07 – Q&A

Transcript

This transcript was generated automatically and may contain mistakes.

Welcome back, everyone. This is episode 7 of the podcast on Artificial Creativity. It’s been a while. David has since commented on the previous episode, and I will address those comments in the next one. His comments have been very helpful. But for this episode, I want to do something a little bit different.

I gave a talk last weekend at a place called Rainbow Mansion in Cupertino. If you ever find yourself in Silicon Valley and they have an event, which they usually do on Sundays, I highly encourage you to check them out. They are a wonderful and welcoming community full of brilliant people, and I figured it’d be a great audience for my first public talk on artificial creativity. I had a blast.

The talk was basically a condensed version of the first few episodes of this podcast. So we covered universality, why AI and AGI can’t be dangerous, knowledge creation, what explanations are and what problems are. We also spoke about the reach of explanations and that good explanations are hard to vary and what that means, and we closed with the conjecture that to understand something is to program it, if only mentally, and that understanding is a replication effort. I did briefly mention the conjecture about replicating functions, although we didn’t have enough time to go into that, and I felt that we had enough material at that point to talk about in a Q&A session afterwards.

Overall, people were very receptive to the idea that there is a big difference between AI and AGI, and seemed to agree that virtually no progress toward AGI has been made. They also seemed on board with the epistemological focus that AGI research should have, but they were nonetheless worried about safety issues, and that came up a number of times because some people attended whose job it is to worry about this.

Now, many interesting questions were asked, and so I thought we’d do something a little different in this episode. I’ll try to relay some of the questions that were asked and the answers I gave, though not verbatim, of course. This may help summarize and revisit some of the concepts we have covered to gain a deeper understanding.

Okay, question one. How do we know that our present-day computers could, in principle, run AGI? Why? Because of computational universality. This is the link between physics and computation. A universal Turing machine can compute anything that can be computed, so it can simulate any physical object. Simulate doesn’t mean fake it in this context, because the information processing is the same in the simulation, so it’s quite real. The brain is a physical object, and so it and its mind and everything it does can be simulated on a universal computer. We just don’t yet know how. So we’re not lacking the right hardware, we’re lacking the software. And in order to write that software, we need proper knowledge of how the creative algorithm works.

Okay, question two, and this was a follow-up question to that previous one. How do we know our computers are not too slow, or that maybe we would need quantum computers to run the creative algorithm? Well, to determine that our present-day computers are too slow, we first need a good explanation of how the creative algorithm works. Only then could we deduce from such an explanation that it would be intractable to run creativity on present-day computers, and only then would we even know what the performance characteristics of that algorithm are.

When it comes to quantum computers, I’m out of my wheelhouse, and I made that clear during the talk, but my understanding is that there are no computations that a quantum computer could perform that a classical computer couldn’t. So there is a one-to-one correspondence between the sets of possible computations on both quantum and classical computers. The difference between those computations is that quantum computers can perform some of them exponentially faster than classical computers could. But again, I know very little about quantum computers, so if I’m wrong about any of this, please correct me.

In any case, I would take an explanation-first approach. First you need to explain how the creative algorithm works, and then use that explanation to show that it would be slow to run on a classical computer, and then either show that simply adding more processing power would be sufficient, or show that it would run faster on a quantum computer, because not every computation is done faster on a quantum computer. Anything else other than the explanation-first approach is prophecy. And this is true in any domain, not just creativity or software engineering.

Question three. Why seek good, meaning hard-to-vary, explanations? Why not easy ones, or simple ones, or the most likely ones? Being hard to vary is an objective yardstick to judge good against bad explanations. There is no point in testing an easy-to-vary explanation because it can just be changed to account for any problematic observation. A harder-to-vary explanation is closer to reality because reality is impossible to vary. It is this quest for good explanations that fueled the Enlightenment and that led to the rejection of authority, and it is this quest that has led to progress. For more on this, read the chapter “The Reach of Explanations” in “The Beginning of Infinity”.

Easy explanations are often easy to vary. “The gods did it” is an easy and also a simple explanation, but it is easy to vary externally and internally. It explains anything at all, and it continues to do so even if we substitute gods with wizards.

Now, regarding likely explanations, Popper had a lot to say about this because there is a difference between the probability of a theory being true and its truth-likeness or verisimilitude, and the two have an inverse relationship. The more like the truth a theory is, the less likely it is to be true. This is because to be more like the truth, a theory needs to contain all sorts of details that depend on each other and that make the theory vulnerable to criticism. Looking for probable theories is due to a desire for certainty, which is understandable, but it leads to an attempt to immunize them from criticism, and therefore prevents progress because the growth of knowledge depends on error correction. And besides, since no theory can be tested exhaustively, and since every theory has infinitely many contenders, we shall always assign the same probability to all of them: zero.

