The "Wittgen=steine" talks continue Friday, June 26th, 3-4:30pm, in NIG lecture hall 3D: 

Thomas Raleigh (University of Luxemburg) will speak on:

Rule-Following and Machine Learning

26.06.2026, 15.00-16.30h, HS 3D

Abstract: Deep Neural Networks trained via machine learning can sometimes appear to be successfully performing some intended function but then surprisingly diverge from the intended function for some new inputs. This sort of misalignment bears a striking similarity to the sort of hypothetical scenario discussed by Wittgenstein – and later Kripke – in which a pupil appears to have successfully learned a rule, such as continually adding 2 each time, but then suddenly starts to continue the series in a divergent, deviant way. In this talk I will explore whether and how the 'Kripkensteinian' rule-following paradox is relevant for current AI technology. In particular I critically respond to recent claims by Arvan (2025) that the rule-following paradox allows for a 'proof' that it is impossible to interpret or align Large Language Models. If there is time at the end, I will also try to say something positive (and somewhat speculative!) about how Wittgenstein's rule-following remarks do bear on the topic of AI and machine learning.