A neural network wired to a fruit fly connectome learned to play Doom

A developer used AWS's open-source Strands coding agent to build a recurrent network on a real fruit fly connectome (49,393 neurons, 9 million synapses) and trained it to play Doom by distilling five ordinary PPO agents. The student matched its teachers, while 12 GRPO fine-tuning runs failed to beat it beyond evaluation noise.

• Network built on the fruit fly connectome: 49,393 neurons and 9 million synapses
• Training took 5 days, 28 commits and about 3,300 lines of code
• The student matched its five PPO teachers but did not surpass them
• 12 GRPO runs produced no gain beyond evaluation noise