| > If consciousness exists on a gradient rather than as a binary state, then each architectural advance might add layers of cognitive sophistication that could eventually support conscious experience. It's 2025 and I'm frustrated that after decades of discussion we still can't get people to be clear about what they mean about consciousness. This article is all about cognitive capacities and behaviors and just assumes that these lead up to/are linked with conscious "experience". The Global Workspace Theory the author cites is about how we put attention on the most important stuff. Yes, one can make an analogy to how AI models today integrate information, but that's in part because Baars was making a cogsci analog to what 1980s AI models were already doing: > Bernard Baars derived inspiration for the theory as the cognitive analog of the blackboard system of early artificial intelligence system architectures, where independent programs shared information. But describing how we highlight information doesn't at all speak to why/how we have a qualia of that highlighted thing. Later in the wikipedia article, Baars' own "theater" metaphor is described, and you'll note it bears a striking resemblance to the "Cartesian Theater" as described by Dennett. This basically just shifts the qualia question: Roughly, who is watching the stage? If a rat can have qualia (and we use them to test depression meds) but not "recursive self-reflection", and a scheme interpreter can have "recursive self-reflection" but not conscious experience, then "consciousness" might not be a binary, but also isn't a "gradient" which implies you just have more or less of it. We have no clear signal from LLMs, no matter how sophisticated their responses, are _experiencing_ anything. I'm not taking a position on the consciousness of models; I think it's genuinely possible that a system of [tokenizing/embedding "perception"] -> [transformer-based generation] -> [recursive self-invocation] -> [actions/"tools" to interact with env], or something similar is potentially a really interesting tool for exploring cognition. But we shouldn't be using LLMs that have been trained on the speech / behaviors of already conscious beings. Consciousness arose in animals perhaps multiple times but not by copying pre-existing conscious creatures. Using language models specifically to examine this stuff muddies the water because we should absolutely expect them to produce text about an internal experience (we gave them examples like this!) whether or not that experience actually exists. |