139 lines
6.9 KiB
Org Mode
139 lines
6.9 KiB
Org Mode
:PROPERTIES:
|
|
:CREATED: [2026-06-01 Mon]
|
|
:WEIGHT: 61
|
|
:ID: f6a7b8c9-0d1e-2f3a-4b5c-6d7e8f90abcd
|
|
:END:
|
|
#+title: Biomimicry in Passepartout
|
|
#+filetags: :passepartout:architecture:neurosymbolic:biomimicry:p150:
|
|
|
|
**Biomimicry in Passepartout**
|
|
|
|
**What already exists (real biomimicry, not metaphor)**
|
|
|
|
| Feature | Biological analog | Implementation |
|
|
|---------+-------------------+----------------|
|
|
| Three-layer reasoning | Reptilian → limbic → neocortex | LLM (intuition) → Screamer (constrained search) → ACL2 (verified reasoning) |
|
|
| Verdict-overrides-LLM | Somatic markers override conscious deliberation | Gate outputs overrule LLM proposals, not the other way |
|
|
| Dream cycle | Sleep consolidation | gbrain dream cycle: replay and re-index daily experience offline |
|
|
| Delegate subagents | Cognitive recruitment | delegate_task — spawns specialized subprocesses for subproblems |
|
|
| Memory as two systems | Declarative vs procedural | Fact store (explicit) vs skills (implicit/procedural) |
|
|
|
|
**What is missing — and how to fill it**
|
|
|
|
***1. Peripheral nervous system (P150 slot)***
|
|
|
|
Biology does not poll. The brain does not run ~while true: check if_finger_hot()~. Dedicated low-power circuits (nociceptors, proprioceptors) monitor continuously and only signal the CNS on deviation.
|
|
|
|
Passepartout polls everything — cron output, filesystem, user messages. A P150 running 72 parallel event-driven monitors would dedicate:
|
|
|
|
- One core to "is the user typing on Signal?"
|
|
- One to "did the weekly model discovery fail?"
|
|
- One to "is ZFS ARC thrashing?"
|
|
- One to "is the test build running longer than usual?"
|
|
|
|
Each sleeps until something meaningful happens. Only then does it signal the symbolic system. Zero LLM involvement for routine monitoring.
|
|
|
|
This changes Passepartout from a system that responds to commands to a system that notices things on its own. The difference between a calculator and a research assistant.
|
|
|
|
***2. Associative activation (spreading activation)***
|
|
|
|
In the brain, activating one concept (ACL2) automatically pre-activates related concepts (SP3, proof, Lisp, verification). No clean-slate search.
|
|
|
|
Passepartout has no equivalent. Every query is a fresh search. A biomimetic fact store would:
|
|
|
|
- Pre-fetch linked pages when one is loaded
|
|
- Prime caches based on current conversation context
|
|
- Use the graph structure to predict what will be needed next
|
|
|
|
The brain does not pre-fetch — it primes — so the next thought is faster. Passepartout could prime its caches so facts most likely needed next are already loaded.
|
|
|
|
***3. Error-driven learning with local credit assignment***
|
|
|
|
The brain does not backpropagate. Errors trigger local corrections at the synapse that made the mistake.
|
|
|
|
Passepartout's Gate decisions today are either right or wrong, but nothing locally adjusts. A biomimetic Gate would:
|
|
|
|
- Track which rules fired during a wrong decision
|
|
- Locally adjust confidence scores of only those rules
|
|
- No global retrain — just the specific rule that fired
|
|
|
|
This is STDP at the symbolic level.
|
|
|
|
***4. Sleep consolidation (dream cycle upgrade)***
|
|
|
|
The gbrain dream cycle already replays daily experience. It could go further during offline cycles:
|
|
|
|
- Replay the day's decisions, identify which Gate checks were slow
|
|
- Regenerate ACL2 proof caches for rules that changed
|
|
- Prune skills that never fired (neurogenesis pruning counterpart)
|
|
- Re-index fact store based on actual usage, not static linking
|
|
- Propose new skills for repeated multi-step tasks discovered during the day
|
|
|
|
***5. Graceful degradation***
|
|
|
|
Biology has redundant fallbacks at every level. Passepartout has single points of failure.
