What happened
Researchers tested whether general-purpose AI models trained on many tabular datasets can estimate full probability distributions—not just single predictions—as well as custom-built tools designed for that specific task. The foundation models won on most datasets and dataset sizes, suggesting that you may not need to build separate specialized models anymore; one general model can handle the job across many different problems.
Why it matters
If general-purpose models can genuinely replace task-specific tools without loss of accuracy, organizations stop needing to hire specialists to build custom solutions for each new prediction problem—that's a significant shift in who does the work and how much it costs.