“If AGI is going to create a beneficial technological singularity, if it’s going to be good for all humanity… it can’t be siloed and beholden only to the narrow goals of a very small group of people.”
In short: Recorded backstage at Consensus 2023 in Austin, Texas, this short interview has SingularityNET CEO Dr Ben Goertzel, who introduced the term AGI, explaining Zarqa, a SingularityNET ecosystem company set up to build the next generation of large language models. He argues that systems like GPT-4 hallucinate and only simulate invention, and that the next leap will come from combining transformers with logical reasoning and evolutionary learning, rolled out on open, decentralised infrastructure.
Dr. Ben Goertzel is a well-renowned figure in the field of artificial intelligence (AI), who emerges not only as a distinguished researcher but as a visionary architect shaping the future of intelligent systems. His multifaceted contributions extend beyond the technical intricacies of AI, encompassing philosophy, entrepreneurship, and a profound commitment to creating machines that not only mimic human intelligence but understand and collaborate with it.
This conversation with Dr. Ben Goertzel was recorded backstage at this year’s Consensus event in Austin, Texas. Consensus is the world’s largest, longest-running and most influential gathering that brings together all sides of the cryptocurrency, blockchain and Web3 community, and it was great to catch up with Ben at the event!
Dr. Ben Goertzel’s academic journey traversed the intricate realms of mathematics, computer science, and cognitive science. His early academic pursuits at Columbia University laid the foundation for a comprehensive understanding of the theoretical underpinnings of artificial intelligence. This culminated in his doctoral studies at Temple University, where he delved into the mathematical foundations of cognitive science, providing him with the intellectual toolkit to approach AI challenges with a unique perspective.
In the late 1990s, Dr. Goertzel co-founded Webmind, an ambitious project aimed at creating a global brain by integrating advanced AI technologies. Although the project faced challenges and was eventually disbanded, it represented a pioneering effort to push the boundaries of AI research and develop systems capable of emulating human cognitive processes.
One of Dr. Goertzel’s seminal contributions to the field is OpenCog, an open-source software framework designed for the collaborative development of artificial general intelligence (AGI). This ambitious project adopts a holistic approach, combining symbolic and sub-symbolic reasoning to simulate complex cognitive processes. OpenCog reflects Goertzel’s commitment to building not just intelligent systems but those capable of reasoning, learning, and understanding in a manner analogous to human cognition.
In recent years, Dr. Ben Goertzel has been at the forefront of merging AI with blockchain technology, giving rise to SingularityNET. This decentralised AI marketplace empowers AIs to communicate, collaborate, and transact autonomously, fostering a democratised and transparent ecosystem for AI development. Goertzel’s vision transcends traditional centralised models, envisioning a future where AI is accessible, collaborative, and beneficial for the global community.
Dr. Goertzel’s contributions extend beyond the technical aspects of AI to encompass profound philosophical reflections on consciousness, intelligence, and the ethical implications of advanced AI systems. His insights into the potential future coexistence of humans and intelligent machines, coupled with considerations of the ethical responsibilities associated with AGI, showcase a thought leader who not only engineers AI but contemplates its societal implications.
Dr. Goertzel’s influence extends globally through collaborations with researchers, developers, and organisations. His commitment to fostering a global community of AI enthusiasts reflects a belief in the collective intelligence necessary to navigate the challenges posed by advanced AI technologies. As a thought leader, he continues to contribute to global discussions on the evolving nature of intelligence, the role of technology in society, and the ethical considerations surrounding AI development.
Dr. Ben Goertzel’s journey in the realm of artificial intelligence is an odyssey marked by visionary pursuits, groundbreaking research, and a commitment to shaping the future of intelligent systems. From his early academic endeavours to founding pioneering projects like Webmind and OpenCog, and his current role in leading SingularityNET, Goertzel’s impact on the AI landscape is profound. As we stand on the cusp of an AI-driven future, Dr. Ben Goertzel’s contributions serve as a guiding light, illuminating a path towards a more collaborative, transparent, and ethically grounded era in artificial intelligence.
Key takeaways
- GPT-4 is useful but shallow. Goertzel praises OpenAI for releasing transformer models Google thought too rough, but says they cannot tell reality from nonsense and merely recombine their training data.
- Logic keeps creativity honest. He wants to pair transformers with OpenCog Hyperon’s logical reasoning to ground statements in truth, and with evolutionary algorithms to invent beyond the data.
- Zarqa grew out of SingularityNET. He describes Zarqa as a dedicated company formed by SingularityNET ecosystem members to build neurosymbolic, decentralised LLMs after ChatGPT.
- OpenAI is already hybridising. He notes OpenAI’s work with Wolfram Alpha and Microsoft’s knowledge graph and says GPT version numbers are becoming a brand rather than a single algorithm.
- Advantage may shift to deep expertise. Once LLMs are combined with other AI modules, he argues, the edge goes to whoever best understands those modules, not whoever shipped GPT first.
- Open AGI serves more people. He says an AGI beholden to a small group cannot serve all humanity, so the code should be open and the deployment decentralised.
- 2029 remains his AGI estimate. Speaking in 2023, Goertzel backed Ray Kurzweil’s 2029 forecast for human-level AI, give or take a few years, but expected superhuman AI soon after rather than in 2045.
- Fear of AI is Hollywood thinking. He accepts the future is uncertain but sees no reason to expect a Terminator outcome if AI is built for compassion and governed by participatory infrastructure.
Frequently asked questions
Dr Ben Goertzel is a cross-disciplinary scientist, futurist, author and entrepreneur who introduced the term artificial general intelligence (AGI) around 2004–2005. He earned his PhD about 35 years before this 2023 interview, leads the OpenCog Hyperon AGI project and is CEO of SingularityNET, a blockchain-based platform for decentralised AI services. He has spent most of his career trying to build benevolent, human-level and superhuman AI.
Brian Rose spoke to Goertzel backstage at Consensus 2023 in Austin, Texas, one of the largest gatherings in the crypto, blockchain and Web3 world. The episode was published on London Real in June 2023. At around 22 minutes, it is a short, focused conversation about Zarqa, neurosymbolic AI, open AGI and Goertzel’s timeline for human-level machine intelligence.
In this interview Goertzel describes Zarqa as a company put together with SingularityNET ecosystem members to build the generation of large language models that follows ChatGPT and GPT-4. Its approach combines transformer neural networks with logical reasoning and evolutionary learning, running on decentralised infrastructure. Brian Rose said it would be based in the Middle East. London Real Ventures invested in Zarqa, so viewers should keep that interest in mind.
Goertzel said such systems cannot tell reality from nonsense, have a weak ability to ground statements in observation and hallucinate constantly. He added that their inventions show repetition and banality because they are simulating creativity by recombining training data. He stressed these flaws do not cancel out the products’ usefulness, but they leave a long way to go before real thinking machines.
As Goertzel explains it, neurosymbolic AI combines neural networks, such as the transformers behind GPT models, with symbolic systems that perform logical reasoning over knowledge. The logic component lets a system reason step by step and ground its statements in truth. He also wants to add evolutionary learning, using genetic algorithms to invent new ideas, and argues the logic is needed to keep that creativity from becoming wild.
Goertzel believes so. He pointed to OpenAI integrating Wolfram Alpha and Microsoft’s knowledge graph as neurosymbolic steps taken quietly behind the scenes, and said serious AI researchers know a transformer with reinforcement learning on top will soon hit a wall. Because of this, he argued, the advantage could move to teams with deep knowledge of the other AI modules being combined.
Asked by Brian Rose, Goertzel said that he introduced the term artificial general intelligence, around 2004 or 2005, long before AI became as popular or as capable as it is now. He said he had been working on AI and towards AGI since completing his PhD roughly 35 years earlier, and that modern computing scale now lets ideas become working systems far faster.
He argues that an AGI beholden to the narrow goals of a small group cannot, almost by definition, serve all humanity. Open code, decentralised deployment and ongoing input from a broad, diverse group are, in his view, the likeliest route to broadly beneficial AGI and to rollout in every country and market. He contrasted this with ChatGPT, where users cannot see what is filtered.
