Cover art for ‘03 – The Current State Of AGI Research’

03 – The Current State Of AGI Research

Artificial Creativity

Listen on Apple Podcasts Listen on Google Podcasts Listen on Spotify

References:

Further reading:

  • David Deutsch, “The Beginning of Infinity”, chapter 1 “The Reach of Explanations”, and chapter 6 “The Jump to Universality”

If you have any questions or comments, please leave them here or tweet at me: https://twitter.com/dchackethal

Subscribe to this podcast via RSS Back to all episodes

Transcript

This transcript was generated automatically and may contain mistakes.

Hello and welcome back. This is the third episode of the series on artificial creativity. Thank you for bearing with me in terms of audio quality. I have a microphone now, and I’m trying it out for the first time, so hopefully the audio is better this time around. If you hear some frogs croaking outside, it turns out that in the springtime, frogs like to populate the pond outside my house, but I hope that won’t get in the way of us doing this episode. I think it can actually sound kind of calming, and I hope you don’t mind.

In the last episode, we spoke about why the creative algorithm, once built, wouldn’t be dangerous, at least no more dangerous than people are, because its repertoire would be the same as that of people, and indeed it would be a person. In this episode, I would like to explore the current state of AGI research. AGI has not been built yet, and we don’t yet know how to build it, and since it’s qualitatively different from all other computer programs, a good rule of thumb is, if it can already be built, it has nothing to do with AGI. That’s why you can disregard newspaper stories that claim AGI is here.

As far as I’m aware, there are three main fields that claim to work on the problem of how to build an AGI. Current AGI research, this is a sort of direct or targeted research, and then there’s neuroscience and philosophy. It’s mostly those three. Now, people don’t like when you dismiss fields like that, and so I’ll immediately concede that we cannot predict how knowledge in these fields will grow, and so they may surprise us by contributing something of value in the future. And even if they don’t, these fields are still interesting in their own right and in regard to the problems they can solve. This is also a good time to point out that even though I quote David Deutsch often, that doesn’t mean that he agrees with me on any of these points.

But I am going to argue that, sadly, these fields are, at least currently, largely poorly equipped to address the problems of how to build an AGI, and therefore most likely won’t make much progress towards this goal. The goal is, as I said, not to build an AI, but an AGI. And to reiterate, an AGI is the same as a general purpose explainer, and a general purpose explainer is the same as artificial creativity. And just to remind you, I use these terms interchangeably. We will look at all three in order. Current AGI research first, then neuroscience, then philosophy.

The reasons current AGI research is limited in its capacity to contribute is that, even though it presents itself as AGI research, it’s really AI research. That is, it explores the space of narrow programs. It makes the mistake of thinking that if you add new features to a growing list of abilities, you will somehow, eventually anyway, have a general purpose problem solver. We saw in the last episode that this cannot be so, but it’s worth repeating. It can’t be so because the repertoire of an AGI is inexhaustible because it’s infinitely big. That means no matter how many things you add to the list, you’re not getting any closer to infinity.

One prominent self-proclaimed AGI researcher is OpenAI in San Francisco. At the time of recording this episode, that is the spring of 2019, their mission is, quote, discovering and enacting the path to safe artificial general intelligence, end quote. We know now that the focus on safety is a red herring, but at least they’re working on AGI, right? Well, looking at the work they’ve done and showcase on their website offers a different picture. Among other things, they have created a musical composer, a text predictor, and a piece of hardware that looks like a human hand that can manipulate objects. You could put all of them together to make a system that can do all three things, but again, that doesn’t get you closer to an infinitely big repertoire.

They have a section on their website called OpenAI Gym, which is an environment to develop and run AI systems. It features examples of very specific algorithms and so-called toy problems, and thus entrenches the alleged values of specificity in the minds of young contributors to the field. As a side note here, John Koza, a pioneer in the field of genetic programming, is a notable difference. His website, I will put a link in the description, states that, quote, while proof of principle and toy problems are occasionally useful for tutorial or introductory purposes, we believe that it is time for fields of artificial intelligence and machine learning to start delivering non-trivial results that satisfy the test of being competitive with human performance, end quote.

