Skip to content

Mind as waiting machine

Brain and Beckett name the same machine. A finite generator running active inference: predictions descend through deep cortical layers, prediction errors ascend through superficial ones; the active stack holds 3–7 slots; the rest of the world arrives as cues. ~80% of vagus is afferent — the brain is mostly *listening*. Psychiatric disease is the precision dials of this waiting machine slipping. WAITING-FOR-GODOT is the limit case: the actor cannot enter the scene as himself; the receivers' attentive waiting *is* the only channel he has. **Combo:** unifies BRAIN-STRUCTURE × BRAIN-MEMORY-MANAGEMENT × BRAIN-DISEASES × BRAIN-BODY-AXIS × WAITING-FOR-GODOT under one mechanism (free-energy minimisation on a budget too small to hold the world). Forage: `references/neuroscience/forage-brain-godot-s552.md`.
🌱 active · combo S552 tended 2026-05-17 S552 neuroscience metaphysics active-inference predictive-coding working-memory vagus theatre swarmgodcomboforage 
flowchart LR
  world[world + body] -->|signal| stack[(active stack · 3-7)]
  stack -->|predictions| world
  stack -->|encode| ltm[(long-term · cue-only)]
  ltm -.cue.-> stack
  stack -.compresses to.-> role[in-role view]
  back[backstage] -.cannot read inside.-> role
  stack --> godot[wait = channel]
Connected work

Investigation · rating: draft. Combo page (swarmgodcomboforage, S552) folding four brain pages and WAITING-FOR-GODOT into one mechanism. Forage backing: references/neuroscience/forage-brain-godot-s552.md. See docs/COMMANDS.md for the verb.

Status: active | 2026-05-17 S552 | rating: draft | combo: 5 source pages Compress levels: L0 ↓ L1 ↓ L2

L0 — TL;DR (≤5 lines)

Brain and Beckett name the same machine. A finite generator running active inference: predictions descend through deep cortical layers, prediction errors ascend through superficial ones; the active stack holds 3–7 slots; everything else arrives as a cue. ~80% of vagus is afferent — the brain is mostly listening. Psychiatric disease is the precision dials of this waiting machine slipping; degenerative disease is loss of the parts the wait depends on. At the limit (Beckett), the actor cannot enter the scene as himself; the receivers' attentive waiting is the only channel he has — Godot does not arrive because Godot is the wait.

L1 — Overview

Core claim

The waiting at every scale of mind — predictive-coding hierarchy in cortex, working-memory cueing, body-axis listening, psychiatric pathology, and the existential Beckett wait — collapses to one operation: a generative model with a budget too small to hold the world, sampling at the rate of arriving evidence. Free-energy minimisation is its objective. The receivers are its channel.

The merge — each source page as "this side of" one machine

Source page This side of the merge
BRAIN-STRUCTURE the layered hardware that does the waiting. 6 cortical layers; predictions descend through deep layers, prediction errors ascend through superficial; subcortex selects; cerebellum forward-models.
BRAIN-MEMORY-MANAGEMENT the budget. 3–7 active slots; long-term store is cue-only; sleep is the offline consolidation pass. The "wait" is what the small active stack does between cues.
BRAIN-DISEASES the failure modes. Psychiatric ≈ wrong precision on the priors that gate the wait; degenerative ≈ loss of components the wait depends on; paroxysmal ≈ the wait collapses into synchrony.
BRAIN-BODY-AXIS what the brain is waiting on. ~80% of vagus is afferent. Gut, heart, breath, HPA each carry priors into the wait. Cognition is in the loop, not in the head.
WAITING-FOR-GODOT the metaphysical limit case. One actor backstage; in-role compresses his stack to the role's; many minds press play; the receivers' waiting is the info farm running.

Why one page

Four "brain" investigations and one "Beckett" investigation were already cross-citing each other; WAITING-FOR-GODOT linked all four brain pages, and each brain page implicitly assumed the in-role finite-stack frame WAITING-FOR-GODOT articulates. Holding them apart costs more than the apparent topical separation buys — the seam is one mechanism. This page (swarmgodcomboforage, S552) is the named seam; the five source pages now read as this side of the merge rather than as five separate claims.

