Can machines think?

Is a question Richard Hamming poses in his book The Art of Doing Science and Engineering. Here is my shoddy answer.

Trivial Solution

The ones made by the Thinking Machines Corporation could.

Nontrivial Nonsolution

A Rough Sketch

We can either assume a physical-only universe, or a partly-physical universe. There are many variations on both categories, of course.

If we assume a partly-physical universe (which is not just trivially non-physical), then something beyond our physical brains is (at least in part) involved with our thinking. The non-physical can either be amenable to our manipulation, or not. If it is, it may be possible to “breathe consciousness” into a machine, or it may not.

If we assume a physical-only universe, then we may either be capable of making machines that can think (in terms of our intellectual capacities and our discipline and drive), or we may not. Silicone may or may not be conducive to thoughts. Computers with their current architectures may or may not be conducive to thoughts.

Probabilities in a partly-physical universe:

Suppose 30% probability of the non-physical being amenable to our manipulation (at a fine enough level to be useful for our purposes).
Suppose 60% probability of being possible to “breathe consciousness” into a machine (many mythologies imbue inanimate objects with life).

Probabilities in a physical-only universe:

Suppose 25% probability of humans being capable of making machines that can think.
Suppose 85% probability of silicone being conducive to thoughts (I have a hard time seeing why our biological brains should be materially more conducive to thoughts than silicone).
Suppose 35% probability of current architectures being conducive to thoughts.

Probabilities of each universe:

Suppose 5% probability of a partly-physical universe.
Suppose 95% probability of a physical-only universe.

Then total probability of seeing thinking computers is:
5% * 30% * 60% + 95% * 25% * 85% * 35% = 8%.
And thinking machines in general, if we ignore the architectural issues of current computers or presuppose an escape from it:
5% * 30% * 60% + 95% * 25% * 85% = 21%.
So, let’s arbitrarily say 15% probability. That’s decent enough to motivate someone with a strong preference for a universe with thinking machines in it.

Observations

I don’t think machines can think already. Thinking is a process, not an outcome: hence existent hacks to accomplish tasks are probably dead-ends, and existent approaches that wrestle with how to think rather than the products of thinking have some chance of success. A reflex exhibits zero thought.

On the other hand, many who claim it’s inherently impossible for machines to think appeal to some mysticism. I don’t think this appeal makes sense as a direct prohibition: it’s possible to envision universes with a non-material component to thought, where it’s nonetheless possible to imbue machines with the same non-material components. For example, if a god has made us in his image, why should we not be able to make things in our image? A pretty bad performance by the god if creation of sentience cuts out at one level – essentially a hacky solution.

Thinking is also a matter of degree. Unless we subscribe to Last Thursdayism, there have been several species of humans. What was thinking like for Homo Erectus? What thoughts might a modern human think if transported back 200K years as a baby to be raised in the stone age, that the stone age humans could not? We’d certainly like to think the modern human would outwit the others by the 200K year margin. Though, perhaps the Victorians outwitted us. What is a dolphin’s thoughts like? A chimpanzee’s? If they pass the mirror test, does that indeed mean they possess a sense of self? Does that mean they have sentience? Some birds have passed the mirror test – but bird brain’s have different architecture from our mammal brains. What differences does that architecture make? If machines can think, can they think thoughts we can’t, due to the architectural difference alone? Does silicone and carbon facilitate different thought-spaces?

If current computer architectures turn out to be resistant to real AI, how likely are we to really investigate other alternatives? We do seem to love our technological Stockholm syndromes.

Conclusion

I think it’s plausible that it’s theoretically possible for machines to think, though I don’t think we have thinking machines yet (e.g. “machine learning” doesn’t cut it). I find the usual excuse for why machines should be unable to think – appeal to religion/mysticism – to be weak, because it fails to address 1) why machines can’t be imbued with the non-physical thought-stuff 2) the thoughts of other animals. Hence, even if we exist in a partly-physical universe, I don’t think that is reason enough to prohibit thinking machines. In both scenarios (purely-physical/partly-physical universes), the question of our human capacities are central, though – maybe we’re too dumb, or lack the mystical powers, to make thinking machines. My rough estimate says 15%, though, which is pretty good chances for something that’d be this dramatic.

Anyhow, here’s some of Hamming’s students’ thoughts on the matter: