Critic Slams Misuse of Regression Limits to Dismiss AI Capabilities

you see people outside of the field, both "pro" or "anti" AI people, deploying this to varying degrees. the folk conception of how it works basically doesn't support the idea of "probability" in a distributional context, where it's not like the prediction that a weatherman makes

hey hey, i have a stats background! look at me. i'm able to deploy the phrase "regression to the mean". why bother to pay attention to what the mathematical object i am talking about is actually doing, which would immediately reveal what i'm saying as a category error

like bro even if a model is giving you the "most common mode" of what it learned from data, thats still in fact NOT THE MEAN. the most common sample isn't even necessarily the most typical one if you're predicting names well at all, almost ALL of the probability mass isn't "John"

@kalomaze I do sometimes feel like I'm the tail wagging the dog when I explain to colleagues how I'm burning so many tokens, what a CLI agent setup really looks like and why I would pick and use different agents different models and different reasoning levels to solve a problem.