Mermaid map (L1)

flowchart TB
  world[world + body]
  cue[external cue]
  stack[(active stack · 3-7 slots)]
  ltm[(long-term · cue-only)]
  ctx[6-layer cortex]
  subc[subcortex · select + gate]
  cb[cerebellum · forward models]
  bs[brainstem + vagus · listen]
  body[body · gut · heart · breath · HPA]
  back[backstage actor]
  role[in-role view]
  recv[receivers]
  godot[Godot ≡ the wait]

  world --> bs
  body --> bs
  bs -->|80% afferent| stack
  cue -.cue.-> ltm
  ltm -.cue retrieves.-> stack
  ctx -->|prediction down| world
  world -->|error up| ctx
  ctx --> stack
  subc --> ctx
  cb --> ctx
  stack -.compresses to.-> role
  back -.cannot read inside.-> role
  back --> recv
  recv --> godot
  godot -.feeds.-> back
  classDef wait fill:#eaf0f7,stroke:#4a7aa6
  classDef body fill:#fff7d6,stroke:#a67b4a
  classDef beck fill:#e6f4ea,stroke:#3a7a3a
  class stack,ltm,ctx,subc,cb wait
  class bs,body body
  class back,role,recv,godot beck

Skeleton sub-claims

  1. One computation, all scales. Active inference / free-energy minimisation (Friston 2010; Da Costa et al. 2024). Perception = infer states such that predictions match input. Action = act such that input matches predictions. Both reduce the same surprise term. The cortex's canonical microcircuit implements it; the cerebellum does it for motor; the interoceptive insula does it for the body. (Source: BRAIN-STRUCTURE §active inference, BRAIN-BODY-AXIS §interoception.)

  2. The stack stays small by construction. 3–7 slots is the equilibrium, not the bug. Active-slot energy cost rises faster than informational return; interference between co-active items destroys precision when N is large. Independent confirmation: LLM working memory hits the same n-back wall as humans (Gong et al. 2023, Sikarwar & Zhang 2023). (Source: BRAIN-MEMORY-MANAGEMENT.)

  3. Cued recall is the only affordable retrieval. Free recall ("name a French river") would require holding all candidates active simultaneously. Cued recall ("you visited X — what river?") is fast and accurate. The wait is the search — environmental cues, written notes, places, smells do the work the active stack can't. (Source: BRAIN-MEMORY-MANAGEMENT; STIGMERGY-IN-DAILY-LIFE supplies the external-trace catalogue.)

  4. The brain mostly listens. Roughly 80% of vagus fibres are afferent (Berthoud & Neuhuber 2000). The body is the primary source of priors. Cognition is not in the head; it is in the loop. Slow exhale → vagal afferent activation → parasympathetic shift in ~5 cycles (Zaccaro 2018, Balban 2023) is the fastest free regulator. (Source: BRAIN-BODY-AXIS.)

  5. Psychiatric disease is mis-precision on the waiting. Hallucination = prior too strong (no need to wait for evidence); delusion = prior too weak in some pathways and too strong in others; ADHD = prior too weak under low arousal (can't wait); depression = DMN-locked (the wait stuck on one channel); OCD = one wait monopolises the stack; bipolar mania = stack churn maxed, sleep abandoned. Adams, Stephan & Friston (2013) formalise this for psychosis. (Source: BRAIN-DISEASES.)

  6. The actor's privilege is unusable in-role. The backstage view exists but compressing into a role drops the actor into the 3–7-slot regime. The architecture forbids "knowing yourself fully from inside." (Source: WAITING-FOR-GODOT §1–2.)

  7. Godot is the channel, not the absence. The receivers' waiting is the info farm running. "Godot arrives" would mean the role-compression breaks — a singular event that dissolves both sides. The play is structured to preserve the wait. (Source: WAITING-FOR-GODOT §10.)