His website was last updated in 2010, so perhaps some would argue that this has happened since, for example, with AlphaGo, but that still puts too much emphasis on narrow applications. I will mention genetic programming again in a later episode because it has interesting parallels to Popperian epistemology, and it may be one of the tools that could help build AGI.

Another company is DeepMind in London. They seem to avoid the term AGI, but their mission statement is clear. Quote, we’re on a scientific mission to push the boundaries of AI, developing programs that can learn to solve any complex problem without needing to be taught how, end quote. So far so good, but they don’t make good on that quote. They write papers in very specific domains, all interesting in their own right. For example, they received quite a bit of fame for their champion Go player called AlphaGo, but they do not work on general purpose problem solving. Some companies claim to be working on more of a general purpose solution, for example, general-purpose language comprehension, but that is still too narrow.

In addition to narrow scope, there is the issue that current AI researchers do not work on the problem of knowledge creation. They work on the execution of particular tasks. As I said in the previous episode, terms like machine learning are therefore particularly misleading and damaging, because they create the impression that knowledge is created at runtime, when in reality, these programs apply knowledge that was created by humans first.

There is some hope, though. It is possible that making a breakthrough in a specific field can open that field up and make its repertoire explode to infinity. But when this isn’t the goal, people tend to go out of their way to constrain the program to only perform the specific purpose it was designed for. For more on this, I recommend you read the chapter “The Jump to Universality” in “The Beginning of Infinity”. Among other things, it offers an interesting parallel to how AI researchers today are discouraged from universality in the domain of creating knowledge.

It tells how in ancient Greece, Archimedes repeatedly extended the Greek number system and each time it ended up almost universal. But it seems he chose to keep arbitrary restrictions, perhaps because, quote, he felt that he had to avoid aspiring to infinite reach in order to make a convincing case, end quote. This is the kind of situation AI researchers find themselves in today. If one of them tells leaders in the field in Silicon Valley that he has an idea for how to build an AGI, they will tell them to narrow the problem down and run it in the OpenAI Gym. This is no small problem. It actively discourages striving for infinite reach in the domain of problem solving.

How about neuroscience? The hope here is that if we study the human brain, or the nervous system as a whole to be exact, we can understand better, among other things, how people learn. So at least there is a real focus on the question of how knowledge is created. But unfortunately, it will fail. That is because neuroscience focuses on the study of the brain, which is hardware. But creativity is not a piece or property of hardware. It’s a piece of software. By the way, this is why robotics research that focuses on hardware is probably orthogonal to AGI research as well, and it’s also why doomsday scenarios of the so-called robot apocalypse are unlikely.

Software is an abstraction. It’s independent of its physical instantiation, and so I think of the brain as hardware, the mind as the operating system, and the creative algorithm as a particular piece of software running on the mind. Sure, this software can’t run without some hardware to run on, but it doesn’t discriminate much between different forms of hardware as long as they meet the following condition. They have to be universal Turing machines, that is, general-purpose computers. Why can’t creativity run on a special-purpose computer? Because creativity itself is general purpose and therefore relies on the computer to support its infinite repertoire.

The human brain is host to creativity. We know this because you and I are creative. And since creativity requires general purpose computation, we therefore know that the brain is, or at least contains, a universal Turing machine. And this universality is the crucial bit. All universal Turing machines, by definition, have the same repertoire. And the laptop I’m recording this on is also a universal machine. That means we already know that it could, in principle, run an AGI, as long as it’s given the knowledge of how to do so in the form of a program. And my brain, or rather my mind, is running an AGI right now as well.

Yet it’s hard to imagine two more different pieces of hardware than a brain and a computer. One is a wet, organic piece of meat, the other a dry, inorganic piece of metal and silicon. Thinking that there is something special about the brain in terms of hardware, for example, the neuronal structure or the so-called massive parallelism, violates the universality of computation, as David writes in his Aeon article.

In the CBC interview with David that I referenced in the last episode, he explains why this matters. Quote, if you want to understand computer programs, it’s completely hopeless to study metal and silicon. End quote. For example, if you want to understand how a web browser works, which is a piece of software, you will learn, roughly, that it works by making requests to servers over the internet and parsing HTML to display the webpages it receives in response to the user. This is a rough but good explanation, and it’s devoid of anything to do with the underlying hardware.

