Interactive Benchmark Replay · T-002

Fruit Fly Brain
vs AI

Watch a MaleCNS-derived connectome-constrained model and an LSTM tackle the same delayed-recall challenge — the task where benchmark performance diverges most.

Benchmark replay. Trial outcomes are drawn from measured benchmark accuracy probabilities (A1-BIO+SA-010: 94.8%, LSTM: 57.6%). The trained models are not running live in the browser.

Delaysteps

T-002 · Delayed Recall

A binary pattern is shown briefly. After a configurable memory gap, each model must recall which pattern it saw. Same input, same delay — different architectures.

Benchmark context · T-002 confirmed results (5-seed · held-out test set)

MaleCNS-derived + SA-010

94.8%

± 6.2% (95% CI)

6,514 train params

MaleCNS-derived (frozen) + SA-010

92.4%

± 5.5% (95% CI)

2,852 train params

LSTM

57.6%

± 26.8% (95% CI)

266,626 train params

Random baseline

50.6%

± 6.6% (95% CI)

0 train params

SA-010 configuration (decay 0.05 / leak 0.02) was selected via exploratory sensitivity analysis. Results are task-specific and should not be interpreted as a universal biological conclusion. Full T-002 results →