|
|
|
|
A biomimetic approach:
|
|
|
|
- Gate offline? Fall back to cached rule set
|
|
- LLM offline? Fall back to smaller local model
|
|
- ACL2 busy? Use previously verified boundaries
|
|
- Never go silent — get slower and dumber until primary returns
|
|
- P150 cores can run degraded modes independently
|
|
|
|
**The P150's role**
|
|
|
|
The P150 (72 Tensix cores, 32GB GDDR6, QSFP-DD 800G interconnect) fills a slot nothing else in the build covers:
|
|
|
|
- Not for fast inference (2x 3090s are faster and cheaper for that)
|
|
- Not for baremetal Lisp Machine (FPGA is the right tool for tagged memory + hardware GC)
|
|
- For ambient awareness, parallel verification dispatch, fact store indexing, anomaly detection
|
|
|
|
The P150 is the system's peripheral nervous system — always-on monitoring behind the scenes.
|
|
|
|
**Revised architecture**
|
|
|
|
| Component | Role |
|
|
|-----------+------|
|
|
| 2x RTX 3090 | Fast LLM inference |
|
|
| EPYC (main cores) | ACL2, Screamer, PDS, Gate orchestration |
|
|
| P150 | Always-on temporal awareness, parallel constraint search, fact store indexing, anomaly detection |
|
|
| FPGA (future) | Lisp Machine (tagged memory, hardware GC) |
|
|
|
|
**Temporal awareness: explicit vs ambient**
|
|
|
|
Passepartout today reasons about time (reading logs, comparing timestamps, understanding "before X happened" from context) but has no sense of time.
|
|
|
|
Explicit (current): Reads a cron schedule, orders log events, answers "when did X happen."
|
|
|
|
Ambient (with P150): Notices the build took 3x longer than usual without being asked, flags that message frequency dropped at 3AM, anticipates the user will want the weekly report before they ask.
|
|
|
|
The P150 makes ambient temporal processing economically viable because 72 independent cores running statistical monitors consume near-zero power. Running the same monitors on the EPYC competes with ACL2 and the PDS. Running them on the 3090s wastes bandwidth on non-matrix work.
|
|
|
|
**Relationship to the Pinker/Marcus critique**
|
|
|
|
Pinker and Marcus argue that neural networks (spiking or otherwise) lack compositional syntax and systematic reasoning. A network that learns "A fires before B" through STDP has learned a temporal correlation, not a rule. It cannot distinguish causation, correlation, and coincidence.
|
|
|
|
This critique does not apply to Passepartout because Passepartout is not a pure neural network. It is a hybrid system:
|
|
|
|
| Problem | Mathematics | Where it runs |
|
|
|---------+------------+---------------|
|
|
| Temporal intuition | Statistical pattern detection | P150 |
|
|
| Compositional time (before/after/during) | Symbolic reasoning | Gate + Screamer on CPU |
|
|
| Sequential patterns from data | ANN attention | GPU |
|
|
|
|
The neuromorphic layer gives the system a sense of time. The symbolic layer gives it understanding of time. Both are necessary. Neither one replaces the other.
|
|
|
|
**What biomimicry means here**
|
|
|
|
The real gains come not from replicating brain details (spiking neurons, STDP, ion channels) but from adopting organizational principles that biology evolved:
|
|
|
|
- Specialized subsystems for different time/resource regimes (PNS vs CNS)
|
|
- Asynchronous event-driven communication instead of synchronous polling
|
|
- Redundant fallbacks at every level
|
|
- Local learning that does not require global retraining
|
|
- Offline consolidation separate from online inference
|
|
- Parallel associative retrieval rather than sequential search
|
|
|
|
Passepartout already adopts some of these. The P150 and an upgraded cron/dream cycle would add the rest.
|