Beyond Tool Use: Multimodal Distillation and the Evolution of Neural Networks toward AI Scientists

AI scientists are frequently regarded as a transition from passive instruments to autonomous research entities. This research contends that the fundamental transformation is mechanistic: neural networks have progressed from task-specific predictors to multimodal distillation systems adept at integrating scientific information, compressing expert knowledge, and facilitating discovery workflows. Scientific AI is transitioning along the spectrum from AlphaFold-style predictions to agentic systems that design experiments, utilize tools, and compose research results. We offer a conceptual framework that connects three stages: predictive neural networks, foundation models, and multimodal distillation-driven scientific agents. We contend that the autonomy of AI scientists should be assessed not solely by the quality of their outputs, but also by their capabilities in hypothesis formulation, experimental design, verification, attribution, and human oversight.