Post Ar7VCesGvXa1AOhvCC by sqrtminusone@mastodon.bsd.cafe
 (DIR) More posts by sqrtminusone@mastodon.bsd.cafe
 (DIR) Post #Ar7U2UfTv2yaf8JX72 by interfluidity@zirk.us
       2025-02-14T20:03:41Z
       
       0 likes, 0 repeats
       
       claude.ai just asked me to explain something. "Would you help me understand where I went wrong in my thinking about the Secondary Income credit entry?"so i did, at least i explained how i think about it.claude.ai "worked through" my explanation, then responded "Thank you for this clarification - it's a much more logical way to approach the problem!"will claude.ai in any sense "remember" my explanation? have i meaningfully taught it anything?
       
 (DIR) Post #Ar7VCesGvXa1AOhvCC by sqrtminusone@mastodon.bsd.cafe
       2025-02-14T20:16:41Z
       
       0 likes, 0 repeats
       
       @interfluidity (not sure if that's a serious question)If you've allowed Anthropic to train on your chats, they might do so (I think they've made it opt-in by default). But even in that case, the stars have to align very precisely for it to reproduce your answer from a single training document.Otherwise, the way is to add the required explanations to the context for future chats where it might be needed.
       
 (DIR) Post #Ar7VL7xOgMcpgK8vZY by benjohn@todon.nl
       2025-02-14T20:18:12Z
       
       0 likes, 0 repeats
       
       @interfluidity I’ll boost, because I don’t know. I find it unlikely as it would need to ultimately become training data in future to adjust its weights and biases. It could become part of a training corpus though?Or it might do some kind of online learning - I doubt that. Or it might tuck it away in to some persistent context?
       
 (DIR) Post #Ar7YmOMO6UteEhWaCO by faassen@fosstodon.org
       2025-02-14T20:56:47Z
       
       0 likes, 0 repeats
       
       @interfluidityCertainly not directly. This is just verbiage it picked up in its training to be good at chatting. LLMs don't remember their input, except in the sense that it is the input of this particular session, along with the rest of the interaction.For it to "remember"  they need to make this interaction part of the training data of either the next version of claude or some fine tuning. Not sure whether they do that and what's in terms of conditions. Might have a minor effect at best.