When will machines become smarter than us, and what happens next? In this clip from his 2019 London Real interview, AI researcher Dr Ben Goertzel tells Brian Rose why he is sticking with Ray Kurzweil’s date of 2045 for the singularity, how quickly things could move once a human-level AI exists, and why the real question is not intelligence but values.
It is a calm explanation of an idea often treated as science fiction, from someone who has spent decades working on artificial general intelligence (AGI).
▶ Watch the full episode: Dr Ben Goertzel – Will Artificial Intelligence Kill Us?
Why 2045?
Brian asks when the singularity is coming. Kurzweil’s prediction is 2045, and Goertzel says “I’ll stick with that.” It could be sooner or later, he says, anywhere from around 2030 to 2070, but he does not think it is centuries away, and he does not think it is right around the corner either.
From one human-level AI to a million
Brian suggests that once it starts, it will happen fast, and Goertzel agrees that is the whole point. He offers a thought experiment. Suppose that by 2025 a team such as Google DeepMind, or his own at SingularityNET and OpenCog, builds an AI that is roughly as smart as a person: it can hold a conversation, prove maths theorems and analyse a biology data set. Unlike the Sophia robot today, it would really understand what it was seeing and doing.
The difference from a human is what you can do next. You can teach it computer science, programming and hardware engineering, then copy it a million times. Half of those copies could work on improving their own intelligence, and unlike human researchers, they could share their thoughts directly in the cloud. Within a few years of reaching human-level AI, he thinks it is “very plausible” that it would invent all manner of new things. Lab work and hardware still take time, he adds, but the AI would be working on speeding those up too.
The values problem
Could it simply take us out of the process? Goertzel says that depends on its value system, which is why he wants AI to learn kindness, compassion, tolerance, inclusiveness and love for all sentient beings. You cannot just program those in, in his view. You give the AI the ability to learn them.
The difficulty is that human values keep changing. He points to suburban New Jersey, where he grew up in the 1970s, when gay marriage was illegal and racism was rampant. Fix an AI’s values at 2019 levels, he argues, and by 2050 it would be like being ruled by a dictator with the values of 1950s America. What he wants is an AI whose values evolve along with ours.
Will it treat us the way we treat animals?
Some of his vegan friends ask the obvious follow-up: what if AI treats us the way we treat less intelligent animals? Goertzel takes the worry seriously. We care about whole species going extinct, he says, but not much about one cow or wolf more or less. By that analogy, a superintelligence might want to keep humanity around as its creators, without caring much about any single person.
He is cautiously hopeful. As societies move from scarcity towards abundance, he argues, they tend to show more care for the environment and for other animals. The aim, in his words, is for AIs to be “super compassionate” as well as super intelligent: kinder to us than we are to each other. So far, he admits, we are not setting the best example.
Who is Dr Ben Goertzel?
Dr Ben Goertzel is an American AI researcher and the founder and CEO of SingularityNET, a project to build a decentralised network for AI. He is one of the best-known figures in artificial general intelligence research, has led the OpenCog open-source AGI project, and worked with Hanson Robotics on the AI behind the humanoid robot Sophia.
This clip comes from his first London Real episode, Will Artificial Intelligence Kill Us?, recorded in 2019. He has returned many times since, including for Why OpenAI Really Fired Sam Altman in 2023.
See everything Dr Ben Goertzel has done with London Real on his guest page: Dr Ben Goertzel on London Real.
More on London Real
Transcript
Transcript: Dr Ben Goertzel – Why The Singularity Could Arrive By 2045
0:00 Brian Rose: When do you think the singularity is coming? Ray Kurzweil says 2045.
0:07 Dr Ben Goertzel: I’ll stick with that.
0:08 Brian Rose: That’s 26 years.
0:09 Dr Ben Goertzel: Yeah. It could be sooner, it could be later. It could be 2030, it could be 2070. I don’t think it’s 2200. I don’t think it’s 2020.
0:17 Brian Rose: [unclear] It’s too soon.
0:20 Dr Ben Goertzel: I don’t know, man.
0:21 Brian Rose: I guess when it starts to happen, it happens real fast.
0:24 Dr Ben Goertzel: Well, that’s sort of the point, right? So suppose, as a thought experiment, that by 2025 Google DeepMind, or my own AI team in SingularityNET and OpenCog, one of us manages to create an AI that is essentially as smart as a human being. So it can hold a conversation like a human being, it can prove maths theorems like a human mathematician, it can analyse a biology data set.
