Fuchan: A Four-Layer Artificial Life Avatar with a Drosophila Connectome as Its Instinct Layer

Abstract

Background: Dialogue agents powered by Large Language Models (LLMs) achieve high fluency in persona-based conversation, yet their internal states are entirely textual descriptions. When such an agent is "angry," the LLM is merely performing anger; no physical substrate corresponds to the emotion. This limitation produces context-dependent instability, non-differentiation of stimuli, and an absence of spontaneous behavior. Proposal: We propose a four-layer artificial life avatar, "Fuchan," in which the male whole-central-nervous-system connectome of Drosophila melanogaster, MaleCNS v1.0 (211,577 neurons, ~25.5 million synapses), serves as the instinct layer. Sensory inputs are propagated through the connectome via sparse matrix multiplication, and the resulting neural activity drives an LLM's language generation. The LLM is confined to verbalizing the instinctual state rather than performing it. Method: We conducted an exploratory observational study of a single session (1 hour 54 minutes: 13,390 telemetry samples; 66 LLM calls) involving a color-matching puzzle and an Ishihara-style color vision test. Results: We observed three emergent behaviors that were not designed: (1) test anxiety, surfacing as natural expressions such as "I'm not confident" and, notably, the negative form "I'm not that scared"; (2) trust building, where the user's encouragement "I won't hate you even if you make a mistake" produced a strong reward pulse (sugar=0.8) and a trust rise from 0.994 to 1.000; and (3) sleep with dream consolidation, in which 38 dialogue log entries were semantically compressed into a single memory ("Color Search with Kurui"). We also observed that biological constraints (the fly's poor red perception) surfaced not as a definitive "cannot see" but as sincere uncertainty, without hallucination. Conclusion: The four-layer architecture, implemented by a single author in less than one month, demonstrates that incorporating biological constraints into a computational system produces qualitatively different behavior from prompt-engineered personas.