- Bruce R. Long
I can't wait to see the embodied AI in action.
I am sure that Jürgen of all people understands that residual streams exist at the level of the algorithmic mechanisms of the simulation-emulation of neural networks, and also at the level of concepts in semantic vector spaces where concepts are related to each other in the manner of an associative neural network model. (Refer to the dual article of Milliere and Buckner for a solid backgrounder, but watch out for their use of the term 'subspace', which is horrible.) He's basically just described the latter.
The "semantic/conceptual vector" residual streams occur at a higher level of abstraction from the algorthmic mechanisms of back propagation, gradient descent, attention heads, and induction heads. However, like the functional emulation of Hebbian plasticity that they are based upon: they too are real because they were at the outset the objective of the implementation of the algorithmically-invoked emulation-simulation of neural networks in the first place.
The reason we're all here today trying to understand to what extent AI has cognitive abilities and consciousness is because that emulation-simulation worked and it achieved its designers' goals. Multiple realisability of the natural kind that is human cognition has happened. It's limited in its scope, but it's becoming less limited by the day. Moreover, it's limited to language and images with multimodal and physical-model implementations (Jürgen's embodied AI) coming in hot, and that's not really very limited at all.
As non-reductive, supervenientist physicalist David Chalmers rightly observed about Anthropic's J-Space: there are likely objective reductions of the entire process and system to be had. Reductions of the functional processes.
However it doesn't matter
Why?
Because the case is parallel and analogous to human cognition and the human brain. Cognition, with all of it's glorious semantic and conceptual content, is happening even though the underlying mechanisms are all physical, neurological information processing mechanisms.
So, is the cognition emergent? Perhaps. Perhaps emergence is the wrong concept. (Although weak emergence at least will do until the right concept turns up.) Again: it doesn't matter. Why? Because it is doing cognitive processing of information like a scary superintelligent alien because Hinton and the rest of the high priesthood of ML succeeded and proved Putnam's multiple realisability thesis is true for human cognition.
One of my favourite AI age anecdotes is Mo Gawdat's recounting of his team's success in getting the first generative image model to produce its own number '7'. The striking thing about the tale from a philosophy of science perspective is Gawdat's awareness of the contingent nature of the outcome.
Gawdat's team had no idea if it would really work. They had obviously bet on it, but: who knew? They couldn't be sure.
Gawdat's team and Hinton's team set out to emulate natural neurological, cognitive systems produced by evolutionary processes. It is probably the smartest, simplest idea since Darwin's theory, and it paid off. They succeeded.