Speaking in 2023, Goertzel said he was sticking with Ray Kurzweil’s extrapolation of human-level AI by 2029, give or take a few years. He disagreed with Kurzweil’s 2045 date for massively superhuman AI, arguing that a human-level AI could copy itself, rewrite its code and design new hardware, so superintelligence could follow within a few years unless the AI chose to slow down.
Goertzel acknowledged an irreducible uncertainty, comparing AI to past leaps such as agriculture, industry and flight. But he said there is no reason to assume a bad outcome is likely. His teams are building AI for elder care, education, art, music and science, designed for compassion and governed by participatory, decentralised infrastructure, and he called Terminator-style fears thinking by reference to Hollywood movies.
Full transcript
Read the full transcript of Brian Rose’s interview with Dr Ben Goertzel
0:00:00 Introduction: Dr Ben Goertzel backstage at Consensus
Brian Rose: This is London Real. I am Brian Rose.
Brian Rose: All right, I’m here with Dr Ben Goertzel. He’s the CEO of SingularityNET, and I want to ask him what I’m sure you want to know, which is: we want to hear more about Zarqa, which is a neurosymbolic large language model from SingularityNET at massive scale. Everyone’s talking about this new project; I’m calling it the OpenAI killer. Tell us about it. What is it, why do it now, and how do you explain it to people, even on the street, who don’t know that what you’re doing is something different from a lot of these other plays, like what we’re seeing coming out of Silicon Valley?
0:03:53 What OpenAI got right, and where GPT-4 falls short
Dr Ben Goertzel: I think the field of AI is advancing so fast it even confuses those of us who are experts in the field, so it’s understandable that people who are end users and observers are a bit dizzy from the launch of new products and the pace of progress. I think what OpenAI has done with GPT-4, and GPT in this whole series of products, is really, really fascinating. I mean, it’s awesome what the systems can do, and it’s gutsy the way they’ve just put it out there at a relatively raw stage.
Dr Ben Goertzel: The core technology underlying ChatGPT and so forth was transformer neural nets, which were developed by Google Brain in Mountain View in 2017–18, right? But Google found it a little rough around the edges. OpenAI was like, well, yeah, but it’s so cool, let’s put it out there for people to use. I think that was an edgy thing to do, and obviously it’s taking the world by storm.
Dr Ben Goertzel: On the other hand, everyone can see the limitations of this sort of system. It can’t tell reality from nonsense. It has very weak ability to ground its own statements in its own observations or in any sort of reality. It hallucinates constantly. And while it’s good at inventing stuff, there’s a certain repetition and banality to what it comes up with, where you can really see it’s simulating invention rather than creating in the same way that people can. It’s just recombining the stuff that’s fed into it.
Dr Ben Goertzel: So I think these limitations are real. They don’t take away from the incredible utility of the product, but they mean there’s certainly an awful lot further to go between GPT-4 and human-like intelligence, or an artificial general intelligence that’s a real thinking machine.
0:06:09 Beyond GPT-4: adding logic and evolutionary learning
Dr Ben Goertzel: So for those of us who’ve been working on building AI systems, and working toward AGI at the human level and beyond for a long time, a lot of what we’re thinking is: okay, they’ve shown something cool. They’ve shown that if you throw incredible amounts of data and processing power at some relatively basic algorithms, you can get some really interesting functionality. So now, what happens if we throw loads of data and processing power at slightly more sophisticated algorithms with a few more aspects to them, and then just keep amping it up, adding in more and more of the advanced AI methods that we’ve been developing over the years and decades?
Dr Ben Goertzel: Say you take transformer neural nets as were used behind GPT-4. You can tweak that in a number of ways that make a big leap forward, even if they’re not all the way to human-level AGI. You can hybridise a transformer neural net, like a GPT-type system, with a logical reasoning system like we’ve been working on in the OpenCog Hyperon project. And then, lo and behold, you have something that can reason logically and ground its ideas and statements in truth, because logic systems are all about truth.
