AI hardware research lab targeting the energy wall in AI compute. Building a new silicon substrate that runs neural networks directly on analog physics rather than digital abstractions, targeting biology-scale efficiency. Has open-sourced Un-0 model weights, training scripts, and ablation code for community research into physics-based models.
DS-ISA is the first instruction set architecture for analog dynamical system hardware, offering a 9-instruction model that could unlock compilers for physics-based AI chips targeting 1000x energy efficiency.