An SLM trained on $8 ESP32-S3

An SLM trained on $8 ESP32-S3(github.com)

55 points by pavelai 10 hours ago | 17 comments

Would probably be better to demonstrate by exsmple how this approach is used to train on sensor data and then use it (as is hinted by the author) instead of acknowledging that the klingon poc is useless.

rbanffy 2 hours ago | |

But would it be as cool as a microcontroller that spits out Klingon?

dannyw 8 hours ago |

Very cool project! Sounds like it was a fun challenge :)

I wonder when we'll start seeing clusters of ESP32-S3s... not sure how interconnects would go though, but I guess the interconnect wouldn't be the bottleneck anyway.

ReactiveJelly 8 hours ago | |

Imagine a Beowulf cluster of those

fsniper 5 hours ago | |

Afaik there are already a few videos on YouTube.

alansaber 57 minutes ago |

Awesome.

andai 7 hours ago |

Very cool project.

> Backpropagation (gradients derived by hand)

What does by hand mean in this context?

Also how did you write the readme? It's a curious blend of human and AI writing.

porridgeraisin 5 hours ago | |

The fact that by hand is emphasized so often and so often (see src/handgpt.h comments as well above backward()) makes me think it's AI.

Anyways, it looks like gradients derived by hand means they didn't use autograd. They have written out the expression for dL/dW themselves.

wikisailor 3 hours ago | | |

Correct. No autograd: the expressions for the gradients are written out explicitly in C. There's a gradient check in tests/ that verifies them against centred finite differences on the published header, worst relative error 1.07e-08.

4gotunameagain 4 hours ago | | |

I think the readme is clearly LLM generated. The tone, the short sentences, the em dashes, lists, sections.. It all feels like LLM.

prplxd_nihilist 3 hours ago |

> Which is no small thing.

I think it is a small thing.

12373bd 8 hours ago |

hIngan motlh puS ruq tuq DujDaq SISwI' nge'vI' vIn SuvwI' yIvwI' qarghtaHvIS SIchoH

rbanffy 2 hours ago | |

vIparHa'bej.

pavelai 10 hours ago |

This small Klingon speaking language model was trained completely on ESP32. The training took 2 days

Number of parameters: 319K

(Disclaimer) The models is tiny and is not a pocket chatbot. It mistakes and is not capable to support conversation, but that's not the goal of the project

The goals of the project is to bring training to edge devices and it worked out

How can one use it? By using solar panels such device could be turned into autonomous meteorological station