8% performance boost with fine-tunning via biology inspired adapters (genbais.com)

We turned cortical tricks into tiny (~1M param) adapters that plug into various neural network architecture models, CLIP, GPT, CNN, etc.. It evolutionarily navigates astronomical searche space of configs in ~1,000 cheap experiments and delivers noticeable (i.e. +8%) accuracy. No retraining the backbone. No crazy compute. Pure neuroscience cheating codes. Pay only if it beats your baseline—no improvement, no payment. Would love brutal feedback from the HN ML crowd — especially if anyone wants to throw their weird VLM at it and see what the brain stuff does