AI for science needs reasoning, not just data
The article argues that while AlphaFold's success raised expectations for AI in science, the data requirements for replicating such breakthroughs are unrealistic for most fields, and AI agents—reasoning systems with tool access—offer a more practical and universal approach to accelerating scientific discovery. AI agents can automate the iterative, uncertainty-driven process of actual research, improve reproducibility, preserve institutional knowledge, and dramatically reduce experimentation costs across all scientific disciplines.
Source: MIT Technology Review