You think I'm defining away the problem, and I think you're defining, uh, in? the problem. And that's the point. Whether there's a problem or not depends on how you define what you're talking about.

If you assume that intelligence, consciousness, whatever, is a non-physical phenomenon then, tautologically, there must be a non-physical process underlying it. No argument. This view has a very long history. One of the more modern players was Descartes with his hyperdimensional pineal portal. Penrose started from the same assumption and assigned the only (possibly) non-physical thing he could think of: wavefunction collapse.

As I said in my first reply, as far as I know Searle himself shies away from such assumptions. He talks about "causal powers" but doesn't speculate about what those might be or how they originate. The Chinese Room argument (which is just one of many in just his one paper) is meant to illustrate that what we think of as intelligence is non-deterministic. That's weak/regular/small-d deterministic since physicists have gone and coined super-deterministic too. If I give you the same exact answer to a question you would doubt my intelligence, as you doubt the Chinese translation book that always gives the same translation. You "tune-out" sometimes when you walk or drive a familiar path. You have "knee-jerk" reactions. You can act "robotically" doing familiar tasks. Sometimes you might "not be thinking." Etc.

Searle's argument is really quite practical. I think he thought that the AI researchers of the time were focusing too much on deterministic computer programs, databases and logic, when our impression of intelligence is very much not that. That's why it seems so strange to me that people would fault chatGPT for not being able to do arithmetic. Arithmetic is not AI, it's classical computation. AI is the ability to suck at arithmetic and make art, dream up hypotheses and say stupid shit.

You can add non-determinism in different ways. Throwing in some (pseudo)random numbers is one approach. That's what makes generative models generative: they either start from random numbers and transform them conditioned on input into the output, or they spice the IO transformation with randomness. They do not give the same output for an input. Nobody claims Google Translate (a deterministic deep learning system) is intelligent or conscious, but some people do claim the GPTs (non-deterministic generative deep learning systems) are.

OR, you can decide that intelligence is non-physical and so it needs to have at least some non-physical element. A concrete implementation of that would be True (with a big T) random numbers, not pseudo-random ones. If you believe that wavefunction collapse is Truly random then that could be a source, a la Penrose. That gives you True non-Determinism. Big D. No superdeterminism or any other kind.

You can also decide that intelligence isn't just deterministic physics plus randomness. That takes you to the world of pineal portals, souls and God.

Searle, I think, at least his Chinese room paper, was talking about brains having enough causal connection to the environment that they were non-deterministic by any of the above. Effectively non-deterministic due to fundamentally deterministic but massive influence by that environment or truly random influence from whatever. And how would you tell the difference anyway?

Untrue, they are poorly defined. If you think that they are well defined, tell me what you think they are and how you would test for them.

https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fen.wikipedia.org%2Fwiki%2F...
https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fen.wikipedia.org%2Fwiki%2F...
https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fwww.ncbi.nlm.nih.gov%2Fb...

There are lots of fuzzy definitions. There are also lots of concrete ones. Your choice depends on a) what you're doing and b) how much you want to believe in magic.

Ah, THAT's why documentation is so poor these days. None of the "software engineers" can type, and tab completion doesn't help when you're writing documentation.

Some of the earlier typing programs didn't emphasise proper shift or space key usage so lots of touch typists didn't learn those properly. I use my right thumb on the space bar much more than I should.

It didn't help that Microsoft kept screwing with the right shift and carriage return keys either.

adding an interpunction or non-letter between every word

I've occasionally met people who don't add a non-letter between every word. They'reveryannoying.

When I was little I used to beg my father to play text adventures with me because he could touch type. One day he told me I had to learn myself, complete with covering the keyboard, and about a week later it was fine. Helped out a lot with Turbo Pascal too. And as a bonus everybody at a place I used to work thought it was witchcraft that I'd use the keyboard without pulling the tray out. Of course, they also thought anything over 30 wpm was witchcraft.

IDEs seem to be mostly a crutch for people who (1) can't type fast enough so stopping everything and hitting tab, or hitting arrow, arrow, tab, is quicker than actually typing a whole word and (2) are so unfamiliar with the system they're working with that they need constant reminders about the names of functions.

Or until they fixed the artifact filters in their new radars.

