How a programming language from 1958 can accelerate self-improvement.
The world is quickly moving towards recursive self-improvement (RSI). The promise is an intelligence explosion, where self-improving general systems become better at improving themselves. Classic literature calls this the singularity, yet reality is relatively slow-motion when contrasted with sci-fi. Model training takes time, and autonomous explorative ML research is not instant either. Improvements will have some level of diminishing returns. Even then, I expect a technology leap bigger than any before within the next ten years.
What if those ten years could be squeezed into one? That is the motivation behind rethinking the self-improvement system through Lisp. Lisp is the programming language of early AI research, designed in 1958. We built S-expGPU, a Lisp framework for agents doing ML research. Lisp fell out of favor because of performance challenges. Hardware and compilers have caught up, and right now our S-expGPU is 1.03× to 2.4× faster than PyTorch in every benchmark we have run.
Speed comes from building only what the next iteration needs, and Lisp keeps that minimal. Instead of months of model training, mere days are enough. Instead of hours-long ML experiments, minutes suffice. S-expGPU is built to minimize the needed scope when agents implement ML experiments. This leads to both cheaper and higher-quality experiments. Because the gains compound with each iteration, the payoff can be exponential.
041 research is focused on hyperbolic geometry which is not a total coincidence. Lisp syntax is a tree, and hyperbolic geometry is the principled approach for trees. The building blocks are exotic, but the goal is better general intelligence. We just have a shortcut.