@mattsheffield@mastodon.social
2026-10-07 09:03 UTC
The other problem is that the independent research shows quite clearly that what people are calling "circuits" or "features" for a concept are not unique at all. There are many possible activations for particular outputs.
As one paper put it: All circuits lead to Rome: https://arxiv.org/html/2605.12671v1
Replies (1)
-
@mattsheffield@mastodon.social 2026-10-07 09:05
And even if the "models have a pain representation" argument were correct, having the ability to represent pain is not the same thing as a physical experience of pain which literally alters the physical bodies of biological entities. Chatbot internalities don't exist across answers. There is not evidence that chatbots have feelings or permanent selfhoods, but because they can write so well, this makes some people be easily misled about them. https://plus.flux.community/p/large-language-models-and-the-textual