AI for Science 公司要把模型能力变成可验证的科学成果，最关键的闭环是专有实验数据、湿实验室自动化，还是模型与实验的迭代速度？

“Robotic labs are autonomous experimental platforms that integrate with AI to perform high-throughput experiments, closing the loop between computational design and physical validation in scientific research (Ytb W2cBTVr8nxU). Integrating robotic labs with AI creates a feedback loop where AI proposes experiments, the robotic lab runs them, and the results inform new AI-driven hypotheses, enabling iterative scientific discovery at scale.”

“Lab-in-a-loop is an iterative research methodology described by Aviv Regev in which laboratory experiments are interleaved with AI modelling. Scientists perturb cells, feed the results into an algorithm that builds a predictive model, test the model’s predictions with new experiments, and refine the model continuously.”

“The wet-lab / dry-lab cycle describes the iterative feedback loop used by Recursion to continuously improve its biological maps and AI models. If unsuccessful – The failed experiment becomes new, valuable training data, especially if the model was wrong, because those data points lie in areas where the map is weak.”

“The wet lab–dry lab cycle is an iterative flywheel where AI-generated predictions (dry lab) guide automated biological experiments (wet lab), and the resulting empirical data retrain the AI, continuously improving model accuracy without prior hypotheses. Recursion’s scale—automated labs performing the equivalent of a PhD’s worth of experiments every 15 minutes—enables the rapid iteration essential for this cycle.”

“The automated scientific loop is the process by which AI agents iteratively perform the entire scientific method: generating hypotheses, selecting and executing experiments (in silico, wet-lab, or via CROs), analysing the data, and updating a World Model—a shared, distilled representation of domain knowledge that predicts outcomes. The aim is to shift the bottleneck from lab execution to scientific taste (see Scientific Taste Bottleneck)—the ability to judge which hypotheses are worth pursuing—a challenge that the platform addresses by letting downstream experimental results and human feedback drive decisions.”

“Experiment-in-the-loop AI is a training paradigm in which real-world physical experiments serve as the reward function for reinforcement learning (RL) agents. By making nature the RL environment, experiment-in-the-loop AI aims to overcome the deficiencies of simulators and the absence of high-quality negative-result data in existing scientific literature.”