Scientific Super Intelligence: From Science Tokens to ScienceRSI

Abstract

Scientific reasoning moves beyond natural language. The search for a discovery runs across sequences, structures, fields, omics, images and measurements, while language models reason mainly in text. This Perspective sets out four proposals from the Scinetics team. First, we define science tokens: modality-appropriate units of scientific information, spanning native discrete symbols, typed values, continuous representations and learned, decodable codes. A shared interface for these representations would let one model read, write and reason across modalities, so that a chain of science tokens operates on scientifically meaningful states, not on descriptions of them. Second, we define Scientific Super Intelligence (SSI) as a joint crossover: one frozen model that outperforms vertical specialists, such as the AlphaFold series on protein structure prediction, on their own tasks, and outperforms general large language models of the same size on broad scientific reasoning. Capacity, published training precedents, data inventories and compute place the first test of that crossover at 100–300 billion parameters and 10–30 trillion training tokens. Third, we propose ScienceRSI (Science Recursive Self- Improvement) as the route to that window. Following the AI4AI pattern, each larger generation is trained on data that the smaller generation before it produced: the earlier model generates candidate reasoning, screens it and keeps what passes, and in science that screening can be done by a simulation, a database or an experiment. Training this way reduces the dependence on scarce human-written scientific reasoning data and shortens each iteration. Fourth, we propose science-token economics, which separates metered computation from the value of evidence-labelled, reusable scientific assets.