Google's TabFM skips per-dataset training and still predicts on unseen tables, matching tuned baselines and cutting pipeline ...
Is there really such a thing as ‘clutchness’?