What's this about?
Wharton professor Ethan Mollick pushed back on the popular claim that LLMs only improve in "verifiable" areas like math and coding. He points to evidence that as models got better at those, they *also* got better at open-ended work — business cases, idea generation, medicine, law — and to a study testing AI on the MBA teaching cases used in business school.
What exactly is the "verifiable domains" argument?
Training methods like RLVR (reinforcement learning from verifiable rewards) work best where an answer can be automatically checked — a math proof is right or wrong, code passes tests or doesn't. The worry is that AI only improves where it can be graded, and stalls in "unbounded" fields (writing, strategy) where there's no clean answer key. Mollick's counter: the data shows gains spilling into those unbounded fields too.
Why is this worth noting?
This is the optimistic mirror image of OrangeAI's same-day complaint that RLVR is *overfitting* to coding — together they frame the week's central open question about whether current training methods generalize beyond checkable tasks. Mollick's "jaggedness" point adds nuance: the first AI-written bestseller may already exist, just not cleanly, because a human still patched the model's gaps.