Intelligence,
tested in motion.
We're studying the connection between neural structure, sensory input, action, and learning. The first experimental environment is deliberately small: Snake.
Can the wiring of a fruit-fly nervous system help control an agent?
We are investigating executable connectome-derived neural systems and how they connect to perception, feedback, and action. We use small, testable environments to avoid confusing an impressive simulation with demonstrated learning.
Explore the Snake environment ↗From pixels to behavior.
Our experimental architecture distinguishes the environment, sensory encoding, neural dynamics, decision policy, and evaluator. That separation is necessary if we want to know what caused an improvement.
Progress, not mythology.
Physical sensory-action pipeline
Infrastructure for screen-derived observation, controller execution, and external input has been implemented and tested.
Connectome-derived execution loop
A MaleCNS-based recurrent neural execution path has been exercised in a physical closed-loop setup. This is not proof of trained game-playing intelligence.
Learning and scientific controls
Separate retina-only reinforcement-learning experiments have achieved limited measurable Snake performance, with accepted and rejected checkpoints tracked in lab records. Demonstrating that the biological connectome itself contributes to learning remains an open research objective.
Claims we aren't making.
We are not claiming a conscious digital fly, a fully autonomous brain that has mastered Snake, or proof that anatomical connectivity alone produced learning. The published connectome is a wiring map, not a plug-and-play game agent.
Retina-only learner results are control experiments, not evidence of connectome-derived learning. Current work is about distinguishing those effects through controlled evaluation.
Built on open science.
Fruit-fly connectomics is a collaborative scientific field. SomSoft is independently experimenting with methods and software inspired by published neuroscience; we are not affiliated with the research institutions below.
Research-stage descriptions above reflect internal experiment documentation reviewed October 2026 and may change as further tests are completed. No affiliations or endorsements are implied.