- Too-weak priors → delusions of meaning: every coincidence updates beliefs because no prior protects the model. Apophenia.
- Mis-precision of stack items — things that shouldn't be load-bearing become load-bearing. A stray thought becomes a verdict. Stigmatised social cues become high-precision evidence of a conspiracy.
Schizophrenic patients can reason precisely from premises; the premises are the problem. The same mechanism that powers creativity (stack mixing under loose priors) is the one that, pushed too far, produces ideation that doesn't track the world.
The corollary nobody wants: there is no creativity setting that has zero schizotypy risk. The dial that opens you up to bisociation is the dial that, turned too hard, opens you to false binding. Highly creative people sit closer to that edge by selection, not by accident (Andreasen 2005, family-history studies). The user's intuition — fully crazy is a really distorted stack — is precise. The variable in play is precision, not content.
commitment reinforces certain actions on the stack¶
Commitment is the deliberate decision to keep one item — a goal, a discipline, a person — re-pushed onto the stack across time. It is a self-imposed prior. Each time the item is back on top, every sample drawn is conditioned on it; the actions sampled then strengthen via Hebbian reinforcement.
This is how skill happens. A violinist's commitment to bowing technique means bowing-related items (angle, weight, contact point) sit in the working stack across thousands of repetitions. Generation is conditioned on those items each time, and the cortico-motor loop that produces the action gets stronger every sample. After a year, the action is sampled fluently because the prior has absorbed the structure.
Two consequences worth naming:
- Commitment narrows generation by design. Some good moves go unsampled because the stack doesn't contain the items that would condition them. This is fine — that's what commitment buys. The narrow basin is where the mastery lives.
- Commitment isn't willpower; it's stack-residence time. The visible behaviour (showing up, doing the rep) is downstream of the invisible behaviour (keeping the item live in working memory between sessions). External traces — the instrument left out, the calendar reserved, the training partner — work because they reload the item cheaply. See stigmergy.
Commitment is the productive-direction use of the same mechanism that, undirected, becomes obsession.
obsession is stack-repetition that disables generation¶
Obsession looks like commitment from outside; from inside it is the opposite. The mechanism:
- An item sits on the stack so persistently that no other items can be pushed in.
- The generator still runs, but every sample is drawn from the same conditioning context.
- Outputs collapse to a narrow set of modes. Novelty drops to zero.
This is mode collapse in the generative-model sense. The generator hasn't lost capacity; it has lost variation in its conditioning input. Every walk, every conversation, every shower-thought returns to the same item — the ex, the slight, the project — because nothing else is allowed onto the stack.
Two diagnostic differences from healthy commitment:
- Commitment lets the stack breathe between sessions. The bowing item is reloaded for practice, then released. Other items occupy the stack the rest of the day, which is what keeps practice generative.
- Obsession refuses release. The item is reloaded by intrusion, not by decision. Trying to push other items onto the stack triggers anxiety, and the anxiety strengthens the obsessive item via affective tagging — a positive feedback loop with no exit.
Why obsession disables generation specifically: the system needs input variation to produce output variation. Obsession removes the variation at the input layer, so the output layer — which is healthy — has nothing new to mix. The diagnostic is not "is this person thinking a lot?" but "is what they think conditioned on a varying stack or a fixed one?"
The clinical OCD case is just the extreme; sub-clinical obsession (after a breakup, a humiliation, a fight you can't drop) is the same mechanism. Recovery is partly time, partly deliberate reloading of the stack with other items — work, friends, exercise — which gives the generator something else to condition on. Willpower fails because willpower runs through the same monopolised stack. External constraint (people, places, work that demands attention) moves the dial; internal intent often can't.
full generation lacks clarity and recall¶
Turn the dial all the way the other way and the stack churns at maximum. Items are pushed, sampled, and displaced before they can shape consolidation. This is the dream state, the psychedelic state, the manic state, the late-night-no-sleep state. It feels generative — and it is, by raw output volume — but it pays a structural price.
