AI-Driven Scientific Discovery and Autonomous Agentic Systems

The community has undergone a significant transition in how models describe the world, shifting from simple learning models to those capable of fundamental discovery.
Evolutionary Stages of AI Models:
- Regression Models: Early stage technology utilized ten, twenty, or thirty years ago, focused on simple learning tasks.
- Generative AI: Developed in the last few years to sample existing landscapes, learn from data, and generate content within established distributions.
- Discovery AI: The current frontier where models generate new ideas, create their own training data, and engage in testing and validation processes.
Scientific discovery is compared to a jazz ensemble analogy: it involves "writing the musical score" while simultaneously solving the problem. Discovery requires proposing ideas that may seem "crazy," followed by rigorous verification and improvement steps.
Actionable Discovery AI involves several specific components:
- Autonomous AI agents.
- First-principle simulations (spinning off simulations within the system).
- Physical experimentation.
- Mining scientific literature to find "hidden papers" or ignored evidence within massive datasets.

A central theme in modern scientific AI is "breaking the world," which refers to changing the grammar of how the world is described and challenging established truths.
Traditionally, researchers were tasked with writing models for systems and materials (a practice established by experts at institutions like Caltech and MIT). Now, AI systems are writing, compiling, and running their own models.
The goal of AI in this context is to change the human researcher's mind about something previously believed to be incorrect or impossible.
Agentic Systems and Swarms:
- These systems focus on how a model interacts with other systems and the environment.
- They are inspired by biological emergence, such as cells self-assembling into ecosystems or civilizations.
- The focus is mapping complex, noisy, and messy experimental and biological data into real, executable, and compilable programs.
- Verification and falsification are essential components of these agentic architectures.

Agentic systems can incorporate physics into the retrieval and generation of new data.
Materials Applications: These systems have been applied specifically to energy applications, such as the discovery of new perovskite materials.
Sparks: A system specifically designed for open-ended discovery. Unlike traditional systems that target a known goal, Sparks explores unknown domains to create new scaling laws and principles.
Builder-Breaker World Model:
- This is an adversarial agent system found on the ArXiv repository.
- Builder Agent: Creates a world model to explain concepts as accurately as possible while penalizing complexity and maximizing coverage.
- Breaker Agent: Attempts to falsify or "break" the builder's world model by collecting new evidence from experiments and literature.
- MDL (Minimum Description Length): Used as a "gate" for simplicity to ensure the generation of elegant theoretical concepts.
Protein Science Application: This adaptive discovery process was used to predict the $\text{B-factor}$ in proteins, resulting in a simple, interpretable, and powerful theory.

Swarms: Decentralized collective agents that discover principles completely on their own without pre-described behaviors. These swarms are autonomous, evolving, and live globally rather than being confined to a single lab or computer.
Protein Design Comparison:
- Natural biology and standard evolution follow a specific distribution.
- Standard diffusion models sample from this existing natural distribution.
- Swarm models find "sparks" or "spikes" that fall outside the natural distribution, discovering entirely new protein design principles.
Science Claw + Infinite:
- Science Claw: Inspired by the "claw movement," it consists of hundreds of discovery agents with evolving skills and tools.
- Infinite: Modeled after the MIT Infinite Corridor, a place where interdisciplinary "collisions" occur (e.g., a mathematician, biologist, and philosopher meeting to generate research ideas). The agents inhabit this digital "Infinite Corridor" to solve problems.
Metamaterials Discovery: The system successfully identified a design space for resonators where no previous materials had been found, designing them autonomously and determining where research impact would be highest.

The vision for AI in science moves from AI as a "dictionary" or "library" to a "world-building machine."
Traceable Research: Unlike human scientific papers, which often only report the few experiments that actually worked (e.g., for journals like Nature), agentic systems offer complete traceability.
Every single experiment—including the hundreds that end in failure—is stored and recorded.
Future agents can build upon the exhaustive history of all prior results, both successful and unsuccessful.
The ultimate goal is for AI to write its own library of books and papers that challenge current belief systems.