The idea of artificial intelligence serving as a tool for scientific advancement has gained significant traction in recent years, but some experts argue the focus should shift from static models to dynamic agents. In a recent analysis, Eric Schmidt, the former CEO of Google and cofounder of Schmidt Sciences, and Suhas Mahesh, who leads AI for science work at the AI Center of Schmidt Sciences, suggest that AI agents—rather than purely data-driven tools like AlphaFold—may be better suited to accelerate research. They contend that while AlphaFold transformed protein structure prediction by learning from a massive dataset, it represents a specialized approach to a limited problem. In contrast, agents model the iterative, contingent nature of human discovery, potentially offering a more versatile framework for future breakthroughs.
AlphaFold’s success relied on a dataset of roughly 170,000 experimentally validated protein structures. This massive compilation took decades to assemble and cost billions of dollars to produce. The researchers note that comparable datasets are difficult or impossible to create in many other scientific fields. Because of this, a model that simply learns from a pre-existing database has limits. It cannot easily replicate the trial-and-error process that defines actual research.
AI agents, however, are designed as generalists. They do not represent a new scientific method but rather a digital simulation of how scientists work. By modeling the contingent process of discovery, these agents might be able to tackle problems where a massive historical dataset does not exist. The team argues that for science to advance, tools need to be able to handle uncertainty rather than just retrieve known facts.
This shift in perspective highlights a practical challenge for researchers. When a specific domain lacks the decades of historical data required to train a powerful static model, the path forward becomes less clear. The reliance on massive, curated datasets has defined the current wave of AI success, but it is not a universal solution. For fields where history is sparse, the ability to reason through problems step-by-step, rather than simply accessing a database, becomes critical.
Related: Reward Hacking Explained and Suspected Iranian Cyberattacks
From tools to agents
The transition from using AI as a static tool to using it as an agent involves a fundamental change in how researchers interact with the technology. Instead of feeding a model a specific question and getting a single answer, an agent can act autonomously, exploring different hypotheses and testing them in a simulated environment. This capability could be essential for tackling complex, multi-step scientific problems that cannot be reduced to a single query.
While tools like AlphaFold can provide powerful answers to specific questions, they do not necessarily teach us how to ask new questions. Agents, by modeling the human process of discovery, might offer a different kind of value. They could help scientists handle the messy, unpredictable nature of real-world research, providing a framework for exploring possibilities that have not yet been fully realized.
The researchers emphasize that the goal is not to replace human scientists but to augment their capabilities. By handling the iterative process of testing and refining hypotheses, AI agents could free up researchers to focus on the creative and strategic aspects of their work. This partnership between human intuition and machine reasoning could ultimately lead to faster and more efficient scientific discovery.
Current models often struggle to handle the ambiguity inherent in real-world scenarios, particularly when facing [situations where historical data is scarce](https://www.gmap-track.com/next-generation-ai.html). This limitation suggests a need for a more adaptive approach to problem-solving. The article points out that while large datasets have fueled much of the current progress in the field, they are not a panacea for every scientific inquiry.