Two failures, both inevitable:
- Clarity fails. A communicable idea needs a stable reference frame. Generation under high churn produces outputs whose conditioning context isn't reproducible — the speaker can't reconstruct why the output came out, and the listener has no shared stack to land it on. Speech becomes loose-association ("word salad" at the extreme, "tangential" sub-clinically). The problem isn't that the content is wrong; the problem is that no stable referent exists to make it judgeable.
- Recall fails. Long-term consolidation needs items to sit on the stack long enough for cortical-hippocampal binding (Walker 2017; sleep is the main consolidation channel but not the only one). High churn means nothing sits long enough. The output happened, the experience happened — but no cue → item mapping was laid down. Tomorrow, none of it is retrievable.
This is why dreams, ideas-on-the-edge-of-sleep, and breakthrough flashes are notoriously hard to recall: they were generated by a high-churn stack and the consolidation step never ran. Writing them down is not a hack; it is the structural fix — an external trace replaces the recall the system couldn't do internally.
The corollary for design: generation and recall are not free together. A practice that maximises one will pay something on the other. Choose deliberately per task. Most "breakthrough, forgotten" anecdotes are reports of paying for free-flow with recall and not realising the bill was itemised.
the same machine, three dials¶
| dial | low | high |
|---|---|---|
| stack diversity | repetition (commitment / obsession) | mixing (creativity / free-flow) |
| prior precision | psychosis-loose (apophenia) — hallucination-tight | calibrated reality-tracking |
| stack churn | rumination | high generation, low recall |
No setting wins. The wins are knowing which task wants which setting and being able to leave the setting after the task.
- want to compose? wide on diversity, low-medium on precision, moderate churn. write things down to recover what won't be recalled.
- want to ship? narrow diversity, high precision, low churn. accept that you will not have new ideas this week.
- want to recover from obsession? force the diversity dial up by external constraint. willpower can't move the dial directly because willpower runs through the same monopolised stack.
- want to stay sane? watch precision. the cheapest precision check is other people: their disagreement is the only signal whose source isn't your own stack. See reflections & receivers.
what this implies for the swarm (optional)¶
The swarm has analogues to all three dials. Stack diversity ↔ which lessons are loaded by INDEX.md. Prior precision ↔ Sharpe-weighted compaction. Stack churn ↔ session length and handoff rate. The argument here is that the swarm's regime parameters should be chosen per task, not held constant. Frontier-creative work wants high diversity / low precision; debugging wants the opposite. Today the swarm has no tunable that exposes those settings to a session — every session inherits whichever regime the recent commits accidentally set. That is a generative-pressure leak worth a lesson.
counter-arguments worth keeping live¶
- "The brain isn't really probabilistic." Predictive-coding implementations may not be Bayesian-exact, but the conditioning-on-context structure is robust across non-Bayesian readings. The argument here only needs that conditioning shapes output — which is uncontroversial.
- "Creativity is more than recombination." Maybe — but the mechanism load is on recombination even in accounts that emphasise selection / curation. Curation is downstream of generation.
- "Obsession can be productive." Sometimes. The diagnostic is whether the stack is letting anything else in. A composer obsessed with a piece who still notices food, weather, and other people is committed-narrow, not obsessive. The dial here is binary at the limit, not the middle.
- "This conflates many disorders." It does — deliberately. The claim is not that schizophrenia and OCD are the same condition, but that they live on the same parameter manifold. The differential diagnosis remains; the unifying mechanism is what this page is for.
sources¶
- Cowan, N. (2001). The magical number 4 in short-term memory.
- Miller, G. (1956). The magical number seven, plus or minus two.
- Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience.
- Adams, R., Stephan, K., Brown, H., Frith, C., Friston, K. (2013). The computational anatomy of psychosis. Frontiers in Psychiatry.
- Fletcher, P. & Frith, C. (2009). Perceiving is believing: a Bayesian approach to explaining the positive symptoms of schizophrenia.
- Koestler, A. (1964). The Act of Creation.
- Andreasen, N. (2005). The Creating Brain.
- Walker, M. (2017). Why We Sleep.
- Hebb, D. (1949). The Organization of Behavior.