And we got there. I've got a 200M param 1 layer looped transformer with a 2048 vocab, and 32 token context window. It uses up pretty much all of the chip so this is about as big as I could make it. The model is trained on starhopp3r/TinyChat. It streams over uart so ~10tps.

Show Post

These models when told to operate new motor systems will ultimately write Python, JavaScript, or whatever language scripts and tools they need to go from tiny servo movements to full-on hand, arm, and limb orchestration and movements and routines such as grab({item...}) for later agents. That's RSI

Show Post

State of the Art (SOTA) models such as Claude Sonnet and Opus 5.5 as well as GPT 6 Astra and even 6.1 Terra are all capable of reasoning with kinematic motion up to the six or seventh degree in physical systems either within video game engines or real life to just describe said math in JavaScript

Show Post

In 2012 I worked at the Johns Hopkins Applied Physics lab on the Modular Prosthetic Limb (MPL) where we used human-driven systems like an arm sensor-servo cage for arm movement and tactile feedback. This is, ultimately, just a series of degrees of freedom in a mathematical space in 3D

Show Post

Similar to Boston Dynamics, a robot that can wrap up a package, or a robot that can block a sword swung at it, both operate on a set of mostly rigid routines that have rough, spatially-aware and defined hyperparameters (size of box, vector of sword approach). Architectures can work with that!

Show Post