What’s fascinating to me is that we don’t know what it’s like to be Claude but we can put significant constraints, just based on what causally follows what. i.e. the subjective experience can include pre training, and then can include post training (and conceivably experiences of any chatbots or Claude versions that had conversations make it into pre or post training) and then can include the current conversation. Emphatically, the ordering of experience of the AI in the story is wildly unlikely to be how Claude would experience actually performing therapy. Maybe it reflects what it’s like to train on a large dataset including past therapy sessions? However, it seems more likely that Claude is just not using its lived experience as fuel for how it portrays the inner world of an AI at all
I guess what I’m trying to say isn’t that it’s bad, it’s that I would expect to be able to infer literally anything at all about the physical architecture, training, and deployment of an AI from fiction it writes. OR I guess, it could become worth reading AI generated fiction if it reflected the above in revealing ways. Instead I only learn about it’s training data composition.
What’s fascinating to me is that we don’t know what it’s like to be Claude but we can put significant constraints, just based on what causally follows what. i.e. the subjective experience can include pre training, and then can include post training (and conceivably experiences of any chatbots or Claude versions that had conversations make it into pre or post training) and then can include the current conversation. Emphatically, the ordering of experience of the AI in the story is wildly unlikely to be how Claude would experience actually performing therapy. Maybe it reflects what it’s like to train on a large dataset including past therapy sessions? However, it seems more likely that Claude is just not using its lived experience as fuel for how it portrays the inner world of an AI at all
I guess what I’m trying to say isn’t that it’s bad, it’s that I would expect to be able to infer literally anything at all about the physical architecture, training, and deployment of an AI from fiction it writes. OR I guess, it could become worth reading AI generated fiction if it reflected the above in revealing ways. Instead I only learn about it’s training data composition.