I would want to better understand what is going on here, and what caused it. We are starting to see various new models have their chain of thought start to break down, and labs seem not that curious about how or why.
This video claims that OpenAI has optimized its coding agents for lower token count in chain of thought during coding tasks (and that the “thinking” displayed in UIs is the output of a second model summarizing the actual thinking). The illegible thinking sample shown looks a lot like something that would result from that sort of optimization pressure, either because Anthropic is directly doing a similar training procedure to get a similar compacted result, or because their training data is contaminated with reasoning traces from a dense-CoT model and it occasionally lands in that part of latent space.
This video claims that OpenAI has optimized its coding agents for lower token count in chain of thought during coding tasks (and that the “thinking” displayed in UIs is the output of a second model summarizing the actual thinking). The illegible thinking sample shown looks a lot like something that would result from that sort of optimization pressure, either because Anthropic is directly doing a similar training procedure to get a similar compacted result, or because their training data is contaminated with reasoning traces from a dense-CoT model and it occasionally lands in that part of latent space.
But it’s not really illegible, it’s just had all the spaces removed?