Never intelligent?

I sometimes see folks, usually not experts but occasionally, take the position that machines/software such at ChatGPT-4 is not “intelligent”, and that it never will be, never could be.

Sometimes this refers to the specific method. Usually it refers to machines in general, in theory, in all possible or all soon to be realistically achievable configurations.

I haven’t seen a good argument for this that I understand. Maybe it’s out there and someone can help me find it.

To be clear, I am referring specifically and only to “intelligence”.

So for example I replied to one person:

1. I don’t think we’ve ever discussed consciousness.

2. I have not commented on “human intelligence”, just “intelligence”, and instantiating it in silicon isn’t a duplication, it just is intelligence, is the view.

3. The view about intelligence is not predicated on replication human biology, so it doesn’t depend on whether or not we have tools for studying neurology at sufficient detail. We still don’t know a lot about that but we have intelligence in machines already, and in some weak form have had it for a long time. It will get better.

Whether one is an eliminitivist about consciousness and qualia or not, and if those things are indeed real and epiphenomenal what makes them come to pass, or if they are not epiphenomenal then are they causal in a way that current machines can’t capture, or would take a lot of knowledge to figure out how to capture and we are far from that, are all interesting questions but distinct and for the most part independent of the issue of machine intelligence simpliciter. It only comes into play if it’s not epiphenomenal and is somehow actually or pragmatically essential for full-blown AGI.

I think the evidence so far is that it’s not, but I grant that this isn’t a settled matter.

As to whether it’s necessary for achieving any intelligence at all, I think the best way to understand the benchmark is operationally, and we have met that condition already. Therefore, the idea or criticism that it might be via a glorified encyclopaedia isn’t a problem in my mind (uhem), because it’s operationally indistinguishable for most practical purposes, and if you think intelligence is a computable function then any finite implementation is equivalent to a lookup table.