The article critiques the notion that large language models (LLMs) can single-handedly solve complex scientific problems, like curing cancer or addressing energy issues. While LLMs have made significant strides, with 66% of physicians and 79% of law firms using AI, the author emphasizes that the challenges of scientific modeling extend beyond language processing. Dario Amodei's vision for AI encompasses not just the ability to analyze literature but also the need for models that can predict real-world outcomes. This distinction leads to two categories of AI: scientists (LLMs) that excel in reasoning and hypothesis generation, and simulators that learn directly from empirical data to model physical systems. The article highlights the importance of simulators, particularly in fields like biology, where complexity defies simple theoretical frameworks. Unlike physics, where principles can often be derived from first principles, biology's interactions are too intricate and variable. The author notes that while physics can be modeled up to millions of atoms, biological systems involve billions of molecules and countless interactions that can't be as easily generalized. The difficulty lies in biology's lack of a concise set of parameters for prediction, making LLMs insufficient for many tasks without the integration of specialized simulation tools and data infrastructures. Overall, the piece argues for a more nuanced understanding of how different types of AI can contribute to scientific advancement. It stresses the need for collaboration between LLMs and domain-specific models to build effective predictions and insights in fields that require deep knowledge of complex systems. This approach acknowledges the limitations of current AI while also recognizing its potential when properly integrated with empirical methods and advanced simulations.