Will AI Save Us From Civilizational Collapse?

Edward Dutton rough transcript, plus Tom Holland and others

Capitolene Temple

AI depends on intelligent programmers, smart people who keep the machines working, and a functioning electrical grid. Many countries have lost such a grid and rely on individual generators to get electricity. We have supply chain problems that we are already too stupid to solve. The number of near misses involving planes is going up as we apply DEI to air traffic control. Airlines have announced they will apply DEI to pilots, and medical schools want people with no interest or ability to be surgeons; anyone other than white, Indian, or Asian males. This is a kind of culturally generated stupidification. It is estimated that by the year 2100, at the current rate of decline, our average IQ will be 85 and we will be unable to sustain this level of industrial and technological development. This was the level of intelligence found in Europe in 1100 AD. The movie Idiocracy is too optimistic. Intelligence is needed to create machines, less intelligence is needed to maintain and fix them, but some intelligence is needed nonetheless, and a functioning infrastructure is required. The problem is caused partly by the smartest most educated women not having children, and only those on welfare and in regular trouble with the police having above replacement levels of children.

Dutton notes that the trouble with things like ChatGPT, artificial intelligence, is that they are trained to solve well-defined clear problems with regard to which they have a great deal of feedback data from which they can generate a model. This is a system where they are successful if they manage to follow the model and unsuccessful if they do not. They are trained to solve a very specific narrow problem. There may, in fact, be a large expansion of solutions to narrow problems and a collapse for the need for low-skilled jobs. [One might add, in moderately skilled jobs too, such as graphic design with the advent of DALI.] If people do not have work and what they regard as a sufficient standard of living this could cause problems and revolutions and lead to a general collapse.

Tom Holland, interviewed by Tyler Cowan, asked why the Greeks and Romans never had an industrial revolution. Holland noted a story about a man who invented unbreakable glass who proudly announced this to the Roman Emperor Tiberius, who, after testing that the glass was indeed unbreakable, promptly put the inventor to death, and buried the secret, for fear of putting glassmakers out of business. The story was probably apocryphal but illustrative of an attitude nonetheless. A steam engine was invented in Alexandria, but instead of being a harbinger of an industrial revolution, was only used to power temple gimmicks, like moving statues. Another story is told of the Emperor Vespasian where someone approached him while they were trying to rebuild the Capitoline Temple that had been destroyed by a fire in 60 AD, saying that he had developed a crane that is an excellent labor saving device. Vespasian refused to use it because it would put the common people of Rome out of work; again, so the story goes. The elites had an interest in keeping the lower orders busy. This strong luddite tendency emerged again at the beginning of the Industrial Revolution.

[Machine learning, like vision for self-driving cars, has the “answer” defined by human beings who live outside the machine. Real intelligence and real novel problem solving does not involve problems with known solutions. Builder bots can generate bots, and teacher bots can test them, with the student bots becoming closer and closer to approximating the human determined solution. The better bots are kept and the worse ones discarded according to human selected criteria. The bots get better, but the bot generators and the bot selectors do not know why or understand how the bots function. Bots cannot justify themselves. That is part of the problem of letting a black box determine things like whether someone should get parole or whatnot. Algorithms cannot justify themselves when asked why we should believe them.]

The current kind of artificial intelligence is not trained to solve novel problems because we cannot design such a thing. [I would say that this is related to the halting problem with puts the kibosh on fantasies of AGI.] With the pyramid of intelligence, at the base of the pyramid you have specialized abilities, very narrow specific things that very weakly correspond to intelligence, and above that you have verbal, mathematical and spatial intelligence, and above that you have general intelligence. AI cannot solve general problems, i.e., novel problems. For example, if we think about a bee, a bee has a very narrow set of skills indeed. A bee can gather pollen, make nectar, and build a hive, but the idea that a bee could become a godlike figure that takes over the world is very unlikely. A bee is not going to be able to solve complex general problems.

Another argument is that it does not take a great deal of intelligence to feed all the information from the history of humanity into these AI machines and from this they will be able to work out what to do and they will have all the knowledge that they have and will be able to develop a very high level of intelligence. Again, this notion treats AI as godlike. The gods will be fed scientific papers, and they will somehow convert this into knowledge and that’s it. But, it does not work like that. Computers have to be trained to do very specific tasks and to train them to do that requires a very high level of intelligence and we are in a situation where intelligence is in tremendous decline.

