The drift to civil war

Musk started SpaceX in 2002, 16 years ago and when he was 30. I have the distinct impression that he thought AI would happen 10 to 20 years later than it is, that he had another 10 to 20 years from now to make the rocket thing work before worrying about the Cyborg Question.

Publicity != frequency. Not even close. You only hear about the stuff people click on. Tesla has such a high profile that everything it does that can be reported will be reported. You read about one Autopilot-related crash in the news and think “wow that shit is dangerous”. You don’t read about all the times it worked.

And city officials are just as moronic as you, which is why when they read about an autonomous vehicle killing somebody they immediately move to make it illegal. Never mind that human-piloted vehicles kill people all the time and no one cares; autonomous vehicle deaths are Big News.

A moronic design decision, to be sure. But it will be fixed. And when it is fixed every autonomous vehicle will receive a nightly patch and none will ever make the same mistake ever again anywhere in the world.

Try that with the worldwide fleet of meatbags.

Again, the computers are *already not* making the catastrophic mistakes of humans. They don’t get distracted, drive drunk, fall asleep, or have aging vision and molassine reflexes. These problems are irreparable. A better emergency braking subroutine is not.

It’s shocking to me that, given the state of AI and how well people who know anything about AI understand how limited it is right now, the field test chosen was *CARS* lol

This said, we’re probably going to need some kind of alternative to human-driven cars before much longer because as the population becomes more impulsive, less intelligent and more irresponsible, the roads get more and more unsafe.

AI’s unlikely to fit the bill any time soon. Part of the problem is our inflated sense of our own intelligence which causes us to under-estimate the importance of moment to moment *sensation* as opposed to cognition. The other part of the problem’s the general decline in engineering and in particular programming. Each generation that passes, the code becomes more and more virtual, emulated and high level. That’s fine, except the distance between hardware reality and programming concepts closes off possibilities that used to be readily apparent.

Everyone knows a computer has a bunch of modes. Password entry mode, active phone mode, watching videos mode, presentation mode, fullscreen game mode, power saving mode. These modes have interactions. No one knows how they interact.

One man was put in charge of the boot process on Linux. He killed the longstanding nohup system for daemonization and created a binary log database because storage space and bandwidth are so scarce.

Desktop Linux has regressed terribly.

There wasn’t as much to get wrong in the old days, but people used to sometimes be able to get things right

In my first week in medical school 46 years ago, a speaker told us to go into surgery or radiology, because there was a computer program being developed in Pittsburgh called “Internist” which would make all the diagnoses and assign the therapies within five years.

These programs have proliferated but are all still garbage. It turns out the way expert clinicians make diagnoses is NOT correctly modeled as Bayesian iterative hypothesis testing.

I have believed, and continue to believe, that self-driving vehicles are an impossibility, and for similar reasons. Cognition remains a mystery, the machine analogs are crude approximations at best, and where human lives are at stake, humans will continue to drive (so to speak).

An illusion of consciousness is consciousness. If there is no consciousness there is nothing to be deluded. The idea of intentionality would never occur to p-zombies in the first place.

Good job getting gulled, though. Very obedient.

Jim you really would benefit from giving Dennett a chance. Yes superficially in the media he’s all about crude cheerleading for Darwin (which is fine) and crude ridiculing of religion (which is stupid), but his actual work is pretty much the thing you’ve been thinking about for some time.

Obviously the oeuvre can’t just be summed up in comments on here but one highly relevant point is this: if the brain’s made of neurons then there is no single, particular place within the brain where it ‘all comes together’. Since neurons operate quite slowly by computing standards, this also means there’s no particular ‘finish line’ in the brain before which something isn’t conscious and after which it is.
This has important consequences for the way we think about consciousness and the way we really should be thinking about consciousness.

“how do you program a car to know that it is about to run into something and to not want to run into it, leading to flexible and adaptive avoidance behavior that is not specifically programmed to deal only with specific situations specifically foreseen by the programmer?”

With the greatest respect, you’re glorifying consciousness beyond what it’s actually capable of. Human consciousness (the best example of consciousness we have, whether or not you want to take third party evidence as meaningful) doesn’t do much of that kind of stuff at all. When we’re driving, half the time we’re in auto-pilot and can barely even remember the thirty minute journey we just completed. Sure we do smart interesting things while driving a car that computer AI doesn’t do, but that’s mostly because computer AI isn’t actually *driving a car*: it’s doing a bunch of abstract, disconnected stuff that has ‘driving a car’ as a side effect.
That’s technically true of us as well, but our relationship with the car’s more direct because we’ve got a ton of sensory information coming in all the time and we care about the ramifications (some of the time anyway, for some of us lol) of a car headed toward us in the opposite direction.
How do you ‘program’ something to care (or ‘care’) about such things? We’re all about surviving to reproduce, which is why all the things you say about the profound effects of reproductive imperatives on behaviour are so adroit.

For a computer AI to do a good job of driving a car, it first needs to want to survive (optionally to reproduce). Simply having a constant in the software of “maintain your own existence” isn’t sufficient because that can easily be over-ridden (no pun intended and that’s also true of living things too, including us): that imperative needs to be reinforced (no pun intended) through repeated connections to EVERY ASPECT of the software: it needs to be completely committed to its own survival – if it stubs its toe, so to speak, its mind should immediately be flooded with concern about *that*, and if it thinks its supply of air (or oil, gas, whatever) is less secure than it should be, that should immediately become its working obsession.

Once the computer AI is utterly obsessed with its own survival, it can find ‘comfortable’ states and from those comfortable states it can take on other less important tasks like driving a car.

