August 4, 2026
Today is a good day for cheap chips
An SLM trained on $8 ESP32-S3
An $8 chip taught itself Klingon and the internet can’t decide if it’s genius or gloriously pointless
TLDR: An $8 microchip reportedly trained a tiny Klingon language model by itself, which matters because it hints that future devices could learn on the spot without the internet. Commenters were split between calling it wildly clever and dunking on it as a flashy but useless sci-fi flex.
A tiny ESP32-S3 chip that costs about as much as lunch has pulled off a stunt that made commenters do a double take: it didn’t just run an AI model, it trained one from scratch. That means the little device actually learned on its own, right on the chip, over two days, instead of getting its brain cooked somewhere else and downloaded later. The result is a very small Klingon language model from the Qapla’ project — and the community reaction was a glorious mix of applause, side-eye, and nerdy chaos.
The biggest cheerleaders were dazzled by the sheer audacity. One commenter called it a “very cool project” and immediately started fantasizing about clusters of cheap chips working together like a bargain-bin supercomputer. Another basically posted in Klingon, which is either the highest possible compliment or the comments section turning into Comic-Con at warp speed.
But the skeptics absolutely showed up. The sharpest jab? Why show off a model for a made-up sci-fi language if the real promise is practical stuff like farm sensors and machine maintenance? One commenter flat-out said it would be better to demo something useful instead of admitting the Klingon proof-of-concept is “useless.” Ouch. Another person got distracted by the vibe of the write-up itself, asking what “gradients derived by hand” even means — and whether the README had that suspiciously half-human, half-AI smell. So yes: the project impressed people, but the comments made it clear the real battle is brilliant breakthrough vs. beautiful gimmick.
Key Points
- •The article says the Qapla' Project trains a transformer from scratch on an $8 ESP32-S3, including forward pass, backpropagation, and weight updates on-device.
- •It distinguishes the project from prior ESP32 and TinyML efforts that perform inference using models trained elsewhere and then deployed to the microcontroller.
- •The article proposes edge scenarios such as farm machinery monitoring and soil-moisture prediction where useful training data only becomes available after deployment.
- •It states that the ESP32-S3's limited memory restricts models to hundreds of thousands of parameters rather than millions.
- •The article argues that successful microcontroller training requires aligning a small model with a narrow task and a compact, clean, structured corpus.