I think this is a popular but ineffective argument. The authors argue (46 page PDF) that "Human cognition—in important instances—is better conceptualized as a form of theorizing rather than data processing, prediction, or even Bayesian updating." I don't exactly agree with this, but it is certainly the most widely-held view. They explain, "Theories provide a mechanism for identifying new data and evidence, a way of "intervening" in the world, experimenting, and problem solving." Theories are, according to the authors, the difference between recognizing critical new evidence, such as the verification of Kepler's elliptical orbits, rather than simply the majority vie. "The LLM can only represent and mirror the predominant and existing conceptions—in this case, support for the geocentric view of the universe—it finds in the statistical association of words in its training data." All this sounds great - until your LLMs start producing theories. Which, in my view, they can. Via Luiza Jarovsky.
Today: 1 Total: 120 [Direct link] [Share]
