(Non-essential section is at the bottom.)
I use a general template (my own though.) to answer. I've mostly used 'thinking' and so on; note that there is a slight difference that you may need to adjust for:
Minsky doesn't take inner experience very seriously, it's just a bunch of physical processes for him. Both "Society of Mind" and "The Emotion Machine" show exactly what Minsky expects, I mention this at the end of the first section. I personally like better the latter.
Searle does take it as an emergent phenomenon with a physical basis, and because this isn't publicly observable (for the most part.), he requires a specific mechanism in place; not just simulation, but a graph-like similarity, as defined and described below. You can check a related paper of his here.
Let's assume that brains and computers do the same I/O computations; the difference being substrate (hardware vs wetware.) but also the graph dynamics (let me explain.), but the "conclusions" and behaviours are very similar.
By graph dynamics I mean that we are not just replacing nodes (neurons), edges (axons), environment (chemicals, inter-synaptic space, distance) and its evolution in time (dynamics) but that those don't exist even as analogous entities when you look at the hardware.
This raises a few questions:
- Can computers simulate the graph distributed computations, and hence think? (here you can split strong AI vs weak AI as defined in the post.)
- Can hardware-computers be like brains, using the right "graph"? (this is different.)
- Can only brains (with neurons, proteins, chemicals, etc.) be like brains ?
Let's ignore the last question (3.), and assume that the substrate isn't important.
[On a note, I suspect Minsky didn't always think that 1. and 2. were equivalent. He may have come to that conclusion later.]
The computations that occur in everyday computers are carried out in electronic circuits that do not resemble a brain i.e they are not wired in the same way, nor compute in an analogous manner, as detailed before.
If the system is different in terms of the graph dynamics, but has analogous I/O (language, movement, overt behaviour.) could it be thinking the same?
The same computer, programmed correctly, can simulate weather, but just as Searle would say, it's does not rain within. This may have been Searle's first idea, but it isn't good. What we are looking for, from the beginning, is the overt behaviour, the I/O part.
We could also simulate the weather by replacing each molecule with a person (or a grain of sand), then following initial conditions and physics laws. It feels somewhat pathetic. It wouldn't rain either, but you may end up throwing a lot of sand around -just as clouds do with water. Does this count as the same weather phenomenon?
Again, when we simulate weather, we are interested in predicting; simulating the laws is enough for that.
When we simulate more complex entities like humans, we are interested in a system that does seems human when interacting, to a certain extent.
Later on Searle probably reviewed his ideas on consciousness, accepted that we are machines (consciousness may be emergent but not inherent to wetware.).
Now, if we make a machine that does digestion using different mechanisms, does it do digestion? What counts?
What counts is the transformations of the food. So for this to be simulated, just like with hearts.
With thinking, the inputs can be readable words, but it's likely that for Searle to buy in, one needs not only a similar kind of overt behaviour (output), but also similar kinds of processes. For that, we need to know what happens in the brain, which we are far from.
Searle would reply 1. No, 2. Yes.
For late-Minsky, I think, it's not so important (albeit desirable) whether the simulation is processing information in an analogous way to what the brain does. But he does think that we need high level theories of how the brain could do it, then computers would think.
Complexity and Flexibility
Two aspects are less relevant but have some importance given the context (80-90s):
- Flexibility: older AI was a very poor and fixed program. It would be specifically doing a computation to solve a problem. For example, it wouldn't deal with noise well at all.
Brains are flexible in comparison: they learn, they deal with noise much better, they are much more interconnected and less deterministic on apparently the same output.
Neural Nets were developing fast at the time, but probably was not known to many.
This isn't an argument for impossibility but just to give an impression of the field. Also, Neural Nets need more data than brains to learn, and it's unclear whether they have strong similarities, apart from some parts -say, of vision and convnets.
- Complexity: brains are far more complex than a program in the 80-90s, the overall state affects processing in such a way that it's non-deterministic; although FFNN-like networks existed since the 30-40s, they were not the main paradigm at all. I'm not saying these are complex enough, but they do deal better with noisy data.
The main aspect is the third one. It involves real vs simulated; real understood as the physics and substrate of a real system, simulated by one of a different substrate (to an extent) but that simulates the physics and also higher level theories. Described below.
