In pursuit of natural harmony in AI.
Nature is the perfect algorithm. It makes the most efficient use of available hardware. On its offspring it grants the privilege of sensors and the responsibility of actuators. It operates in cycles. It chains one season's outputs into the next season's inputs.
Supermodel labs exists to bring AI systems in closer alignment with nature. We're doing this through systems projects and deep learning research. Today's AI systems operate in siloes. Model architectures are growing increasingly efficient but run on all of the same infrastructure. Hardware, meanwhile, is growing increasingly heterogenous but measured against a uniform yardstick. The world's most difficult tasks are inherently complex, layered, and best served from entirely distinct knowledge domains.
We're a lab focused on bringing natural harmony and AI research to its roots: hardware and software. We're from Apple and Together AI. We study nature as a source of principles for designing technical systems whose physical foundations and computational capabilities evolve together. Our mission is to translate those principles into an AI system that is more natural, efficient, and cohesive, building toward an advanced world informed by the intelligence of the natural one.
Projects
Aquaduck Aquaduck is a distributed inference project focused on bringing big models to little laptops. Through its hybrid AI architecture and approach to pipeline parallelism, it activates ordinary laptops and devices into a unified inference layer capable of advanced machine learning techniques like model chaining, model routing, and distributed compute, to maximize the utilization of the idle compute in latent hardware that enterprises and individuals already own.
Research
Pre-Generation Prediction Of Large Language Model Output Token Length Using Prompt Embeddings Accurately estimating the number of output tokens a large language model will generate prior to inference can improve cost estimation, scheduling, and resource allocation for downstream systems. In this work, we study the task of predicting output token length before generation using only the input prompt and its tokenized context.