Post #1490147

@pbloem@sigmoid.social @ngaylinn@tech.lgbt Two relevant points about this, the first one meta/rhetorical: The recent piece by Dan Kagan-Kans in an effective altruist community is pushing the narrative that "next token predictor" is the wrong way to think about LLMs, in an attempt to deflate Bender et al.'s "stochastic parrot" framing (as if the stochasticity were what's at issue, when it is not). So, this argument shape is being deployed in bad faith by certain actors. Caveat emptor"Risk" can be quantified. "Uncertainty" generally cannot. It makes sense to say there's a risk a coin will come up heads when you've bet on tails. It generally does not make sense to say there's a risk a chatbot will make a "bad decision" when that decision touches the real (large) world, unless you've so constrained what you mean by "decision" and "bad" that you're effectively not touching the real world anymore.