CMU Assistant Professor Aran Nayebi argues J-space in LLM activations lacks the global recurrence required for consciousness theories
AI researcher Jack Lindsey countered that LLMs can emulate recurrence within forward passes
AI researcher Jack Lindsey countered that LLMs can emulate recurrence within forward passes
Source Posts on X
Cluster sources
Definitely agree that the J-space *dictionary* is the same, but what I meant was that what varies is the active sparse content, which is a bit different than the persistent workspace hypothesized by GWT (at least as I understand it). Moreover, the ignition and broadcast analogues in 4.1-4.3 are quite different in kind from what GWT posits involving competitive access, limited-capacity global availability, and global broadcast to many specialist systems. This is not to treat GWT as any sort of "ground truth" for sentience (I don't think it is!), but it's a bit of a far cry from saying there's a GWT in LLMs for those reasons, and I would expect a lot of these features to arise in basically any network architecture that solves a predictive task. In fact, I think your prior introspection work is a much stronger-backed claim in support of what people may intuitively map onto "sentience". But just my two cents 🙂