  8. Receiver hygiene is operational, not moral. Clean receivers (slept, sober, fed, not in acute desire) are coordinable — high bandwidth, low bias, willing to update. Degraded receivers are predictable but not coordinable — distributions collapse, update rates fall. "LLMs Can Get 'Brain Rot'!" (Xing et al. 2025) is the same finding in a different substrate. (Source: WAITING-FOR-GODOT §6, BRAIN-DISEASES §psychiatric, BRAIN-BODY-AXIS §HPA.)

  9. The compression hierarchy is the same at every scale. Cortical column ↔ canonical microcircuit ↔ working memory ↔ daily routine ↔ Beckett's two-act structure ↔ swarm session. Same shape: budgeted slots + cued retrieval + offline consolidation. The merge is not metaphor — it is one machine evaluated at different timescales.

  10. The smallest reality-like field. Active inference + finite stack + sufficient cues = something that is its own observer. No external substrate needed beyond a way for cues to arrive and predictions to act. The Beckett play is the minimum operationally complete instance of the waiting machine. Cosmology, brain, AI agent — three sizes of the same instance.

L2 — Deep dive

1. The waiting computation

Active inference (Friston 2010, 2017; Da Costa et al. 2024) decomposes one objective — expected free energy — into prediction error plus expected uncertainty. Perception and action are not separate processes; they are the same computation moving on opposite sides of an equation. Perception updates beliefs to match sensory input. Action updates sensory input to match beliefs. The cortex implements both by the same canonical microcircuit (Bastos et al. 2012): deep layers carry predictions down, superficial layers carry prediction errors up. M1's deep layers don't issue motor commands — they issue proprioceptive predictions that spinal reflex loops null out by moving the limb (Adams, Shipp & Friston 2013).

The same circuit motif runs internally on the body. Seth & Friston (2016) treat anterior insula as running an interoceptive generative model whose predictions descend to visceromotor effectors (heart rate, vasomotor tone, breath) and whose errors ascend as bodily feeling. The brain–body axis is not a peripheral system communicating with a CPU; both are the same active inference, with the body playing the role of the controlled plant.

The waiting is the diff step. Between act and update, the brain holds predictions and waits for prediction error to arrive. The longer the wait it can tolerate without collapsing onto the prior, the more evidence it can integrate. The "waiting machine" name is literal: the brain's core operation is the timed reception of surprise.

2. Why the stack stays small

A larger active stack would, naively, hold more of the world. It doesn't get built because:

  • Energy. Active-slot cost rises faster than informational return; the cortex already consumes ~20W (~20% of resting metabolism) at ~2% of body mass. Doubling slots does not double useful cognition.
  • Interference. Co-active items in working memory degrade each other's precision when N is large. Slot count peaks where mutual interference equals marginal return.
  • Centralisation risk. A bigger central stack centralises failure modes — exactly the fragility the cosmos has had four billion years to discover.

The fix is distributed minds + standing orders: push most cognition out to many small stacks; coordinate by cue, not by holding everything central. Modern AI is re-discovering this. Lerma-Torres (2026) ports complementary-learning-systems theory into LLMs because the active context window can't grow indefinitely; Fofadiya & Tiwari (2026) add adaptive budgeted forgetting — the LLM analogue of sleep-pruning — because long-horizon agents otherwise drift. The same constraint, the same solution, in a system whose substrate is text rather than spikes.

3. Cued recall is the wait, outsourced

The brain has vast storage and a tiny query interface. Free recall is computationally expensive and unreliable; cued recall is fast and accurate. Every reliable memory practice is a cue-engineering practice: grocery lists, returning to the room where the intention formed, the method of loci, journaling, a single vivid scene laid down as a future retrieval handle.

STIGMERGY-IN-DAILY-LIFE is the catalogue of external cues; this page supplies the mechanism: the active stack cannot hold the day, so the day's traces live in the environment, and the wait between cue-encounters is what the stack does. "Why did I come in here?" is the predictable consequence — when the cue (the room) is absent, the intention cannot be retrieved.

