2026-10-05 15:51 UTC
How can you "steal" reasoning? You could steal the product of reasoning, perhaps.
Imagine they made a chess model, and accused a rival of stealing its "chess reasoning" from transcripts of games played against it.
That, BTW, is exactly how LLMs play chess. They don't know where the pieces are, or what the rules are, and they don't plan moves ahead.
Based on the sequence of moves so far, and the huge corpus of games they're trained, they simply predict which move is most likely to come next.
Replies (2)
-
@jasongorman@mstdn.business 2026-10-05 15:51
And that, folks, is "multi-step reasoning" in LLMs. They're not trained to reason. They're trained *on* examples of reasoning.
-
@benh@mastodon.scot 2026-10-05 16:04
@jasongorman@mstdn.business something I have been wondering for a while: when OpenAI wants to train a new bigger model with more layers and more nodes; do they start from scratch with raw training data? Or do they follow this 'distilling' process of querying existing models.