Another way to look at it is this. As I have said, the computer I’m recording this on would not need to change its hardware in order to run the creative algorithm because it’s already universal. Suppose one day we know how to write it, and you have two computers in front of you. Their hardware is identical. They both have the same memory and processing power, the same architecture, and so on. You’re told that one of them has their creative algorithm installed, but you don’t know which one. You’re asked to find out only by comparing the two computers’ hardware. But they’re identical.

If apart from the creative algorithm, both computers have the exact same software installed, you may be able to read the zeros and ones off each hard drive and compare them to reverse engineer the creative algorithm. But this requires a lot of working knowledge of how software interacts with hardware that, as far as I’m aware, we have for computers, but not yet for brains. And besides, even if we did, I’m not sure how you would find someone who is not creative to compare with the person who is.

So to sum up the neuroscience piece, as long as the underlying hardware is a universal Turing machine, it can run a creative mind. The hardware specifics do not matter as long as it meets this one criterion. This kind of universality allows software to transcend previous hardware limitations and become platform independent. This is why I don’t expect big contributions to AGI to come out of neuroscience. We can skip the hardware and should focus on the software instead. This is why I said in the first episode that creativity is a software project.

As a side note here, the fact that creativity, like all software, is, as we may call it, platform independent, is a really good thing. It means that in theory, we could one day leave our human bodies behind and transfer to a more stable, more slowly aging, and less disease-prone hardware. I think anyone who’s ever had more than a cold in their life can appreciate the benefits of that.

On to philosophy. I said in the previous episode that since creativity is about knowledge creation, we need to study how we know. That’s the field of epistemology. And, like I said, the biggest and best contributions to the field, after about 2,000 years of standstill, by the way, have been made by Karl Popper and David Deutsch. Unfortunately, Popperian epistemology is largely unknown or undervalued, and most people in the field focus on so-called justified true belief, even though Popper showed that this is the wrong way to think about knowledge. But most people just ignore him. This is why philosophers who are stuck thinking about justified true belief will not be able to contribute unless they understand Popperian epistemology or replace it with something better.

Again, all of these fields are interesting in their own right, but it’s when we think they make genuine contributions to the field of AGI, even though they don’t, that we may become complacent and a breakthrough may be absent for decades. I think this is why, in the roughly 70 years of computer science so far, we have made little progress towards AGI. And it’s why I think that any effort not directed toward epistemology is futile.

There is a silver lining, though. Especially AI research and neuroscience take a lot of specialization. So the good news is that if you want to contribute to the field, you don’t need to study machine learning. And you don’t need to get your PhD in neuroscience and work at a lab. You need philosophy. In particular, you need to solve an epistemological problem, because we do not know how to build an AGI without understanding how knowledge creation works.

Unfortunately, philosophy gets a bad rap in the mainstream for being too theoretical or navel-gazing. And to be fair, there is a lot of bad philosophy out there that is just that. But people generally do not realize that there are real philosophical problems that need solving, and that building an AGI is one of those. But I think the problem can be phrased without stating it in philosophical terms, by packaging it as a software engineering problem instead.

This is why software engineers who have not yet worked on AI and are not familiar with justified true belief could be good candidates to contribute. I say that because software engineers have already made contributions to epistemology without realizing it, by finding rules governing the improvement of the structure of programs. Good programs are hard to vary. And it turns out that this applies to explanations as well. We will come back to this in a later episode, but if you would like to dig into the concept of hard-to-vary explanations, as well as the folly of justified true belief now, I recommend you read chapter one of “The Beginning of Infinity”, called “The Reach of Explanations”.

Thank you very much for listening. In the next episode, we will explore and try to answer the question, how do we know? We will also study some of the remaining problems in that field, particularly as they relate to creativity. If you have any questions or comments, please leave a note or tweet at me. Again, a link to my Twitter profile is in the description. See you next time.


What people are saying

What are your thoughts?

You are responding to comment #. Clear
Markdown supported. cmd + enter to submit. You have free speech here. You’re responsible for what you write. Terms, privacy policy
Your real name is preferred.
This small puzzle helps protect the blog against automated spam.

Sign your comment with GPG to create or add to a public profile with all your comments.

Paste a detached signature of your comment.

              
Paste your public-key block if you haven’t before. You consent to your key’s contents, including your name, being displayed to the public.

              

Preview

Playing ~