🤖 Analyzing AGI: What would the world of artificial general intelligence look like?
Also: 5 Quick Questions for … economist Jesús Fernández-Villaverde on AI and economic growth
Quote of the Issue
“AI will usher in an age of abundance, assuming that we're in a benign AI scenario.” - Elon Musk
The Essay
🤖 What would the world of artificial general intelligence look like?
My humble contention: A big reason why opinion polls suggest considerable concern about recent AI advances is that our culture has produced few visions of a positive future with supersmart computers. So we all wonder: What will our society be like in a world where computers can do much of what we currently do (weak artificial general intelligence), can do all of what currently do (strong AGI), and, finally, can do what we can’t even imagine doing (artificial superintelligence)? Now whether each of those performance levels is possible is another question — and would eventually require big advances in robotics, as well as AI. (Some would consider strong, human-level AGI to be superintelligence.)
Yet we’re still thinking about such questions. And worrying about the possible answers. By now, the GenAI craze has reminded us all of the many negative cultural visions about technology, especially as presented by Hollywood. Skynet, the self-aware AI in the Terminator franchise, is probably the one that gets cited most frequently.
Of course, the many incarnations of Star Trek offer a very take different take on AI. Clearly, the main computer on the Enterprise and other Star Fleet vessels is highly advanced and able to both retrieve information and analyze that information. But it isn’t sentient, although its performance in the virtual-reality holodecks shows it could easily pass as human. What’s more, the main computer seems aligned with human goals with nary a worry about it taking over or acting destructively. We can probably conclude it is a powerful AGI, able to do everything humans can do within a broad sector of responsibility; one version is good at helping run a starship, and another assisting doctors in sickbay, such as the Emergency Medical Hologram on Star Trek: Voyager.
One more example: the AI in The Expanse book and television series, which takes places 300 years in the future when humanity is spread throughout the Solar System. AI in The Expanse is both everywhere and nowhere:
One of the reasons humanity's expanding presence in the solar system is made possible is because of all of the invisible tendrils of sophisticated automated intelligence that are integral to most of the tools present in this interpretation of the future. Many of the transports, research equipment and heavy machinery all represent advancements in synthetic intelligence. There are medical bays that can diagnose ailments via scans and blood work with the speed of thought as well as provide automated surgical treatments. Gunships traverse the space ways by independently routing courses, at times piloting themselves and use complex targeting algorithms to lock onto precise locations at vast distances. AI exists everywhere in The Expanse, its just incorporated into the housing of every daily tool.
Indeed, many Expanse fans have commented on how little attention is paid to AI despite being ubiquitous. Here’s author James S.A. Corey (the pen name of authors Daniel Abraham and Ty Franck) on the issue:
What we have is uncommented automation. It's all around the characters all the time but it's uncommented because it's unremarkable to them. The Roci [spaceship] is constantly described as 'smart', and [crew member] Naomi is always giving it complex tasks to work on. The med bay is basically a computerized hospital requiring almost no human intervention. If you mean AI as in self aware or sentient machines? Yeah, we avoid that because we're both sort of bored by it. Humans are far more interesting
“Uncommented automation” is a great way of thinking about a comforting vision of AI: a tool that helps humanity without fundamentally changing our civilizational structure or our relations with each other. I should point out, however, that in The Expanse most people seem to be unemployed and are paid a universal basic income. Then again, I think the economics of the show are screwy. Given the apparent technological sophistication of AI, even a world of more inequality should be richer. There shouldn’t be tent cities of UBIers as there are in The Expanse. Still, humanity is running the show and it’s a world we recognize. A world with sentient, self-improving AI where the economy is doubling every month or every week isn’t one I think economists have much to say about nor is it often depicted in science fiction.
It’s AGI rather than such superintelligence that’s examined in a great new piece in The Economist:
For a start, we suppose that AI will be benevolent, controllable and distinguishable from humans. We also suppose that human culture will not be radically altered by technological progress to the point that people begin to love or even worship ais. Instead, we imagine AI as a tool: a virtual, super-smart, dirt-cheap bot. We assume that constraints on the widespread use of ai, such as energy limits, will be resolved. None of this is guaranteed, but it helps make an exercise like this possible.
What will we do in this world? Well, according to the piece, we won’t stop working altogether because work has become a "consumption good," offering far more utility than the income it generates. Work can provide a sense of purpose, social interaction, and personal satisfaction. For example, some people enjoy working in creative fields, such as writing or art, because it allows them to express themselves and share their ideas with others. Others find satisfaction in helping others, such as by working in a healthcare or social service field — and humans might still be better at showing empathy. And still others simply enjoy the challenge and stimulation of working in a demanding field. Humans might specialize in, quoting sci-fi author Iain Banks, “the things that really [matter] in life, such as sport, games, romance, studying dead languages, barbarian societies and impossible problems, and climbing high mountains without the aid of a safety harness.”
What’s more, as seen in both Star Trek and The Expanse, politics will still be a human career. While some policymaking tasks may be delegated to AI — finally Milton Friedman’s dream of an automated Federal Reserve System might be realized — humans will likely continue to play a central role in decision-making. There might also be a human premium on some activities because humans are doing them. From the piece: “Artificial diamonds, which have the same molecular structure as those from the ground, trade at an enormous discount—around 70% by one estimate. In the future, items with a ‘made by a human’ tag might be especially desirable.”
Of course, it’s probably as difficult to predict life in the world of AGI as it is to predict how technology in general will affect our lives or, say, what jobs it will create. But there’s no good reason to paint purely dystopian visions. Right, the balance is way too far to the negative.
