Dumb Intelligence
By Jessica Lake
Today I started by tidying up a paper title Brain versus Mind and ended up in a different universe.
Long term, I am exploring how the human nervous system might work. But I approach it from a logical layer a few degrees removed from neuroscience.
My aim is not to muck with neurons. It is to identify a logical architecture that produces the neurobehavioral we observe. That allows me to apply a lifetime of electrical engineering and computer science to the understanding.
Moreover my neurodivergence—ASD, SDAM, and Aphantasia—places me in a peculiar position. I experience life once and recall knowledge.
To quote what oft said by me:
Experiences are the examples.
Knowledge is the rule you learn.
In my lifetime that distinction keeps appearing everywhere.
One thread I revisited today was language. I love language. On and off for my life I have treated language as an organic telecommunications protocol. Not a grammar. Let us look
at a stateful exchange. The speaker initiates a handshake with the listener to establish the basis of transferral. After they synchronize shared state to effect an efficient exchange. Lastly, comes the highly compressed conveyance.
This led me back to Strategic vision. That shared map of reality that expedites F2F exchanges.
For months I have been badgering myself with the question:
Why would the brain integrate a panoramic model of reality only to rip it apart again?
In other words if navigation is satisfied by local subsystems, why bother to integrate everything into a Strategic Vision merely to rip it asunder?
The traditional answer is the ascendancy of consciousness. This bugs me since it leads toward the Homunculus loop.
Instead, I entertained alternate possibility:
Strategic Vision exists to permit communication synchronization.
There I have said it. Not language; but involved party synchronization.
Consider the coordination of a herd of wildebeest in the absence of language.
One animal senses predator.
One animal freezes.
The others freeze.
One animal runs.
The others run.
Alarm state propagates.
The synchronization of the herd state predates language.
Later, language is a mighty refinement plastered atop synchronization.
This led me to proposed hierarchy:
Strategic Vision
Shared State
Synchronization
Communication
Language
Culture
At this juncture I said what sounded trivial:
The easier state synchronize of speaker and listeners, the more efficient overall communication becomes.
That observation immediately connected language, culture, religion, customs, law, friendship, and civilization.
All of the above can be considered as means that foster synchronization between individuals.
Then I stumbled upon a marvel. And I coyly confessed:
“I have a secret. Promise not to tell.”
“No. The secret is not about language!”
The secret is:
Evolution never asks questions.
I had been meditating on whether language en vase was “weak.”
Dhua!
That is an engineer’s question.
Evolution questions nothing.
And that correction reorganized everything.
X and Y no longer needed to exhibit cause and effect. Culture did not evolve to support language. Nor did language evolve to support culture.
Instead, anything that improved the synchronization stateful language could improve corporal fitness. And natural selection favors fitness. The anthropomorphism of “intelligent design” comes afterward.
That led me directly into a reorganization of my understanding of Darwin’s theory of evolution.
Admittedly evolution is stochastic.
Only in hindsight does it appear intelligent.
That said, it sure looks like it is intelligent.
Vision improves.
Skull shape changes.
Posture changes.
Head musculature changes.
Everything observed seems indisputably to be goal directed.
What could give this apparent intelligence?
To begin with selection acts holistically on the specimen.
In systems there is often a sticking place that slows everything else down. This is a bottleneck. Once you fix it the performance skyrockets. Until you find the next bottleneck.
Take dinosaurs and feathers. The first to glide away from a predator had an obvious selective advantage. This overwhelming benefit suddenly alters the worth of allele that aids flight.
lighter bones
larger wings
stronger flight muscles
better balance
When the selective benefit is big things happen fast. Evolution exhibits apparent intelligent.
A healthy population of a species possess a sizable reservoir of allele. Let us define affinity as a statistical index that rate the potential for natural selection. You can gauge each allele as beneficial, neutral, or detrimental to natural selection.
What I believed above is wrong.
What I discovered is that natural selection operates on the combinations of allele in an individual. Individually a pair of allele might be detrimental. But together that may be judged a benefit. A novel allele might recast the affinity of any number of extent allele.
A better understanding of evolution became one of the major insights of the day:
X: In evolution there is no cause and effect.
