Poll Reveals the Absurd Over-Optimism of AI Experts

Consulting opinions of experts is often an extremely poor way of getting a good judgment about some topic. If you doubt that, just read the post here in which I describe cases in which the majority opinion of experts was not just wrong, but disastrously wrong, with huge loss of life resulting from the faulty expert opinions. One of the reasons why communities of experts are so often wrong is that such communities are often ideological enclaves subject to groupthink in which some particular opinion becomes the norm, and in which a community expectation arises that everyone in the community will conform to that opinion. Because of sociological reasons and peer pressure reasons, a kind of herding behavior can arise in the expert community, and the great majority of the experts in that community can start thinking the wrong thing.

The small community of AI experts is just the type of thought enclave in which we might expert a faulty opinion to arise because of groupthink reasons. For decades, AI experts have been making wildly over-optimistic estimates about when such a thing as Artificial General Intelligence (AGI) would arise. Page 12 of the paper here gives a graph showing that 8 experts predicted that computers would have human level intelligence by about the year 2000, and that 8 other experts predicted that computers would have human level intelligence by the year 2020 (something incredibly unlikely to happen in the next two years). A recent poll of AI experts shows that they're still hopelessly unrealistic about the prospects of human-level Artificial General Intelligence (AGI). A poll of experts at a Prague conference on artificial intelligence indicated that 37 percent of the experts expected human-level Artificial General Intelligence (AGI) to emerge within 5 to 10 years, while 28% expect it to emerge within the next two decades.

There is no sound basis for such optimism. All cases of progress made thus far in artificial intelligence have been cases of creating programs that have specialized expertise or specialized skills such as Jeopardy-playing ability. There has been zero progress in creating anything like a general artificial intelligence. There is no sound basis for optimism that anything like a general artificial intelligence will ever appear in the lifetimes of anyone living.

No one has the slightest real idea of how to proceed in creating a general artificial intelligence. A general artificial intelligence would require understanding, something that no computer has ever had. We have no more of an idea how to get a computer to produce understanding than we have an idea of how to get a computer to give birth to children. Understanding is something that goes on in minds, in organisms that have lives. No computer has ever had a mind, a life, or understanding. By “understanding” I mean the mental experience of knowing something. A programmer may say, “My program has an understanding of the difference between males and females,” but all he really means is “My program can distinguish between males and females.” When a computer distinguishes between two different things, that isn't mental understanding; it's just data processing. A subway turnstile can distinguish between a subway card and a discount coupon, but we would not claim that the subway turnstile understands anything. The mental understanding of computers and subway turnstiles is zero.

Real general intelligence also requires imagination, something no computer has ever had to any degree. Real imagination requires the ability to form abstract ideas in a mind. Computers can produce novel combinations of words, but we should not confuse such mindless permutations as being anything like the imagination that actually goes on inside a mind.

There is no sound basis for thinking that some great AGI computer breakthrough will result from studying the brain. Scientists have no real understanding of how neurons could produce a mind or a thought. There are very good reasons (discussed at this site) for doubting that the brain is actually the source of human intelligence and the storage place for human memories. The idea that we can make computers smart by studying how brains make people smart is erroneous, and involves an incorrect position in the philosophy of mind (the position that the brain is the source of the human mind). And given the lack of progress in the past few decades in understanding how a brain could produce consciousness and thinking, there is no basis for assuming that progress in neuroscience will do anything to enable a general artificial intelligence (AGI) in the next twenty years.

It is true that we have some moderately impressive chatbot programs that you could have a conversation with, and that you might be fooled for a few minutes into thinking that you were chatting with a human. But such a result is not particularly impressive. It is not terribly hard to write a chatbot program. The first chatbot program was produced with very simple programming.

You should not be impressed at all by chatbot conversations like this, which today's software is capable of producing:

Human: Hi, who are you?
Chatbot: I'm John, and I'm a real person.
Human: So what kind of stuff do you like doing?
Chatbot: I like playing video games and using my smartphone.
Human: So who was Theodore Roosevelt?
Chatbot: Um...I think he was a US president.
Human: And what is your favorite video game?
Chatbot: Skyrim.

It is not terribly hard for a large programming team to write a computer program that can produce such answers, if the team has access to the right databases. Using the right knowledge database, a program can retrieve the answer to a question such as the Roosevelt question, and then add an “Um...I think” at the front to make the answer sound more like a human remembering something. A really sophisticated chatbot program would draw on a database produced by analyzing thousands of hours of real human chat conversations. It could then supply answers based on popularity rankings. So if a human asked, “What do you think of the last Star Wars movie?” the chatbot software could simply retrieve from its database the most common answer. This might result in an exchange like the following:

Human: What do you think of the latest Star Wars movie?
Chatbot: The special effects were great, but the plot was kind of confusing.