0:58 Brian Rose: A bit like your robot, Sophia.
1:01 Dr Ben Goertzel: And you put it in a robot body. I mean, that’s a separate problem. It could operate many robot bodies all at once. But unlike Sophia right now, it can really understand what it’s seeing and what it’s doing, fully, at the level a human can. So suppose you get there: how far are you from a true singularity? Because this AI can then copy itself. You can make a million of those.
1:29 Brian Rose: And improve itself.
1:30 Dr Ben Goertzel: Yeah, right. Because once the AI is as smart as a human, OK, let’s teach it computer science. We can send it to school. It can learn maths, it can learn programming, it can learn hardware engineering. And then you can copy a million of those. Of all million of those, maybe half are working on improving its own intelligence, but they’re not working in isolation like humans are. They can share thoughts directly, because they’re all just running in the compute cloud.
1:58 Dr Ben Goertzel: So it seems very plausible that within a few years of getting to that human-level AI, it’s going to invent all manner of amazing new things. It doesn’t mean it’ll be instantaneous. Doing lab research still takes time. Building new hardware still takes time. But of course it can be working on how to speed that up: how to more fully automate manufacturing and laboratory research.
2:23 Brian Rose: And then it could take us out of the process as well, possibly.
2:26 Dr Ben Goertzel: It could. I mean, that depends on the value system that AGI has. And this is why it’s important to give it values of kindness, compassion, tolerance, inclusiveness, love for all sentient beings. We want to get these positive values into the AI as much as we can.
2:51 Brian Rose: And we think we can program that into AI?
2:54 Dr Ben Goertzel: I think you teach it. You program it with the ability to learn that.
2:57 Brian Rose: But then won’t it unlearn that if it doesn’t suit it? Or is that what we don’t know?
3:02 Dr Ben Goertzel: It’s very subtle, because human values are a moving target. The values of suburban New Jersey in 1975, when I was in elementary school, are not the same as the values in suburban New Jersey right now. Back then, gay marriage was illegal and virulently opposed by almost everybody, and racism was very rampant there. So human values are evolving all the time. By the values of medieval Europe, you and I and probably almost everyone listening to us deserve to burn in hell forever.
3:40 Brian Rose: Yeah.
3:41 Dr Ben Goertzel: So if you gave the AI exact human values from 2019, then by 2050 it’s going to be horrible. It’d be like having a dictator with the values of 1950 America or something. So you want an AI that will have evolving values, in a way that somehow respects the ongoing evolution of human values.
4:06 Brian Rose: And that hopefully still respects it and doesn’t just make its own version up.
4:09 Dr Ben Goertzel: Yeah, which is very subtle. Some of my friends who are vegans and animal rights activists are like: well, what if the AI treats us the way we treat less intelligent animals? And think about it: we care about the extinction of species. Not as much as we should, but in general we don’t want to make entire species of subhuman animals extinct. But we don’t care much about one cow, sheep, pig or wolf more or less. We care about the genome.
4:43 Dr Ben Goertzel: So if you took that analogy, the AI would like to keep humans around because we’re the creators, we’re the ancestors, we have our own unique aesthetic value. But by that analogy, it may not give a crap about one human more or less, any more than we care about one wolf or pig more or less.
5:03 Brian Rose: We didn’t keep any Neanderthals around.
5:05 Dr Ben Goertzel: No, but we weren’t as reflectively intelligent then as we are now. And I think there is an argument that as human cultures advance more and more toward abundance and away from scarcity, there’s more caring about the environment, there’s more compassion toward non-human animals. There was a lot in Stone Age society, it sort of went down in the Industrial Revolution, and now it’s certainly going up again.
5:34 Dr Ben Goertzel: But I guess the point is, we want our AIs not only to be super intelligent, we want them to be super compassionate. We want them to be more compassionate to us than we are to each other, because humans are killing each other for no good reason all the time. So we want them to be morally better than us in some sense.
5:58 Brian Rose: And we hope they are.
6:00 Dr Ben Goertzel: On the whole, we’re not setting the best example, in terms of the way human society is being regulated right now, nor the way we treat non-human life forms on the planet, nor the applications that we’re deploying AI toward. Like, what are all the AI PhDs [unclear]
