Author: Curt Doolittle

CURT: Q: –“Why can’t AI’s really reason?”–

Well, Think of AI’s as either associating, or associating and predicting via pattern matching. That’s equivalent to auto-association in the brain’s hippocampal region.

But ‘reasoning’ requires ‘wayfinding’. The entirety of the neocortex is an evolution of lower level cortical tissues that consist of only three layers, all of which are devoted to motion in space. (wayfinding). At some point the neocortex was formed by doubling over the three layers itnto six, creating potential for more abstract pattern matching. Our higher functions all evolved out of wayfinding. The hierarchy of memory evolved the capacity to wayfind not only with space but with thoughts, ideas, and language.

So, how do we take LLMs and add reasoning BELOW their processing, to reflect the wayfinding that the brain employs when reasoning?

We imitate this with chain of thought or chain of draft agents driving the LLM’s processing. THere is some emergence of a world model somewhat equivalent to episodic memory in the most advanced LLMs with the largest number of tokens, parameters and memory.

But compared to the brain, we would ask the AI’s to either perform sequential and comparative or parallel ‘hypothesizing’, then compete those parallel ideas for greatest reward shortest time, least effort, greatest certainty, lowest risk, and work through the steps necessary to change state from A to B and so on. Without adversarial competition (a market), and recursion (theorizing at each step) it would be difficult to imitate the brain.

In our work at NLI and Runcible we categorize the problem and then kick off a chain of thought to produce the answer. But this is because we are largely trying to determine the testifiability and reciprocity of an statement. It’s a WHAT question. HOW questions, require wayfinding.

HERE IS THE NEUROSCIENCE:
The human neocortex consists of approximately 180 distinct cortical regions per hemisphere (360 total) according to the most detailed recent maps (Glasser et al., 2016). Within each region, neurons are structured into approximately 1 million cortical columns (macrocolumns) per hemisphere, each macrocolumn containing around 100–120 minicolumns. Minicolumns, each containing approximately 80–120 neurons, serve as fundamental computational modules that detect specific patterns of input.

These cortical regions perform specialized functions (perception, attention, language, decision-making, and motor actions), yet they operate upon a common principle of hierarchical signal processing. Information processed within minicolumns aggregates upward to macrocolumns, then cortical areas, culminating in a distributed yet convergent representation projected into the hippocampal-entorhinal system, which indexes events as episodic memory.

Step-by-step causal process of indexing and prediction:

Sensory Integration & Cortical Hierarchy:Sensory inputs propagate upward through cortical layers (minicolumn → macrocolumn → cortical region → interregional network).

Each cortical region extracts specific dimensions (features, relations, contexts) through progressive abstraction.

Hippocampal Indexing of Episodic Memory:
The hippocampus receives highly processed, multimodal cortical outputs via the entorhinal cortex.
It rapidly encodes sequences and contexts, creating a sparse, unique neural pattern—an episodic index.
These episodic indices do not store complete memories; rather, they provide “addresses” linking cortical representations to allow future retrieval and auto-association.

Auto-association & Prediction:
Later, partial activation of sensory or conceptual input activates the hippocampal index.
This reactivates the entire cortical ensemble involved in the original episode (auto-association), allowing the brain to “fill in” missing information from partial cues.
Such pattern completion forms the neurological basis for prediction, enabling anticipation of sensory inputs and outcomes based on past episodic associations.

How does the brain perform ‘wayfinding’ (transforming present states into desired future states)?

Wayfinding refers to the neural process by which the brain navigates through conceptual, physical, or abstract spaces, turning present circumstances into intended future outcomes. The mechanism involves the integration of hippocampal indexing, cortical maps, and predictive simulations into goal-directed paths:

Hippocampal Formation & Cognitive Maps:
The hippocampus constructs spatial, temporal, and conceptual maps, maintaining relational positions through place cells (spatial), time cells (temporal sequencing), and concept cells (relational/semantic mappings).
These “cognitive maps” represent dimensions of experience—whether physical space, temporal sequence, or abstract conceptual structures—allowing mental navigation.

Entorhinal Cortex & Grid Cells:
Grid cells in the medial entorhinal cortex generate structured positional grids across dimensions of representation, providing metrics or scales for space, time, and concept positions.
These grid cells act as coordinate systems, tracking relative positional changes and transitions between states.

Mental Simulation & Prediction:
Cortical minicolumns encode incremental transitions (small state-changes), macrocolumns aggregate transitions into meaningful sequences, and regional networks simulate complete transitions through hierarchical integration.
The frontal and parietal cortices, in collaboration with hippocampal-entorhinal circuits, run simulations by iteratively activating episodic indices corresponding to potential future states.
The brain evaluates alternative simulated paths against internally represented goals or externally motivated states.

Valence & Goal Selection:
Valence (emotional or motivational value) attached to episodic indices or conceptual representations biases simulations toward states of increased reward or reduced cost.
Goal-directed behavior emerges from iterative mental simulations where multiple future states are evaluated against predicted rewards or penalties, leading to preferential selection of optimal paths.

Execution via Motor Outputs:
After a path is selected, cortical motor and premotor areas activate motor columns and minicolumns in motor cortex, translating internally simulated transitions into real-world actions.
This loop repeats in real-time feedback cycles, continuously revising predictions, simulations, and actions as the organism moves closer to its intended goal.

Operational Summary of Wayfinding:
– Cortical minicolumns encode present state.
– Hippocampal indexing encodes episodic references.
– Grid/place/time cells provide coordinate frames (positional dimensions).
– Mental simulation evaluates multiple potential future states.
-Valence weighting directs attention toward optimal state transitions.
– Motor columns execute real-world transitions, continually corrected by sensory feedback loops.

Thus, wayfinding emerges neurologically as a recursive feedback process integrating memory indexing, predictive simulations, and continuous environmental updates, enabling human cognition to bridge present and desired future states effectively and adaptively.

Reference (as requested):
Glasser et al., “A Multi-modal parcellation of human cerebral cortex,” Nature, 2016: https://t.co/HNRKKEYuJ3

Cheers