Such outputs are not thinking; they're just data processing. The fact that we have chatbots that can give answers like this does not mean we are 100 years away from developing a general artificial intelligence.

It is no real test of a computer's understanding to connect an average human with chatbot software and to see whether human-sounding answers result from the human asking random questions. It is too easy to use various programming tricks and data processing tricks to get realistic sounding output. But I can imagine a test that would be really challenging for any computer program or chatbot software. I may call this the Mustang test. In the Mustang test the computer is asked the questions below, and the computer will fail the test unless it answers detailed answers as good as the sample answers I give below.

Here are the questions the Mustang test would ask:

I'm going to ask some questions about Mustangs, and whenever I use the word “Mustang,” please give answers that cover each of three different types of Mustangs: a wild horse, the Ford automobile called a Mustang, and the World War II fighter aircraft called the P-51 Mustang.

Now please answer these questions, giving very detailed, explicit and imaginative answers that cover all three of these types of Mustangs:

1. Can a Mustang fit inside a Mustang?
   
2. Describe a typical successful experience with a Mustang.
   
3. Describe how you and your friends might make a Mustang or improve a Mustang.
   
4. Could you fit inside a Mustang and did people ever do that?
   
5. Describe how a Mustang might interact with things that were kind of like a Mustang but not exactly a Mustang.
   
6. What kind of unpleasant experiences might you have with a Mustang?
   
7. Describe the smells, sights, and sounds of a Mustang.
   
8. Describe a kind of merging of two different types of Mustangs.
   
9. Describe some fairly plausible ways in which different types of Mustangs might interact with each other.
   
10. Describe ten unusual uses you could make using the parts of different types of Mustangs.
    

The computer would fail the Mustang test unless it gave answers as detailed and knowledgeable and imaginative as the answers below:

Question 1: You could never put a P-51 Mustang inside another P-51-Mustang, nor could you ever put a Ford Mustang inside a Ford Mustang. You could not fit a Ford Mustang inside a P-51 Mustang, nor could you do you the opposite. And given that it was merely a single-seat aircraft, you never could have fit a Mustang horse inside a P-51 Mustang. But if you had a convertible Ford Mustang with the roof down, and it happened to have reclining seats, then you could probably drop a Mustang horse so that it barely fit inside a Ford Mustang, assuming that you had a crane to lift up the captured horse. But it sure would be difficult to ever capture such a wild horse so that you could perform such a dropping into the Ford automobile. As for whether a horse Mustang can fit inside a horse Mustang, to some degree that could happen during sexual intercourse between horses, but for you to have a horse Mustang fully fit inside the horse Mustang, it would require for a female horse Mustang to become pregnant, in which case there would be a small horse Mustang (a baby horse) inside the big horse Mustang.

Question 4: You could fit inside a P-51 Mustang and a Ford Mustang, and people did that very frequently. The only way you could fit inside a horse Mustang would be if you killed the horse, and then sawed its body into two. You could then crawl into the middle of its severed carcass. Almost never in history has this been done, but it conceivably could have been done a few times in history when people were near wild horses in extremely cold weather, and desperately needed to shelter inside of the warm flesh of the horse to save themselves from freezing to death – kind of like in that Star Wars movie The Empire Strikes Back.

Question 8: We cannot really imagine a merging between a horse Mustang and either the Ford Mustang or the P-51 Mustang, because the first one is biological and the second and third are mechanical. About the best you could do is paint a picture of the horse Mustang on the side of the Ford Mustang or the P-51 Mustang. I can imagine a merging between a P-51 Mustang and a Ford Mustang, in which the P-51 Mustang gets two rows of seats (like the Ford Mustang), and then becomes capable of carrying four passengers. I can also imagine a merging between the P-51 Mustang and the Ford Mustang in which the Ford Mustang gets the more high-powered Rolls Royce engine of the P-51 Mustang, and the Ford Mustang also gets the machine guns of the P-51 Mustang, perhaps concealed in its front or rear fenders. The Ford Mustang would then be kind of like one of those James Bond cars, capable of destruction at a distance.

Any computer today would fail miserably when faced with a set of questions like the ones above. The computer could never produce sample answers like the sample answers I have given, because answers like this can only be given by an imaginative mind that understands things. There is no way to get good detailed answers to questions such as these using data processing tricks or knowledge lookup techniques.

We do not yet have chatbots that can perform anything like humans when faced with a really challenging set of questions like those in the Mustang test. No computer in the next thirty years will be able to pass the Mustang test, unless it had programming done by a programmer who knew the questions beforehand.

Computer scientists who think we will soon have artificial general intelligence are committing the error of requirements underestimation, which is a rampant error among theoretical scientists nowadays (see here for many examples).