PhAI Labs
PhAI Labs is developing Discovery Foundation Models (DFM), a research paradigm that defines discovery intelligence as a learnable, executable capability for AI systems. The company's work spans three independent research directions: ScienceBuddy, an interactive AI research partner; ScienceIDE, which converts real scientific code into executable environments; and JEPA-Anything, a cross-domain world model for predicting intervention outcomes. JEPA-Anything reduced single-step intervention prediction error on Interventional Pong by 34.8%.
- Artificial Intelligence
- AI Agents
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
AI systems today handle many research-adjacent tasks—retrieving literature, generating code, and processing data—but they do not participate in genuine discovery. Public text data is exhausted for model training, while the most valuable signal—scientific data with a verified outcome (a hypothesis confirmed or rejected, an experiment that worked or failed)—cannot be scraped or synthesized; it must be produced through real research.
Solution
DFM is a research paradigm proposed by PhAI Labs that defines Discovery Intelligence as a learnable, executable, and evaluable capability of model systems. It formalizes seven capabilities spanning the full discovery loop: identifying valuable unknowns, converting them into researchable questions, building and revising problem representations and hypotheses, calling scientific tools and experiments, validating judgments against external evidence, and depositing reusable research structures. DFM is not a single model or software framework but a paradigm instantiated through three independent research directions: ScienceBuddy (an interactive AI research partner that learns from scientist feedback), ScienceIDE (an executable environment for real scientific code, tasks, and validation), and JEPA-Anything (a cross-domain world model for predicting state changes and intervention outcomes).
Target Audience
Primary audience includes AI researchers and machine learning scientists working on foundation models, reinforcement learning, agents, and scientific intelligence, as well as domain scientists across biology, physics, chemistry, and clinical research who need AI systems capable of participating in real discovery processes.
Features
- Formal capability definitions for seven discovery abilities, from identifying valuable unknowns to depositing reusable research structures
- ScienceBuddy: interactive AI research assistant that records scientist feedback, tool calls, and research trajectories to enable Recursive-in-Recursive Self-Improvement
- ScienceIDE: converts real scientific code, workflows, and validation criteria into runnable, measurable, reproducible environments; trained PhAI-IDE models (4B, 9B, 72B) show improvements on held-out code repair tasks and general benchmarks
- JEPA-Anything: orthogonal predictive factor decomposition (OPF) enabling unified latent world models across seven domains—vision, biology, clinical trajectories, control, molecular dynamics, physics fields, and meteorology—with reported improvements on all 10 dynamics tasks and a 34.8% reduction in single-step intervention prediction error on Interventional Pong
- DiscoveryRSI: research trees with hypothesis credit optimization, harness replay, and rediscovery mechanisms for recursive self-improvement; DiscoveryRSI-27B achieves the highest rate (6.3%) of hypotheses with concrete connections to post-cutoff published research among compared models
- RSI-Anything: world-model-guided candidate selection and adaptation for recursive self-improvement pipelines, raising code repair success from 26.3% with a static verifier to 31.2% after world-model adaptation
- DFM Scientist Collaboration Program connecting the paradigm to real laboratories for domain judgment, data, and verification conditions