Dev Suggests Simulating a Fake Fly and Comparing It to the Real Brain Scan Until Behaviors Converge
ctjlewis · x · 2026-09-06
ShimazuSystems shares relevant experience: he built an RL system for a zero-gravity space drone with very similar behavior. His method: run the model, watch it live, use only measured stimuli (no visual RL), measure responses, and eliminate failures against a programmatically ideal arrangement.
He suggests using a model to analyze how a fly moves, simulating a programmatic 'fake fly,' and iterating until the fake fly's behavior matches the real one — without changing the base brain-scan data. He still suspects much of what causes fly behavior may not be derivable from scan data alone.
More from Fun
- Burkov jokes Terence Tao is 'the only mathematician left' on X feeds — burkov · 2026-09-06
- Dev builds complete Roblox idle game entirely with agentic tool Roblox Build — rms80 · 2026-09-06
- Astra recreates Apple exec Craig Federighi in Blender, sparking AI 3D buzz — Dimillian · 2026-09-06
- Ofir Press shares the 2016 EMNLP rejection of his now-canonical weight-tying paper — OfirPress · 2026-09-06
- Redditor's AI Agents Start Feuding With Each Other — symptomsofdementia · 2026-09-06
- "I'm something of a Mozart myself": conducting an agent orchestra from the terminal — prateekj · 2026-09-06