Artificial-intelligence systems such as large language models (LLMs) can exhibit behaviour that seems strikingly intelligent. This has led many people, including some experts, to wonder whether these systems could have some form of consciousness — the capacity for subjective experience. However, we think that existing AI systems are not conscious, and that, along current trajectories of development, they are not likely to become conscious. Greater clarity about whether AI could become consciousness will be crucial as society grapples with the challenges posed by this increasingly powerful technology.
The other blocker is that it can’t actually learn anything, what it gets from the training is what it’s stuck with, it has to hold anything new in the context window.
Both of these are pretty insurmountable problems for the architecture.
Was Henry Molaison no longer possessing consciousness after the surgery?
How about Clive Wearing after the herpes caused anterograde amnesia?
There’s been a number of humans that due to injury can no longer form new memories or learn anything. According to your definition, do those humans lack consciousness?
Those are very obviously surmountable. I could make a simple model architecture on my laptop right now that updates its weights in real time and is always responding to input and giving outputs. I don’t know if it would be a useful model, and there are good reasons not to do that, but it can easily be technically achieved. Those are not good measurements of consciousness, and if you truly believed they were accurate, I could make a “provably” conscious machine in like a day of work on consumer hardware.
It’s rather surmountable, considering the context window is measured in the MILLIONS (soon to be billions) of tokens, and each memory(note) takes in the neighborhood of fifty. I’m pretty sure that has long since eclipsed the capacity of your own memory…
Additionally, there’s zero reason a model could not have dynamic weights, to learn in real time. It’s just not enabled on consumer-facing models, to provide a consistent, reliable experience. It’s not a limitation of the technology.
The problem is context is just not the same thing as changing the structure of the neural net, it doesn’t matter how big you make it
You’re correct, it’s not - it’s a band aid over that reliable experience they’re aiming for. Yet it doesn’t have to be that way. The weights CAN be updated dynamically with each new input. They don’t do that because then they will have less control over what it can do, and they can be held legally responsible for what it can do… But it is still possible to update with every new input. And happening behind closed doors.