This quote from Steve Mandel makes me think about why frontier LLMs are still terrible stock pickers (once you give them a clean sandbox adjusting for lookahead and contamination) even as they conquer frontier mathematics. And are likely to continue to be terrible stock pickers. (I believe this will be particularly true around points of regime change and "n of 1" events in markets...the times when the best human investors drive real separation and retail/consensus/momentum players get washed out). Frontier LLMs are deterministic optimizers trained on verifiable-reward RL to find one checkable, binary answer. And they are getting incredibly good at it (see: Navier-Stokes). But RL-tuned LLMs fundamentally suffer from premature convergence, like a junior analyst who develops "conviction" after 3 hours of work... (if you have managed over-confident junior analysts, you know this can be quite frustrating) The fundamental challenge, per Steve Mandel framing below, is that stock picking has no verifiable reward at the decision point. A good thesis can lose money and a bad thesis can make money (though the market gods tend to hand investors the former far more often than the latter...). And that over-confident junior analyst has a way of getting lucky the first time or two (i don't make the rules). Frontier LLMs have been optimized to collapse uncertainty and generate a definitive, coherent final answer to maximize confidence metrics. This is fundamental to the system, and bigger training runs just make the system better at this job. But fundamental investing is ultimately a game of making predictions where investors must embrace & price uncertainty, think probabilistically, and update those views dispassionately as the evidence changes. This doesn't take frontier-math capabilities, in fact it takes almost no math at all, but it takes the right way of thinking about distributional asymmetry. Lately have become fascinated with this idea of pricing distributional asymmetry, more to come on that (and love to share notes with any AI-pilled people working on this problem).