Another idea is related to Ray Kurzweil ideas, that we will have exponential growth. All we need to do is to get what we have now and scale it up with sufficient hardware, as though general intelligence is just lots of little intelligences combined. The trouble is that there is chasm between what we have now, machine learning, and artificial general intelligence. Computers can only be taught to solve narrow specific problems and have specialized abilities. And to the extent that they succeed, they have done it. It is a closed loop system. AGI would involve computers identifying what they perceive as problems and then solving those problems. It is an open loop system. Essentially, you have to get a tool to work out how to use itself, almost to have a sense of consciousness. You could argue that the problems we have to solve are essentially evolutionary ones and AI is not evolutionary and therefore it is going to find no purpose other than what it is told to solve. If it is told to solve evolutionary problems then perhaps it can go about solving them, that’s fine. But, again, that is not a matter of the computer thinking for itself. That is again the computer doing a very narrow thing that it is taught to do. You can ask ChatGPT right now what should we do about dysgenics, and it might really tell you, based on what other people have written on the topic, but that does not mean that we could persuade people who live in a highly individualistic society to do what it says. [For instance, the bottom third or so of society would have to stop having children and religious types are likely to be averse to embryo selection and genetic engineering.]

Will the Chinese save us? No. Says Dutton. They are already in dysgenic fertility. They are proto-woke. They are just a few decades behind the West. The unnatural colored haired group, the hallmark of the spiteful mutant, in Russia and China are few in number, but they are there, and they will continue to grow.

Previous related discussion from Determinists Strike Back, Part 4.

A.I. Is On The Wrong Path. You Can’t Get There From Here.

Gad Saad interviewed cognitive scientist Gary Marcus who argues that what is described as A.I. mostly involves “look up tables.” If the answer cannot be found there then, since the machine cannot reason, it is stuck. If, for instance, you saw a man carrying a stop sign and you had never seen that before, you would probably know how to respond to that novel situation, but a computer would not. Marcus comments that driverless cars are supposed to be on the horizon because, after forty or fifty year’s work, scientists can now get cars to stick to their lanes. “We’re almost there,” they say. But, Marcus comments, a ladder to the moon will never get you to the moon. You can get a little bit better driverless car, or a program for detecting sarcasm, so they can seem good in a directional way, but really, you are no closer to your actual destination. With regard to A.I., Marcus compares the situation to climbing K2 when you really want to climb Everest. The only way to get up Everest is to go back down K2 and start again. Since, as I will explain below, A.I. as it currently exists lacks any understanding of what it encounters, it will never get there. At the core of the problem, since intelligence cannot be reduced to algorithms, writing more algorithms will never get you to where you want to go, no matter how clever and sophisticated. Large language models, for instance, are “autocomplete on steroids” says Marcus. They are predicting the next words in sentences, but it is just not the right solution. K2 not Everest. You ask, “What do you like to do in your spare time?” And the AI says, “I like to spend time with friends and family.” It has just found those words in its database, but has no idea what “friends” or “family” means. If you ask it, it will just assemble some words from its database again. Since we do not know how a child learns language, how it connects, for instance, the concepts “go” and “went” and thus understands the concept of time, we cannot get a machine to do it.

Machine Learning is Not Philosophically Interesting – addition

This video, How Machines Learn, was sent to me by my computer programming son. Algorithms determine what YouTube videos will be suggested to you. In the old days, we gave bots instructions that humans could explain. But many things are just too big and complicated for humans to program. Out of all the financial transactions going on, which are fraudulent? Of the octillion videos that exist, which eight should the bot recommend? Also, related to the comment above, we do not know how people and even little children learn to distinguish a 3 from a bee. So, we build a bot that builds bots and one that teaches them. Builder bot assembles more or less at random. And teacher bot cannot tell a bee from a 3 either. If it could, we would just use teacher bot. The human gives teacher bot millions of pictures of 3s and bees and an answer key (look up tables) as to which is which. Teacher bot tests student bots. The good ones are put to one side. The bad ones discarded, as judged by the answer key. Builder bot is still not good at building bots, but now it takes the remaining bots and copies them while making changes and new combinations. The teacher has thousands of students and the test involves millions of questions. As the student bots improve, the grade needed to survive to the next round gets higher. Eventually, the bot is pretty good. But, neither the builder bot, teacher bot, the human nor the student bot itself knows how the bot is doing what it is doing. The student bot is only good at the types of questions it has been taught. It is great with photos, but cannot handle videos and is baffled if the photos are upside down. And that’s the problem. Humans can easily handle something upside down, and something it has never encountered along those lines, but where humans have not anticipated a particular scenario, the bot will likely not be able to make up for the deficient training by improvising. No complete set of rules can be written for driving situations, just as the workers at the front desk of a hotel cannot be given written instructions covering all eventualities. They need discretion to be able to do their jobs well. Concerning the bot, things that are obviously not bees, it is confident that they are. All the teacher can do is include more questions to include the ones that the student bot gets wrong. More data = longer tests = better bots. The human directs teacher bot how to score the test.