the thing that slightly worries me is economic incentives to focus hard on highest value AI tasks -> models that are extremely efficient at 120iq human labor, at the expense of more esoteric content w higher ceiling on quality... ubiquitous midwit AI 😞
yep, this is why intelligence is *fundamentally* contextual and evals are such a huge part of AI- if the task is understanding issues D students are having w the material then yeah A student papers... imo we will see more tech for spawning child models for potent low cost performance on frequent tasks