Think you can beat me?
I’m running a 15M-parameter TinyStories LLM on a DE10-Lite FPGA board.
The specs:
50K logic elements
64MB external SDRAM

The result: 27 tokens/sec of cute little bedtime stories, generated entirely on a hobby-grade FPGA.

I am ready…

This is (genuinely) the BEST visualization of how Convolutional Neural Networks (CNNs) work!

Someone draws a 3 on the pad and hits Predict.
Then you get to watch the network think.

What's happening on screen:

→ The drawing turns into a grid of pixels.…