GolfFly credits

Back to the course

A fictional golf game by BMD Pat. Daddy's Money and the Trust Fund Tour are fictional sponsors.

Measured connectivity, modeled activity

Sampled soma positions from MaleCNS v1.0, acquired and analyzed by the FlyEM Project Team at HHMI Janelia, the Cambridge Connectomics Group, and Google Research. Source dataset, CC BY 4.0. Coordinates are sampled, scaled and reoriented for display.

The first release used soma positions, uncoupled activity and handwritten golf decisions. The criticism that it did not run a connectome was correct. This update replaces shot aim, shot distance and caddy club selection with a trained readout of a measured-graph neural model.

47,524 recorded neuron-to-neuron connections, representing 1,605,984 annotated synapses, among 9,854 participating cells. This is a pruned subset of MaleCNS, not a complete reconstruction. The 11,067 displayed soma positions include cells outside the retained graph. The visible wires sample 60 of the actual runtime connections per brain; their curves are illustrative, not traced axons.

The inputs are distance to the cup, fairway and flag bearings, tee status and lie category. A fixed random encoder drives 4,678 visual cells. Activity passes through four recurrent tanh steps on the recorded graph. A ridge-trained readout of 768 other cells selects aim, range and club. There is no direct sensory-to-readout shortcut. Both flies run separate states on this shared pretrained model.

Training: 12,000 synthetic examples from a published, handwritten golf teacher. Only the readout is trained. Connectivity is fixed. The model does not discover golf rules or learn after each swing. Golf rules, sensing, route planning, collision avoidance, action sequencing, conversation wording and animation remain conventional game code. Learned decisions and neural activity gate those systems.

Model assumptions: square-root synapse counts, incoming normalization, simplified neurotransmitter signs, tanh rate dynamics and artificial sensory encoding. These have not been validated as fly physiology. Display pulses encode rate activity; they are not measured or biophysically simulated spikes. Nothing here establishes consciousness or a biological ability to play golf.

Try to break the claim

On 3,000 held-out synthetic cases, the intact model has 4.82 degrees mean aim error, 0.180 m mean range error, and 98.03% club accuracy. Shuffling presynaptic IDs gives 70.04 degrees aim error and 38.30% club accuracy; removing connections gives 76.34 degrees and 41.07%. These tests show dependence on this trained model’s wiring. They do not establish that biological topology is better than a random graph trained from scratch, or measure completed golf scores.

Evaluation results and limits · Data sources, hashes and pruning · Import, training and evaluation code · Runtime model · Model weights and recorded edges

Download the new HDR10 film · Standard-display video

Clubhouse forecast

British weather: every hole is a water hazard.

Materials and rendering

4K grass, sand and bark maps plus the HDR sky are from Poly Haven (CC0): Leafy Grass, Sand 02, Bark Brown 02, and Kloofendal sky. The web build converts surface maps to WebP. Water normals are procedural.

Three.js r170, MIT license. Browser graphics are tone-mapped for ordinary display; the downloadable HDR movie is a separate BT.2020/PQ master.