swat-or-buy
Ad creatives scored by a MaleCNS retina/lamina running in the browser. You pick first; the fly “BUYs” glance salience and swats the rest. Deep buy/swat cells are silent under LIF (graded optic lobe); the live show uses honest glance/boredom.
Neurons are simulated — leaky integrate-and-fire or better — with documented, tuned parameters.
Dispute this rating →The evidence
No evidence, no level.Neurons are leaky integrate-and-fire with explicit rest, membrane tau 20 ms, synaptic tau 5 ms and exponential subthreshold evolution.
https://github.com/oskarmalmwiklund/swat-or-buy/blob/a9876261150534fbe747d468e6491dec954c8044/swatorbuy/fly/kernel.cpp#L29-L30Spiking uses a -45 mV threshold with a refractory check, and firing resets the neuron to rest and clears its conductance.
https://github.com/oskarmalmwiklund/swat-or-buy/blob/a9876261150534fbe747d468e6491dec954c8044/swatorbuy/fly/kernel.cpp#L46The brain that scores every ad is constructed with its weights frozen, and every simulation step is called with learning disabled.
https://github.com/oskarmalmwiklund/swat-or-buy/blob/a9876261150534fbe747d468e6491dec954c8044/swatorbuy/protocols.py#L24-L31The project's own model card states no training happens and that the KC-to-MBON plasticity rule is present but frozen.
https://github.com/oskarmalmwiklund/swat-or-buy/blob/a9876261150534fbe747d468e6491dec954c8044/swatorbuy/fly/prepare.py#L103The inherited plasticity rule returns before touching any efficacy when the frozen flag is set.
https://github.com/oskarmalmwiklund/swat-or-buy/blob/a9876261150534fbe747d468e6491dec954c8044/swatorbuy/fly/rule.py#L45-L46
MaleCNS v1.0 run as leaky integrate-and-fire cells at 0.1 ms in a C++ kernel with explicit rest, membrane tau 20 ms, synaptic tau 5 ms, a -45 mV threshold, refractory period and reset-to-rest, so membrane voltages evolve under a documented neuron model. Not L2, and this one needs saying carefully: the repo does contain a real dopamine-gated anti-Hebbian KC-to-MBON plasticity rule inherited from its upstream (swatorbuy/fly/rule.py, adapted from Huang/Luo et al. 2024), but it is explicitly switched off in every scoring path - the scoring Fly sets weights_frozen = True at construction and passes learning=False on every step, and the project's own model card records 'training: None. Weights stay at the compiled baseline; the KC-to-MBON07/11 plasticity rule is present but frozen.' No synapse changes during a run. The README's own caveats also state there is no neuromodulation, and the model gate section honestly reports that everything past the lamina is silent under every stimulus tried.
Record
Nearby
All projects →Ratings on thelearningfly.com are proposals, not verdicts. Every one of them is arguable in public.