Let us suppose a novel allele appears. Within any generation an individual possess some combination of allele. You can consider the benefit of the novelty only in concert of an individual’s combination of allele.
Regardless of the order that the phenotypes bellow appear in a population flight becomes possible.
lighter bones
larger wings
stronger flight muscles
better balance
Apparent Intelligence
A flaw in our localized approach of the nervous system exists.
How can prediction loops resolve disagreements without intelligence?
A common solution in when sensor disagree is a “voting” mechanism.
Each subsystem contributes a ballot with these properties:
action proposed
confidence in proposal
Peril without resolution
urgency of resolution
The method of arbitration is heuristic. Resolution emerges from an arbitration algorithm.
And best of all no homunculus required.
The same pattern kept reappearing:
Evolution:
Variations → Selection → Organism
Reasoning:
Candidates → Constraints → Conclusion
Action:
Predictors → Arbitration → Behavior
Communication:
Individuals → Sync→ Shared State
Again and again, coherence emerged from distributed processes.
That brought me back to the challenge of consciousness.
For years I have whittling away responsibilities from the consciousness:
free will,
decision making,
action selection,
prediction,
memory reconstruction,
reasoning.
Each time the system continued functioning. To my mind the conscious mind increasingly resembles an observer. No offense.
Apparent Intelligent
Then I lighted on a fascinating possibility.
It is a common fallacy that evolution requires an intelligent designer. Some mouse got lucky seems a poor way to build a better mouse. Yet Darwin showed that evolution itself is not intelligent.
What if human cognition works the same way?
Not because cognition is evolution. But because evolution is mechanistic but is judged by many as intelligent.
That became the key connection. Not that evolution is the answer. Just that it proofs a mechanism can appear intelligent.
At this point I stopped thinking about ideas and started thinking about predictors. Predictors seemed a much better than models. It is the validity of predictions that is the goal.
What is our prediction loops imitated evolution. We seemed to have all the pieces.
Knowledge is prediction.
Predictors compete.
Reality serves as the proving ground.
Accurate predictors prosper.
Inaccurate predictors generate anomalies.
A predictor that failed now succeed before. Therefore we keep them available.
This suggested a new architecture:
Proving Grounds
Organism
Mechanism
The mechanism builds prototypes.
The proving grounds evaluates them.
And the mechanism adapts.
Predictor Topology
At first I imagined predictors existing on a flat plane.
That felt wrong.
Hierarchies seemed a suitable form for predictors. Predictors feeding predictors. Higher-level predictors influence the value of lower-level predictors.
A bottleneck in systems is a pain point. It is the worst offender in distributed architectures.
In semantic memory a bottleneck would likely be over or under constraints.
When we break a bottleneck:
latent possibilities become favorable,
competing predictors are reevaluated,
previously neutral alternatives become adaptive.
Again:
No directed graphs ergo nothing propagates.
As a thought experiment I imagined accelerating evolution by, say, a million-fold.
To the observer, the process would look exactly like intelligence. Yet underneath there remains only variation, selection, retention, and adaptation.
This led to the fascinating possibility:
What if cognition operates through a similar mechanism?
Not because cognition resembles evolution. But because evolution demonstrates that intelligence-like outcomes can emerge from non-intelligent processes.
I am not trying to build intelligence.
My ambition is identify a mechanistic architecture that to the observer appears intelligence— That is apparent intelligence.
The exact underlying implementation is secondary. Neurons are one implementation. Silicon could be another. Quantum computing.
It’s the logical architecture that matters.
This leaves me with a set of constraints:
No magic.
No homunculus.
Local operations only.
Environmental evaluation.
Adaptive reorganization.
Prediction as fitness.
Retention of alternatives.
Bottleneck breaking.
Hierarchical predictors.
Stability and plasticity.
The strongest constraint may be this:
The apparent intelligence must emerge from the mechanism.
Intelligence itself cannot be used as an explanatory primitive.
That is where I ended the day.
Not with an answer.
With a possibility.
If evolution can generate the appearance of design through mechanistic adaptation, perhaps cognition can generate the appearance of intelligence through mechanistic adaptation as well.
The frightening part is not that it might be true.
The frightening part is how natural it suddenly seems.