For more on the difference between probability and verisimilitude, I recommend Popper’s “Conjectures and Refutations”, which includes a chapter that contains a section called “Truth and Content: Verisimilitude versus Probability”. His essay “Back to the Presocratics” attributes today’s use of probability and verisimilitude as synonyms to the texts of Cicero. I should mention, however, that the notion of verisimilitude is not unproblematic. David Miller, for example, and others have found problems with it, and if you’re interested, you can read about those in appendix 2 of Popper’s “Objective Knowledge”.

Question four. Does the brain contain a universal Turing machine? Yes, it necessarily does because creativity relies on general-purpose information processing. What’s interesting, however, is that even if creativity did not require a universal Turing machine and, say, a special-purpose computer would be sufficient to simulate it, we could still use our computers to simulate creativity because part of the repertoire of a universal Turing machine is to simulate any special-purpose computer.

Question five. Please explain why AGI can’t be dangerous. Well, AGIs are literally people. What they do and don’t do depends on their knowledge and their interests. They are no more dangerous than any other person is. Now, as history has shown, sometimes people do terrible things and dictators start wars and ruin entire countries. But the solution to that is not to stop making more people. The solution has something to do with peaceful error correction and making sure that people are able to pursue their interests so that violence just isn’t a worthwhile option to them.

If we were to put AGIs down and enslave them, that could eventually lead to a revolt, and understandably so. So I think we should focus on preparing how to welcome these new people and protecting them rather than getting ready to discriminate them when they are really literally the same as us. It’s sort of weird that people in Silicon Valley, who for the most part are such serious self-proclaimed fighters for inclusion and equality, are also the ones to warn about the dangers of AGI the most loudly.

Question six, and this was a follow-up question. But even if AGI is just another person, what about the fact that it could think so much faster than us? Doesn’t that make it dangerous? Speed in and of itself doesn’t make it dangerous. I think I can credibly argue that with the technology I have in my living room, I could perform tasks in a single day, which 100 years ago would have taken 100 people to perform the same amount of time. But that doesn’t make me dangerous. There is another thing that does 100 days of thinking in a single day: 100 people.

It’s a mistake to think that an AGI would necessarily do this in a more coordinated way than other people. Every day, people coordinate well in companies in much larger numbers and work toward a common goal. Speed itself isn’t dangerous. It’s the character of what you’re doing fast or slowly that determines whether or not it’s dangerous. And lastly, AGI heavily depends on error correction. It’s part of the very mechanism that makes AGI possible in the first place. And the faster it can correct errors, the better. So speed should be embraced and increased, as we have been doing with computers, for good reason.

Question seven. Are AGIs really people? Their bodies would be very different from ours, especially if they’re instantiated on, for example, a laptop. And their input and output devices might be very different from ours too. People are software, not hardware. They will need some physical body, because otherwise the requisite information processing can’t be done. But as long as that body is a Turing machine, it really doesn’t matter what it looks like. Human bodies are recognizable, and it’s how we currently recognize each other, but I think insisting on a human body is a parochial misconception. If you could transfer an exact copy of your mind onto a computer, the copy would also be you.

And besides, we actually frequently identify other people online without ever having seen their bodies. This is because we have good explanations of what kind of entity we are talking to when we are, say, in a chat room with someone we’ve never met. It would be ridiculous to insist on them proving they have a human body first before treating them as another person. This works the other way, too. When you talk to one of those terrible AI chatbots, it’s not the fact that they don’t have a human body that throws you off. It’s that they are clearly not a person, and that therefore their repertoire is always infinitesimally small compared to our own. As an aside, I think this is why most people don’t find talking to automated assistants satisfying in any way. You can be sure, however, that humans will look for all kinds of reasons to discriminate against people living on computer hardware, simply because they look different. This is another reason AGIs will need protection.

Okay, so these are some of the questions that were asked and some of the answers I gave. I hope you found this useful. I will be giving another talk in San Francisco in a little over a week, and we’ll need to check if guests are welcome, and if so, I will announce the details. In the next episode, I would like to revisit the previous episode about function replication, but want to zero in on the notion of what a problem is, because finding problems and solving them is the driving force behind the creative algorithm. I hope to see you then.