Curated. I think I had appreciated each individual point in this post before, but had not really tied them together and extrapolated.
Like everyone else I am fairly confused about the state of LLM cognition, and how it is likely to play out. I’d thought about “people are making mistakes anthropomophizing LLMs” and thought about “the content currently output from LLMs is influencing the next generation of LLMs” (while disagreeing with some that this is particularly relevant for whether we ultimately end up with particularly hostile/unaligned/deceptive AIs when it’ll matter most).
But the argument here is a bit subtler and focused on a different failure mode than I normally think about.
I think this post does a pretty good job threading a narrow line between “noting what AIs tend to say as a kind of evidence” while making sure to disclaim that this isn’t “the AI’s opinion”. I’m somewhat worried that many people reading the article will come away somewhat confused about exactly how that needle was thread, but, don’t think there’s much else Jan could have done about that (maybe putting the disclaimers before each AI bit rather than after?)
Curated. I think I had appreciated each individual point in this post before, but had not really tied them together and extrapolated.
Like everyone else I am fairly confused about the state of LLM cognition, and how it is likely to play out. I’d thought about “people are making mistakes anthropomophizing LLMs” and thought about “the content currently output from LLMs is influencing the next generation of LLMs” (while disagreeing with some that this is particularly relevant for whether we ultimately end up with particularly hostile/unaligned/deceptive AIs when it’ll matter most).
But the argument here is a bit subtler and focused on a different failure mode than I normally think about.
I think this post does a pretty good job threading a narrow line between “noting what AIs tend to say as a kind of evidence” while making sure to disclaim that this isn’t “the AI’s opinion”. I’m somewhat worried that many people reading the article will come away somewhat confused about exactly how that needle was thread, but, don’t think there’s much else Jan could have done about that (maybe putting the disclaimers before each AI bit rather than after?)