@tesmaia@mastodon.gamedev.place
2026-10-05 20:26 UTC
@jasongorman@mstdn.business The entire concept of "reasoning" models was built on top of the discovery that by prompting the damn things to explain a "chain of thought", it gets some better results. What they're doing is fishing for context that then shift the probability distributions around. It's not deeper than "oh hey this works, lets do more of it and build the models around it".
Replies (1)
-
@tesmaia@mastodon.gamedev.place 2026-10-05 20:28
@jasongorman@mstdn.business Of course that means that the "reasoning steps" is genuine output that needs to be considered when distilling a model from prompt+output pairs, you can't expect to get similar results when skipping a big part of the output.