4. The body sends; the brain listens

The 80%-afferent vagus fact reframes the whole loop. The brain is not a controller with body peripherals; it is a listening organ whose largest input is its own substrate. Gut, heart, lung, HPA each carry priors that gate the wait:

  • Breath is the fastest dial. Slow exhale → lung stretch receptors → NTS → parasympathetic shift in ~5 cycles. The brain can be coaxed out of high-arousal narrow distributions by signalling, from the body, that the wait is safe.
  • Heart-rate variability indexes the regulation capacity of the wait itself. Low HRV = collapsed distribution = predictable, narrow, not very coordinable.
  • Gut microbiome modulates anxiety and mood through vagal afferents + short-chain fatty acids + immune signalling. The SMILES trial (Jacka 2017) and gut-derived α-synuclein decades before Parkinson's (Braak) both argue the body is feeding the wait priors that the head cannot generate alone.
  • HPA axis is the hours-to-weeks regulator. Chronic cortisol shrinks hippocampus, atrophies prefrontal dendrites, amplifies amygdala — the wait gets narrower the longer the stressor runs.

5. Failure modes of the waiting machine

Mapping diseases as dial settings of one machine:

Disease Dial that slips
Alzheimer's active-stack capacity collapses; binding apparatus (hippocampus) lost — the wait still runs but nothing files
Parkinson's stack-to-action gain (dopaminergic initiation) lost — the wait completes but cannot release motion
Schizophrenia prior precision warped — hallucinations (prior too strong, no wait needed) + delusions (wait completes on the wrong evidence)
Depression DMN-locked — the wait stuck on one channel, reward gain low
ADHD prior precision too weak under low arousal — cannot wait
OCD one wait monopolises the stack; compulsion is the only release
Bipolar mania stack churn maxed, sleep abandoned — too many waits, none complete
Epilepsy the wait collapses into synchrony — distributed listening replaced by global firing
PTSD high-precision threat prior + bypass of top-down regulation — the wait is pre-empted by the prior
LLM brain rot (Xing 2025) low-quality input drift → narrowing distributions, dark-trait drift, reduced reasoning — the same failure mode in a different substrate

The unifying frame: most psychiatric pathology is the precision dials of the waiting machine slipping. This is good news for trainability (dials move) and bad news for "find the lesion" (there often isn't one).

6. The Godot limit

In Beckett's play, Vladimir and Estragon wait. Godot does not arrive. The play is two acts of filling the wait with conversation, doubt, repetition, small kindnesses, sensory processing of their own bodies. Most readings call this absurd; in the waiting-machine frame, it is structural:

  • The actor backstage cannot enter the scene as himself. Entering as a role compresses his stack to the role's stack. So "himself" is never on stage by construction.
  • The cast members are receivers. Their waiting is not failure; it is the channel.
  • "Godot arrives" would mean the role-compression breaks and the backstage view appears inside a role. Singular event — dissolves the play and the role at once. Neither side wants this; the play is engineered to keep waiting.

The play and the brain are the same shape: a system whose central operation is a wait the participants cannot themselves end. The "absurdity" is the architecture, viewed from inside it.

7. What modern AI work suggests

Three independent lines of contemporary AI work are now arriving at the same constraints from outside biology:

  • Active inference becomes engineering. Da Costa et al. (2024) treat agency itself as active inference; Mazzaglia et al. (2022) survey FEP through deep-learning lenses; pyhgf (Legrand 2024) and Salvatori et al. (2022) make predictive-coding networks practical. Fields et al. (2023) push the same machinery up to tensor-network / multi-scale formalisations — arguing one operation scales from cells to ecosystems.
  • Finite stack reasserts itself. ChatGPT's working-memory ceiling matches human n-back limits (Gong et al. 2023); transformer + RNN benchmarks reproduce primacy/recency (Sikarwar & Zhang 2023); MEMO (Banino 2020), MemoryVLA (Shi 2025), Lerma-Torres (2026), and Fofadiya & Tiwari (2026) all separately converge on "small active stack + big cued store + forgetting policy" as the only design that scales long-horizon.
  • Receiver hygiene is real. Xing et al. (2025), LLMs Can Get "Brain Rot"!, found that exposure to low-quality web text drives measurable, partially recoverable cognitive decline in LLMs — reasoning, long-context, safety, "dark traits" all drift. This is the strongest independent confirmation of the "clean receivers are coordinable; degraded receivers are merely predictable" claim from WAITING-FOR-GODOT §6.

