@hatsunama — 16x breakout

Loading video…

We have just entered a future where you can literally glue super intelligence to anything for like $10.

@hatsunama

3.2K followers · usually 1.1K views a post

Views

10.1K

Likes

672

Comments

106

Why it worked

The video uses a strong, counter-intuitive hook about 'gluing super intelligence' to hardware for $10, which immediately captures interest. It then delivers high-value, technical information in a concise, fast-paced format that feels like a genuine discovery.

Summary

The creator explains how a new microchip, the ESP32-S3, enables running AI models locally and offline. He details the technical specifications and the architectural breakthrough that allows this small chip to handle large models.

Structure

1Introduction to the ESP32-S3 chip
2Technical specifications of the chip
3Explanation of the architectural breakthrough
4Benefits of local AI execution

Call to action

Hit that save button.

On-screen text

Local Ai MEGA UPGRADE

Transcript

We have just entered a future where you can literally glue super intelligence to anything for like $10. I know I sound completely insane. Hit that save button. Let's get it. This is ESP32-S3. It is a new microchip and it can run a 29 million parameter model completely offline. It has 512 kilobytes of fast SRAM and 16 megabytes of flash. I know, I know, I just gave you those numbers and you're like there's no way that AI is gonna fit on that chip Hatsu, there's no way. But wait, there's more. A developer just forced a 28.9 million parameter model on there and yes, that is about 100x larger than ever previously done. It generates coherent stories at 9.5 tokens per second and runs on as much electricity as an LED light. The breakthrough is actually architectural. Most of the models parameters actually live in a giant embedding table. What they did is they moved 25 million parameters into flash and then memory mapped it. The chip only pulls about 450 of those bytes per query and then all of the fast thinking stays on the SRAM. Fully offline. Zero cloud. Complete privacy.