A Neo Lab for next-generation intelligence

PhAI Labs is an AI research lab working toward real scientific discovery. We do two things at once: train models, and put those models inside real research where their outputs get tested.

The team is built by people who can run the experiment and train the model: a science background at Fudan; AI research at Stanford, Oxford, Princeton and Peking University; first-author work in Cell, Nature and Science; open systems used across robotics, reasoning and multimodal learning. Individual profiles are on the Team page.

Intelligence has run out of internet. Published text is static and already trained on; synthetic reasoning only restates what a model already knows. The one input the frontier is short of is scientific data with a verified outcome attached. It cannot be scraped. It has to be produced.

Scientific discovery is the process that produces it. A hypothesis confirmed or rejected, an experiment that worked or failed: each is a signal the model has never seen. That is why we call scientific intelligence the third scaling law. A model that learns inside real research gains more than one fed the internet a second time.

So we are not content to make AI faster at existing research tasks. We want systems to take part in the fuller process of discovery, and to form new capabilities by doing so.

Model capability does not turn into discovery on its own. What is missing is the connection: turning the questions scientists care about into questions a system can research, turning the system’s judgement into tool calls and experiments that can be executed, and turning what the experiment returns into evidence that can test the judgement.

Our work is building those connections. Data trains the model; the model designs the next experiment; what comes back from the bench, including every failure, returns as reward, as a harder benchmark and as the next question. The two directions turn together, and each cycle lowers the cost of the next unit of valid data.

Scientists are not spectators in this loop. Domain judgement, data and verification conditions come from real laboratories, which is why we started the DFM Scientist Collaboration Program.

DFM (Discovery Foundation Models) is neither a single model nor a software framework that forces every system through one pipeline. It is the research paradigm, capability definition and system organisation that PhAI Labs proposes around Discovery Intelligence: through formal definitions, process diagrams, system instances and real scientific cases, it turns “discovery” from a vague notion into a capability of model systems that can be learned, executed and evaluated.

DFM is the overall paradigm of PhAI Labs' technical route. Around it we work in three independent directions: ScienceBuddy on scientists' feedback and a continually evolving agent workspace, ScienceIDE on real scientific code, tasks and environments, JEPA-Anything on cross-domain world-state prediction. ScienceBuddy, ScienceIDE and JEPA-Anything are three fully independent modules and research directions: independent in product form, usage, papers and implementation, with no required calls or dependencies between them.

Looking ahead, the three could form a composable research path: ScienceBuddy collects feedback and research trajectories from the interaction between scientists and agents; ScienceIDE organises that data and experience into reusable scientific environments for task execution, reinforcement learning and model training; JEPA-Anything explores unified state prediction and simulation across multiple scientific environments and interventions. That is a possible future combination, not an integration that exists today, and it does not change their independent standing. The capabilities DFM is about arrive progressively through these works; it is not a finished model.

Our long-term goal is Recursive Self-Improvement: intelligent systems that keep learning, verifying and improving through real tasks and environmental feedback. Science is the strictest test of that goal, because reality does not accommodate the model.

How we work is simple. Start with real questions. Verify through real feedback. Improve through collaboration. We do not overstate capabilities that are not yet released, and we do not describe plans as results. Data is used only within an authorized scope, and outcomes are determined by contribution and written agreement.

We want each exploration to become the starting point for the next discovery.