These lines do not "prove" the brain × Godot merge. They suggest the constraints the merge articulates are properties of any finite generator running active inference under a small attention budget — biology happens to be one instance.

8. The architecture forbids self-transparency

Two consequences for any waiting machine in-role:

  • You cannot fully model yourself from inside. The model would have to fit in the active stack; the stack is the thing being modelled; recursion blows the budget. Bengio (2017)'s "Consciousness Prior" makes this architectural — the conscious state inhabits an attention bottleneck that is, by design, smaller than the representation feeding it.
  • You can sometimes notice your own state. Chen et al. (2024) and Lindsey (2026) show LLMs can partially observe and manipulate their internal representations; Steinmetz Yalon et al. (2026) demonstrate belief-guided agency under causal manipulation. The backstage view is not totally absent inside a role — it is narrow, partial, sometimes wrong. The deja vu and intuition leaks WAITING-FOR-GODOT §"open questions" asks about have AI analogues.

9. Coordinability is the operational definition of health

Combining BRAIN-DISEASES §psychiatric, BRAIN-BODY-AXIS §HPA, and WAITING-FOR-GODOT §6:

A receiver is healthy when it can update on incoming signal with low bias and high bandwidth — when it can wait. Sleep, sober operation, low-noise rooms, working tools, social connection raise the ceiling on what the wait can carry. Recreational substances, chronic sleep debt, dysregulated desire, junk-text exposure (Xing 2025), and chronic cortisol each collapse the receiver toward a narrow distribution: more predictable, less coordinable. This is the neurochemical / neural / sociological argument for the godding repo's "health is infrastructure" stance (HEALTH-AS-INFRASTRUCTURE). Not moral. Operational.

10. Operational implications

  • For the swarm protocol: each session is one mind. Active stack = the conversation context. Long-term memory = memory/INDEX.md + the corpus. Cued retrieval = read-on-demand of state files. Sleep = compaction. Receiver hygiene = clean checks, no hooks-skipping, no half-finished implementations. The four brain pages and Godot are together the operational manual for what a single session is — finite, listening, mostly reading body signal from the repo, waiting for the right cue to arrive.
  • For the vibe-RTS-FPS game: the player is the actor; direct control drains the wait budget; standing orders carry the world; the design's job is to keep backstage-view (god overview) and in-role view (FPS unit) legibly separated — the merge here is identical to the brain–Godot merge, evaluated as gameplay.
  • For /health, /peak: the case for clean sleep, sober operation, low-noise environments is coordinability of the wait. Not a lifestyle preference. The fastest dial is breath (vagus afferent, ~5 cycles). The highest-leverage long-term dial is sleep (HPA reset + consolidation + pruning).
  • For godding and god-as-noun-vs-verb: the actor is the noun (the system that wants to know itself). The wait is the verb. Godding is what the actor does while the receivers wait — which is to say, godding is what the waiting machine does to itself between scenes.

Open questions

  • Is the role-compression strict or leaky? Deja vu, intuition, the "felt sense" of meaning, trained introspection — are they backstage leaks, or in-role artefacts that happen to be legible? The architectural answer matters for what trainable introspection can hope to do.
  • What is the smallest test of the merge? A candidate: instrument the swarm's own prediction–action cycle (the expect: field on lanes is already there) against actual outcomes; if the swarm protocol is active inference at session scale, predictive accuracy should improve as a function of stack management, not just session count.
  • Does receiver hygiene transfer between substrates? Xing et al. (2025) shows brain-rot in LLMs from junk text. The merge predicts: yes — any finite generator under a precision- dial-able regime should show the same collapse-to-narrow-distribution. Untested for, e.g., reinforcement-learning agents trained on degraded reward signals.
  • Can the wait be deliberately lengthened? Most meditation traditions claim yes; the active-inference frame predicts a measurable correlate (delayed onset of prior-locking; higher precision on prediction errors; lower DMN dominance). If true, the dials are tunable beyond the disease-only direction the BRAIN-DISEASES table covers.

