Researchers from the Berlin Institute for the Foundations of Learning and Data (BIFOLD) and Google DeepMind have demonstrated that artificial intelligence models for molecular simulations can achieve state-of-the-art results without explicitly built-in physical laws. The study, titled "How simple can you go? An off-the-shelf transformer approach to molecular dynamics," challenges the long-held assumption that machine learning models for molecular dynamics must encode fundamental principles like energy conservation and rotational symmetry directly into their architectures. Traditionally, models rely on these inductive biases to ensure reliable, physically meaningful predictions. In contrast, the team developed a novel model called MD-ET, which utilizes a standard edge transformer architecture with minimal adaptation. This approach deliberately strips away constraints regarding equivariance and energy conservation, instead relying on the model to learn physical behaviors purely from vast amounts of data. The research was published in The Journal of Chemical Physics by authors including Max Eissler, Tim Korjakow, Stefan Ganscha, Oliver T. Unke, Klaus-Robert Müller, and Stefan Gugler. To train MD-ET, the researchers utilized approximately 30 million molecular structures from the QCML database. After pretraining on this massive dataset, the model undergoes fine-tuning on specific target systems. The results indicate that MD-ET successfully learns to predict forces that are approximately equivariant, with deviations falling many orders of magnitude below typical force magnitudes. In NVT simulations, where temperature and particle count are held constant, the model achieves stable performance even in few-shot settings, demonstrating competitive or superior accuracy against established benchmarks. However, the study also reveals limitations when applied to NVE simulations, which require fixed energy with no thermostat. In these scenarios, energy conservation is only approximately learned and proves sensitive to molecular size and numerical perturbations. The researchers observed that for larger structures, the model can exhibit runaway energy increases, a significant drawback of the unconstrained approach. Consequently, the authors advise that simulations using nonconservative forces must be carefully validated on a case-by-case basis. The findings contribute significantly to the ongoing debate regarding the necessity of physical inductive biases in AI-driven chemistry. MD-ET suggests that sufficiently expressive general-purpose architectures trained on large datasets can approximate physical laws without explicit hard-coding. While this simplifies model design and optimization, it also clarifies the current boundaries of the technology. For many practical applications, the study shows that relying on data-driven learning is viable, though users must remain aware of potential stability issues in energy-conserving contexts. This work opens new avenues for developing flexible molecular simulation tools while highlighting the need for rigorous testing when physical constraints are not enforced by the model itself.