Synthetic Sciences

AI systems have made rapid progress in mathematics because that domain has fast, trustworthy verification: a model plus a proof checker is enough. The empirical sciences lack this, creating two problems. First, experiments are slow, expensive and approximate, so models cannot brute-force the solution space the way they can in math; progress instead depends on experimental taste, the ability to judge which experiments are worth running and interpret their results. Second, even a model with taste needs compute, instruments, simulators and physical facilities, and almost none of that infrastructure is built for autonomous agents to operate.

Synthetic Sciences builds long-horizon evaluations and training infrastructure that test whether models can form hypotheses, run experiments, interpret failures and make progress under resource constraints, then works out how to train those capabilities at scale. The lab is starting with machine learning, where experiments run on compute and existing infrastructure can be used, before moving to physical sciences where agents need purpose-built labs to operate instruments (the first, called Lab 0, is coming online soon). Its named projects include FrontierRSI, which evaluates agents on autonomous ML research and development; OpenScience, an open-source AI workbench combining literature, code, experiments, compute and results; Atlas, infrastructure described as the record agents write to and the library they search; and Ascent, a platform for agents that plan experiments, run them across compute and report back.

AI systems are already producing breakthroughs in mathematics with new discoveries arriving in rapid succession, because math has fast verification; the lab argues the empirical sciences need analogous infrastructure and evaluations to reach similar breakthroughs.