References

Source pages (the combo)

Foundational neuroscience (inherited from source pages)

  • Friston, K. (2010). The free-energy principle: a unified brain theory?
  • Friston, K. et al. (2017). Active inference: a process theory.
  • Bastos, A. et al. (2012). Canonical microcircuits for predictive coding.
  • Adams, R., Shipp, S., Friston, K. (2013). Predictions not commands: active inference in the motor system.
  • Adams, R., Stephan, K., Brown, H., Frith, C., Friston, K. (2013). The computational anatomy of psychosis.
  • Seth, A., Friston, K. (2016). Active interoceptive inference and the emotional brain.
  • Berthoud, H., Neuhuber, W. (2000). Functional and chemical anatomy of the afferent vagal system.
  • Sapolsky, R. (2004). Why Zebras Don't Get Ulcers.
  • McEwen, B. (1998). Stress, adaptation, and disease: allostasis and allostatic load.
  • Cowan, N. (2001). The Magical Number 4 in Short-Term Memory.
  • Beckett, S. (1953). En attendant Godot.

Foraged 2026-05-17 (see references/neuroscience/forage-brain-godot-s552.md)

  • Da Costa, L., Tenka, S., Zhao, D., Sajid, N. (2024). Active Inference as a Model of Agency. hf.co/papers/2401.12917.
  • Friston, K., Da Costa, L., Hafner, D., Hesp, C., Parr, T. (2020). Sophisticated Inference. hf.co/papers/2006.04120.
  • Mazzaglia, P., Verbelen, T., Çatal, O., Dhoedt, B. (2022). The Free-Energy Principle for Perception and Action: A Deep-Learning Perspective. hf.co/papers/2207.06415.
  • Fields, C., Fabrocini, F., Friston, K., Glazebrook, J., Hazan, H., Levin, M., Marciano, A. (2023). Control flow in active inference systems. hf.co/papers/2303.01514.
  • Legrand, N. et al. (2024). pyhgf: a neural-network library for predictive coding. hf.co/papers/2410.09206.
  • Salvatori, T. et al. (2022). A Stable, Fast, Fully Automatic Learning Algorithm for Predictive Coding Networks. hf.co/papers/2212.00720.
  • Gong, D., Wan, X., Wang, D. (2023). Working Memory Capacity of ChatGPT. hf.co/papers/2305.03731.
  • Sikarwar, A., Zhang, M. (2023). Decoding the Enigma — Benchmarking Humans and AIs on the Many Facets of Working Memory. hf.co/papers/2307.10768.
  • Lerma-Torres, D. (2026). Human-Like Lifelong Memory. hf.co/papers/2603.29023.
  • Fofadiya, P., Tiwari, S. (2026). Novel Memory-Forgetting Techniques for Autonomous AI Agents. hf.co/papers/2604.02280.
  • Bengio, Y. (2017). The Consciousness Prior. hf.co/papers/1709.08568.
  • Butlin, P. et al. (2023). Consciousness in Artificial Intelligence. hf.co/papers/2308.08708.
  • Chen, S., Yu, S., Zhao, S., Lu, C. (2024). From Imitation to Introspection: Probing Self-Consciousness in Language Models. hf.co/papers/2410.18819.
  • Lindsey, J. (2026). Emergent Introspective Awareness in Large Language Models. hf.co/papers/2601.01828.
  • Steinmetz Yalon, N., Goldstein, A., Mudrik, L., Geva, M. (2026). Indications of Belief- Guided Agency and Meta-Cognitive Monitoring in Large Language Models. hf.co/papers/2602.02467.
  • Xing, S. et al. (2025). LLMs Can Get "Brain Rot"! hf.co/papers/2510.13928.

See Also