It is an interesting paper. But even though it references Tononi's integrated information theory, I don't think it says anything about consciousness. It is just the name they gave to part of their model. They refer to a "consciousness vector" as the concatenation of vectors representing perceptions and short and long term memory, so really just a state machine vector. They show that their model, which also includes models of space and time, improves the task completion rate of robots tested in natural language using LLMs. It also shows just how far advanced China is in the AI race.
Any LLM that passes the Turing test is conscious as far as you can tell, as long as you assume that humans are conscious too. But this proves that there is nothing more to consciousness than text prediction. Good prediction requires a model of the world, which can be learned given enough text and computing power, but can also be sped up by hard coding some basic knowledge about how objects move, as the paper shows. If you are looking for answers to the mystery of phenomenal consciousness, you need to define it first. The test should be appropriate for humans, animals, and machines. Of course nobody does this (including the authors) because there isn't a test. We define consciousness as the difference between a human and a philosophical zombie. We define a zombie as exactly like a human in every observable way, except that it lacks consciousness. If you poke one, they will react like a human and say "ouch" even though they don't experience pain. But of course we are conscious, right? If I poke you in the eye, are you going to tell me it didn't hurt? Then what is it? What you actually have is a sensation of consciousness. It feels like something to think or recall memories or solve problems. Likewise, qualia is what perception feels like, and free will is what action feels like. These feelings are usually a net positive, which motivates us to not lose them by dying. This results in more offspring. Feelings have a physical explanation that we know how to encode in reinforcement learning algorithms. If you do X and that is followed by a positive (negative) signal, then you are more (less) likely to do X again. On Sat, Jun 15, 2024, 8:34 PM John Rose <[email protected]> wrote: > > For those of us pursuing consciousness-based AGI this is an interesting > paper that gets more practical... LLM agent based but still v. interesting: > > https://arxiv.org/abs/2403.20097 > > > I meant to say that this is an exceptionally well-written paper just > teeming with insightful research on this subject. It's definitely worth a > read. > > *Artificial General Intelligence List <https://agi.topicbox.com/latest>* > / AGI / see discussions <https://agi.topicbox.com/groups/agi> + > participants <https://agi.topicbox.com/groups/agi/members> + > delivery options <https://agi.topicbox.com/groups/agi/subscription> > Permalink > <https://agi.topicbox.com/groups/agi/T32a7a90b43c665ae-M5c35f67aa947a63004e35e44> > ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T32a7a90b43c665ae-M6b99887dcd5633d89566be07 Delivery options: https://agi.topicbox.com/groups/agi/subscription
