Beautiful Interactive Neural Nets running locally on ESP32s.
This idea was born out my realization that much of modern discourse around improving model performance is adding more parameters. In the era of LLMs, models with millions of parameters are deemed obsolete in favor of billion- and even trillion-parameter models. Even in vision, with the advent of billion-parameter Vision Transformers that exceed human performance at many classic image tasks - it made me wonder - how small can we make an effective model? And how can we present this model in a compelling, beautiful way? This project is a foray into those questions.
Requires: TFT_eSPI library, defailt ESP32 libraries (ESP-NOW for networking)
For EACH module you wish to set up, wire according to the pinout in User_Setup.h (in the TFT_eSPI library files). Wire backlight to Vcc, as backlight functionality is not useful here.