No computer engineer in his right mind is going to approach this problem from that perspective: they’re always going to prioritise driving first, so these cars are never going to be good drivers. They’re going to be, in a manner of speaking, “reckless drivers”.

“how do you program a car to know that it is about to run into something and to not want to run into it, leading to flexible and adaptive avoidance behavior that is not specifically programmed to deal only with specific situations specifically foreseen by the programmer”

It’s obvious that you don’t know a thing about machine learning.

You have a “consciousness” program with the ability to measure success, of which you the programmer set the definition, and you have a “subconsciousness” program that just does random shit. When the “subconsciousness” program accidentally does something that you want (and have successfully specified), the “consciousness” program rewards it with the programmatical version of dopamine.

Later, a “reproduction” program takes the most successful “subconsciousness” program, randomly varies its configuration, and runs another hundred trials.

There is no way for an artificial intelligence to do something not specified for by the programmer as there is no way for a biological intelligence to do something not specified for by the eonic struggle for survival on the planet Earf.

But those behaviors *were* specifically programmed into the genetic code of the tarantula. By billions of years of natural selection at every level of existence. It stands to reason that systems with vastly less evolutionary experience will be vastly less flexible.

Man is just now, in the last decade or so, begun to graft processes organic into silicon thoroughly mechanical. And having incredible success in a very short period time. No one is concerned that what currently passes for artificial intelligence will take over the world and rule as an inhuman god. They are concerned about the rate of improvement, which is astronomical. The silicon machines are accumulating evolutionary experience many orders of magnitude faster than biological machines ever did.

Self-driving cars would be possible for quite a while if governments were capable of making the necessary infrastructural investment. Governmentally, grafting computer vision into a car is a really nasty workaround for real rails between cities, virtual rails within cities, forcing upgrades from human- to computer-driven cars, the separation of two-ton hurtling missiles from pedestrian traffic, and the merciful euthanasia of suburbia in general.

Clearly it will not take a fully organic-competent AI to critically disrupt human civilization, human economy, and human society. It is already happening. The smartphone, powered by Facebook and Youtube algorithms, is a cultural catastrophe. Newly built factories are almost entirely un-peopled. The only desirable profession left is software-stuff, and that fact is not entirely, or even mostly, the result of governmental interference.

It is easy to identify the glaring holes in the overall fabric. It is much harder to notice the many portions that are already much stronger.

“Machine learning takes a general rule, and generates millions of extremely specific rules by throwing trial and error at billions of specific examples. Evolution has had billions of specific examples, but is not giving the tarantula millions of extremely specific rules.”

Granted, the state-of-the-art is still using search trees. But if I understand the AlphaGo paper well enough, they’re using search trees as the storage medium between the simulation and the performance, with neural networks doing the generation and generic search algorithms for the retrieval. Clearly, we are not yet in the realm of neuronal memory, but this is most certainly not my grandfather’s brute-forcing.

“No it will not be fixed. It was not a moronic design decision, but accommodation to the reality that computers just cannot make these kind of decisions.”

I tap keys on a slab-like board in ridiculously abstract ways to rearrange electrons running around in silicon manufactured to nanometer precision, for which function I earn several times the average first-world income. For a trivial fee, I can wirelessly connect to the Great CNS from nearly anywhere on the planet. If I wonder whether a zebra is white with black stripes or black with white stripes, I can reach into my pants, whip out my 5-inch wonder^H^H^H^H^H^H flat black obelisk, talk to it, and listen to it as it speaks the answer to my query aloud.

These things feel like magic… feel like the future… feel like intelligence.

On my half-inch thick, two-pound, magical clamshell gizmo exquisitely carved from a solid block of flawless aluminium, which probably contains many orders of magnitude more computing power than existed in the world when you were a child, I read that computer systems are beating humans at chess, are beating humans at trivia, are beating humans at Go. Meanwhile, they are now capable of walking around inside and outside, untethered, in autonomously exploring their environment; they are now capable of recognizing faces; they are now capable of picking things up and putting them down.

These are remarkable things. They are amazing achievements. Truly, we are living in extraordinary times. The ways in which humans are superior to computers are steadily eroding, from the top down: from the things that only humans can do to the things that any biological organism can do. And while you deny what is before your very eyes.

“AI programming throws ten billion random trials and errors at a hundred million examples, to generate ten million rules. Animals are not working off ten million pre programmed reflex reactions.”

That’s true— they’re working off ten million nonillion pre-programmed reflex reactions. Every bit of natural selection endured by every one of their ancestors is embodied in them at the cellular level.

Note that the rules are being generated by information-processing entities similarly undergoing ruthless natural selection for performance characteristics.

“There is something more to “knowing” than mere data. A chess playing computer has total data about the chessboard, but does not know it is playing chess. Knowing enables us to perform feats like proving that Euclid’s parallel postulate cannot be derived from the others, which mere data and mere algorithms could not do.“

What is this “knowing”? “I think, therefore I am”? In light of recent DeepMind developments, I think it is becoming more clear than ever that Descartes was not radically self-doubting enough. The chess playing computer has total data about the chessboard, but does not know it is playing chess; if we do not know what consciousness is, then how can we be conscious?

On the walk into my building at work the other morning the pleasant fragrance of Honeysuckles entered my consciousness. Now I’m quite sure we’ll be able to build an AI that can “spectrum analyze” the chemical signatures in the air and “recognize” the presence of honeysuckles, but will they find the “aroma” “pleasant”? To ask is to recognize the absurdity of the question. I’m also sure we’ll be able to program the AI to pretend they find it pleasant, but do they really feel anything at all?