Modeling the Self with Neuronal Tokens

Achieving true superintelligence requires generative AI models capable of deeply modeling both human and non-human agents. While state-of-the-art generative models like GPT-4 can already parse emotions, sentiments, and psychological traits with impressive accuracy (Carneros-Prado et al., 2024; Peters et al., 2024), they currently lack cohesive, iterative psychological models capable of dynamically evolving with each interaction.

Existing generative models demonstrate a fundamental capability to simulate personalities through role-playing tasks, maintaining character consistency across multiple conversational turns (Ahn et al., 2024). However, this ability remains limited to transient interactions rather than long-term, evolving representations.

In my recent work, I've built upon this foundation by developing detailed psychological profiles encompassing personality traits, emotional dispositions, triggers and responses, decision-making styles, and personal values. This structured profiling provides generative models with an explicit, cohesive grounding--essentially a rudimentary but comprehensive proxy model of the "self."

When contemplating synthetic minds, some researchers aim to replicate human neural architectures neuron-for-neuron, a method suggested by experts like Yann LeCun. However, this strategy is problematic. The human brain contains roughly 86 billion neurons, yet the cerebral cortex--responsible for most higher cognition--houses only about 16--25 billion (Herculano-Houzel, 2009). The majority of the brain's neurons manage bodily processes unrelated to higher cognition--such as sensorimotor coordination (~69 billion neurons in the cerebellum alone) and autonomic functions like digestion (around 100 million neurons in the enteric nervous system; StatPearls, 2022). Replicating these networks is unnecessary and computationally inefficient for artificial intelligence.

Moreover, a fully simulated mind is biologically impractical. Human brains simplify interactions using internal models of self and others, leveraging mirror neurons and predictive coding mechanisms to efficiently understand behaviors and intentions without exhaustive simulation (Salvatori et al., 2023).

Inspired by this, I propose the concept of "neuronal tokens"--structured, psychological profiles encoded into tokens that generative models use as high-fidelity self-models. These tokens enable models to dynamically project and update detailed psychological states efficiently. They can be human-readable formats, such as JSON psychological profiles, or specialized machine-readable embeddings optimized for rapid retrieval and iteration.

This approach significantly enhances efficiency and flexibility, moving us closer to realistic, continually evolving synthetic cognition without impractical computational burdens. Future research should explore optimal token serialization methods, memory augmentation strategies, and rigorous evaluation protocols to ensure privacy, flexibility, and alignment with intended behaviors.