Despite the century-long study of star-forming regions, we still do not have a clear observational picture of how star clusters actually assemble:
Is there a switch between hierarchical and monolithic formation? 
When do young stars decouple from their parent gas clouds? 
Current formation theories rely heavily on simulations, which often offer contrasting predictions. But with new data from multi-epoch surveys, we can finally push the boundaries of observational science and address the questions from a data-driven viewpoint.
I will briefly introduce these open questions and why they matter, and then discuss how multi-epoch NIR surveys like VISIONS, combined with new machine-learning tools, are letting us measure the motions of deeply embedded young stars and even the ISM itself. I will show what these measurements are telling us about cluster formation and how this will improve in the era of JWST, Roman, and Vera Rubin.

Artificial Intelligence (AI) has changed our way of assessing, processing, and distributing information, and Large Language Models (LLMs) like Chat GPT, Gemini, Claude and others could be seen as a replacement of intellectual performance previously only available through human intelligence. In the scientific community, many of us have been using AI and LLMs to accelerate their scientific productivity. However, it is unclear whether LLMs are actually needed to do high-quality science or simply yield higher productivity without substantially extending knowledge? This informal discussion was triggered by a discussion I had with Jason at the ESO guest house in Chile in January. I would like to discuss whether science actually profits from LLMs or if these tools rather undermine novel, quality research by design.