Recent research highlights the effectiveness of diffusion models in sampling biomolecules by leveraging equilibrium molecular distributions, enabling both direct sampling and the derivation of forces acting on molecular systems. These models have demonstrated significant utility in molecular dynamics simulations, as they facilitate a more accurate representation of molecular behavior. However, despite their strengths, there remain inconsistencies between the energy-based interpretation of the learned scores and the training distribution used in these models. This discrepancy suggests that while diffusion models are powerful tools, further refinement is necessary to fully reconcile the theoretical energy framework with practical training outcomes. The ongoing investigation into these inconsistencies is crucial for improving the reliability and accuracy of molecular simulations using diffusion models. Overall, the current evidence supports the view that diffusion models are effective for biomolecular sampling, yet challenges persist in aligning their energy-based interpretations with empirical data.