5QQ
💡 5 Quick Questions for … economist Jesús Fernández-Villaverde on AI and economic growth
One staple topic of this newsletter is the slowdown in productivity growth the United States has experienced since 1973. Thankfully, the 2020s have seen the arrival of a new technology in deep learning artificial intelligence algorithms that may just be the next General-Purpose Technology. Like the steam engine or electrification before it, AI could perhaps jump-start productivity and economic growth. For more on that topic, I reached out to my AEI colleague Jesús Fernández-Villaverde. Fernández-Villaverde is a professor of economics at the University of Pennsylvania, where he serves as director of the Penn Initiative for the Study of Markets. He’s also the John H. Makin Visiting Scholar at the American Enterprise Institute.
1/ As an economist, how do you think about the cause of economic growth?
If you take my class at Penn, you will never heard the word “capitalism” in the whole semester because capitalism somehow highlights that what matters for economic growth is capital, and it is not. What matters for economic growth is ideas. Take the richest economies in the world today: We are around 30 times richer in per capita terms that when modern economic growth started in the early 18th century. We didn't become 30 times richer because we accumulated capital. Accumulating capital allows you to become twice as rich or three times as rich, but not 30 times.
The way you can think about ideas is that a new idea is combining previous ideas. Let me imagine that I invent ice cream, a great contribution to human welfare, at least from my perspective. And you invented the idea of green tea. Ice cream existed for centuries, green tea existed for centuries, and in some moment in the 1950s, someone comes up with idea of green tea ice cream, which now has become very popular and is one of my favorite flavors. The point I'm trying to make is that we took two ideas and created a new one.
2/ And how will AI help us innovate better by combining ideas?
What artificial intelligence is fantastic at is searching for patterns among all existing ideas and helping us to become so much better at sorting them out and recombining them. What is the problem we have faced over the last 20 or 25 years? It has become harder and harder to even know the ideas that already exist. I try to teach graduate students what the frontier of the field is. In my area of economics, once you finish the regular PhD courses, in 1996 or 1997, it would take you 12 months to learn all that’s out there. Now it takes you at least 24 months. It's much harder just to understand everything, to keep track of everything. What artificial intelligence is great at is helping you explore what already exists, find it out, and combine it in much more innovative ways. So artificial intelligence can definitely help us grow more.
3/ Is this an opinion you held two or three years ago, or has it changed because of generative AI?
I see three scenarios. Scenario one is that there are fewer people and we are running out of ideas. Let me call that the pessimistic scenario. One year ago I gave a 50 percent probability to that negative scenario. There is a middle-of-the-road scenario where we keep growing more or less at the same speed as before. One year ago, I gave it 40 percent probability. Artificial intelligence is so great, we grow much faster: I gave that 10 percent probability. After ChatGPT and everything we have seen over the last few months, I have reduced the 50 percent to 40 percent, the 40 percent to 30 percent, and [raised] the 10 percent probability of growing much faster to 30 percent.
It’s a significant increase in my forecast. Now, my forecast is still that the most likely scenario is we slow down. I'm cautiously more optimistic than one year ago, but I'm still not wildly optimistic.
4/ When you estimate a 30 percent probability of “much faster” growth, how fast do you mean?
Let me give you a benchmark number. When I start my class in economic growth at Penn, I tell the students there is only one number you need to learn: average economic growth in the US in per capita terms since the end of the Civil War in 1865 to today is 1.9 percent. The middle-of-the-road scenario is we continue at 1.9 percent. The pessimistic scenario is we fall to 1 percent. Optimistic, let's say 3 percent in per capita terms.
Now, this is a huge difference. If something grows at a 1 percent per year, it doubles in 70 years because of compound growth. If it grows at 2 percent, it’s 35 years. If it's growing at 3 percent, it means you multiply by two every 23 years. That's a huge difference. So even growing to 3 percent growth thanks to artificial intelligence is a complete game changer.
Imagine a genie shows up over here and tells me, “Jesus, with a probability of one, we are going to grow at 3 percent in per capita income for the next 50 years.” What happens with the Medicare and the Social Security entitlement problems? This is more or less as significant for the US economy as the problem of taking out the trash for my kitchen tonight: something that can be done in five minutes. The problem for Social Security and Medicare is if we continue growing at 1.9 percent, or even worse, dropping to 1 percent.
5/ Some people are making much more bullish predictions based on what AI can do today. Why do you only assign a 30 percent probability to that optimistic scenario?
Why am I only giving it a 30 percent probability? Because the US economy provides you with tons of services. Some of them are very high-tech, like my cell phone, Netflix, Amazon Prime, and those can be potentially revolutionized by artificial intelligence. The US economy also depends crucially on a lot of manual labor. But we are not even close to anything that can clean my bathroom. That's called Moravec’s paradox. We already know how to code a program that takes derivatives, but it’s difficult to automate something as simple as: open the trash can in my kitchen, take the trash bag out, tie it up, take it outside to the trash container, come back and put a new bag in. I can force a six-year-old kid to do it without any problem, but we don't have computers that do that. And the problem is, a lot of our modern economy is about that.
It's a challenge. I wish AI will get us out of a lot of these problems. It's not going to get us out of as many of them as some optimistic people think. And it doesn't really fix all the social issues of having a society that is just old and tired and doesn't feel like doing anything about the future.
Micro Reads
▶ Taming the stars - John Myers, Works in Progress |
▶ Governance of superintelligence - Sam Altman, Greg Brockman, and Ilya Sutskever, OpenAI |
▶ What Makes A.I. Chatbots Go Wrong? - Cade Metz, NYT |
▶ A.I. Needs an International Watchdog, ChatGPT Creators Say - Gregory Schmidt, NYT |
▶ The Emergence of Economic Rationality of GPT - Yiting Chen, Tracy Xiao Liu, You Shan, and Songfa Zhong, Arxiv |
▶ Do new Alzheimer's drugs signal the end of the condition? - Clare Wilson, New Scientist |