but I do wonder about the confidence with which you can totally dismiss the notion
For the current tech, 100%.
These are static systems. They don’t update themselves while running. If nothing else, a system of consciousness has to be dynamic. Also, the way these models are trained is unlikely to produce consciousness even if it theoretically could.
Assuming that they are seems like a leap, but since we don’t really know exactly what consciousness is,
We don’t technically have a definition for what it is, but we have some criteria. Consciousness is an emergent property. So theoretically a system could become conscious unintentionally if it is complex enough. But again, it requires a system to be dynamic, to be able to change and grow on it’s own.
Nerual nets are just trained on data. LLMs specifically are trained on the structure of language, which is the only reason they work as much as they do. We can’t train meaning or understanding, but being able to churn out something resembling information is a byproduct of training language because language is used to communicate information.
The issue that a lot of people have is they assume that something is intelligent/sentient if it can produce language, which is what we have seen in nature, but while it takes intelligence and maybe sentience to create/develop nothing says that intelligence or sentience is required to “use” language.
LLMs do one thing: Produce the next word for a given context. It does not matter how big we make it or what the underlying complexity is. The models just produce a word. The software running the model adds the word to the context and executes a new loop with the most recent context. It runs until it hits a terminating token that the current output is “finished”.
Even for the models that are considered the “thinking”/“reasoning” models just have additional context tokens for the “thinking” section that basically force the model to generate more context which, thanks to the way language is constructed, can constrain the output, but it’s only ever outputting the next word.
Which is exactly my point. A biological brain, human or otherwise, is incredibly efficient for what it does. It’s also effectively infinitely parallel which is impossible to do with the current tech.
In order to even attempt or approach a system that could be remotely considered “conscious” we would need something that is way more efficient just because of logistics. What they are trying to do with the current hardware has basically reached the practical maximum of scalability.
Hardware footprint and power are massive constraints. The current data centers can’t even run at full capacity because the power grid cannot supply enough power to, and what they are using is driving energy costs up for everyone. On top of that, a bio brain is way more dense. We would need absurd orders of magnitude more hardware to come close with the current tech.
And then there is the software. Nerual nets are a dumbed down model of how brains work, but it is very simplified. Part of that simplification are static weights. The models do not update themselves during execution because they would very quickly muck up the weights from training and basically produce nonsense. They don’t have feedback mechanisms. We train them on one thing. That’s it.
In the case of LLMs, they are trained on the structure of language. We can’t train meaning because that requires unimaginable orders of magnitude more complexity to even attempt.
If AGI or artificial sentience is possible it will never be done with the current tech. I would argue the bubble has likely set AI research back decades because of how short sighted and hamfisted companies are pushing it has soured public perception.