They're lying about lying! Now some people will say, what if they're lying about lying about lying, but I know that's not the case. If anything they're lying about lying about lying about lying!

Those aren't really reactionless drives. They're pushing on the Earth. The Earth is the reaction mass.

It works fine. Windows 10 does too.

40% of desktop computers sold in 2015 had SSDs, and lots of those got upgraded too. But Windows 11 requires CPUs that didn't exist until 2017 and 2018, and weren't dominant until later still, so we're not really talking about "10 year old hardware."

That's a reasonable summary, although I'm not sure it's exactly what Searle himself meant. The weak point is pretty obvioius though:

2. Intelligence (by any reasonable understanding of it) involves understanding the deeper meaning, the context, the nature and qualia of the thing being considered. The semantics as Searle puts it.

"Deeper meaning" is a big red flag. If you start by assuming "deeper meaning" is something mystical, uncomputable, or non-physical then indeed, there is, by definition, no possiblity of "connecting it" to anything computable or physical. QED. This approach isn't very interesting though, and hasn't been very fruitful. God directs the planets in the sky as He wills. Go home and pray, there's no more to be done here. Ironically, such assumptions imply that these things are beyond understanding, "deeper" or otherwise.

If, on the other hand, you define intelligence in a usable way, you can start asking interesting questions.

Consider the method that poster mentioned but reverse it. Take an individual neuron and study its response to all possible inputs. This is practical to do for reasonable subsets of "all possible inputs", and in fact done all the time. There's randomness that makes the precise output unpredictable precisely, but nothing really surprising, nothing we can't pretty easily duplicate in a device or software. Now start connecting more than one of those neurons together. The IO function gets more and more complicated, but all we're doing is adding fairly simple, well-characterized units, either biological ones or synthetic. At some point, with the biological units at least, we all say "oh, now it's 'intelligent.'" That point differs between people, but if it gets big enough all of us agree it's intelligent.

This is perfectly straightforward if you define "intelligent" as an IO transfer function of sufficient complexity. If you do so, then there's a nice proof that an artificial neural network of at least two layers, sufficient number of units, and nonlinearity that can form a spanning set (all the common ones except identity work fine) can learn any transfer function.

There's no reason not to adopt this definition, and lots of reasons to do so, except that lots of people don't like it. There's no room left for magic. People had the same problem with Newton's mechanics and the "clockwork universe." People still bring up quantum mechanics to save them from it, but all QM adds, even in the most generous interpretation, is randomness. Which, by the way, ANNs love.

You can also go the other way. We can make devices that replace part of the brain and restore functionality. I don't believe any artificial hippocampuses have been implanted in humans yet, but there are a fair number of animal, including monkey, studies.

Then you can ask real questions. How complicated is the transfer function for various levels of intelligence? How can it most efficiently be learned? What configurations of basis functions are most efficient? One important result that's come out of questions like these, and the reason for all the current AI stuff, useful and stupid, is that computational depth can provide exponential gains in efficiency. Thus "deep learning."

What intelligence is, like consciousness, is poorly defined

It's not really. Both words have good definitions. People who claim there aren't any good definitions just don't like any of them because they aren't mysterious enough or make some things intelligent/conscious they don't like.

People would not generally consider someone working through it as intelligent, which would suggest that no machine could be what people would consider to be intelligent.

You are making an enormous leap there, the same one people assume Searle did (he didn't). You could also work through the input/output function of a single neuron by hand but most people wouldn't consider that intelligent. They also wouldn't consider the single neuron itself intelligent. But put enough of them together, with "enough" being part of that definition problem up above, and they would. Why not the same with your program?

The dissatisfaction you mention is the cognitive dissonance between everything we know about the world and our deeply held belief that we are special and possess souls, free will, magical brains, second substance, whatever.

Yes. Static but not stable is what I intended by "you've got to really, really want it."

You could also come up with more or less plausible stable configurations if the strength of dark energy were related to the total energy of the universe, or to the size of the universe with something less than the constant per unit volume the cosmological constant model assumes.

Lord of the Flies then. Or any internet community ever.

When people start calling it a "community" seems to be the first sign of the apocalypse. Right after that comes "for the good of the community" and then it's definitely time to move.