Dr Ben Goertzel: You can plug in an evolutionary learning system, which uses genetic algorithms, genetic programming, different sorts of AI algorithms which are good at inventing new things that leap further beyond the data that was in there. Now, you need this logic part if you’re going to do the creativity part, because otherwise it’s going to come up with more and more weird, wild stuff.
0:07:57 How SingularityNET’s AI teams gave rise to Zarqa
Dr Ben Goertzel: What we found, though, is this is something that we know how to do in SingularityNET’s AI team, because we have a bunch of people in the SingularityNET project and ecosystem who have been working on AI since forever, on neural nets, logic systems and all the rest. We did find, when we started looking at how to build the next thing beyond ChatGPT, which uses logic and evolution along with transformers and which runs in a democratic and decentralised way on the SingularityNET AI platform, that it’s sort of its own thing. That’s its own project.
Dr Ben Goertzel: And this leads to the company Zarqa, Z-A-R-Q-A, which we’ve put together with a bunch of SingularityNET ecosystem members. We’ve put this together specifically to focus on building the next generation of LLMs after ChatGPT and GPT-4.
0:09:04 Two camps in AI, and coining the term AGI
Brian Rose: Yeah, it’s been interesting, because it seems like there are almost two camps. You’ve got the large language model, narrow AIs that are in this race to make these the ultimate, and then there’s you over here, who have been doing this for, for you, 30 years. And you have been thinking about how we can actually make them smart, make them so they can reason, to where it’s not just regurgitating a big LLM; it’s actually able to make intelligent decisions about that information, and then try to move to that idea of AGI, which is a term you coined, right?
Dr Ben Goertzel: Yeah, I introduced that term. I got my PhD, what, like 35 years ago or something, so I’ve been working on AI and working toward AGI since then, although I introduced the term AGI more recently, like 2004 or 2005. That’s still long before AI became as popular, or as functional, as it is now. And the scale at which we can do these things now is just unprecedented, which is really why we’re able to move so fast.
Dr Ben Goertzel: You still need the right ideas and algorithms, which comes down to the right human brain power. But given the scale at which we can operate, you can bring ideas and algorithms to practical reality and useful functionality so much more quickly and easily than was ever the case before.
Brian Rose: And that’s because of the processing power now, and we know these LLMs can deliver something.
Dr Ben Goertzel: It’s processing power, it’s memory, it’s networking. It’s the whole modern computing and communication infrastructure, and then the huge amount of data that’s provided through the internet.
0:10:43 Is OpenAI already going neurosymbolic behind the scenes?
Brian Rose: Do you think the world is starting to wake up and realise that we can’t just go to ChatGPT 7, 8, 9, 10, 11, 12, and that we need to actually take a step back, do what you’re doing and really think about this neurosymbolic approach, an approach that looks more towards AGI rather than narrow AI? Or do you think everyone’s still caught up in this ChatGPT moment?
Dr Ben Goertzel: I think that people who are deep into AI R&D fully understand that a transformer neural net with some reinforcement learning layered on top is going to hit a wall pretty soon. And if you look at what OpenAI is doing, they’re integrating Wolfram Alpha, they’re integrating Microsoft’s knowledge graph. So they’re already doing a bunch of things that you would classify as neurosymbolic, just quietly behind the scenes.
Dr Ben Goertzel: The thing is, GPT-4, 5, 6, 7 has become sort of a brand name now. It doesn’t mean that it’s really just a GPT algorithm in the technical sense; they can add a lot of other components to it. But I think once you get into doing that, the competitive advantage doesn’t necessarily go to the people who first came out with the GPT model. It could go to people who have deep understanding of these other AI modules that are to be combined with the transformer neural network, which comes down to the neural-symbolic bit, and what Zarqa is doing: building neural-symbolic LLMs that are decentralised, at massive scale.
0:12:18 Why SingularityNET’s team is primed to scale up
Brian Rose: You believe that your team at SingularityNET that you’re putting together to run Zarqa is uniquely qualified to do this. Maybe I’m putting words in your mouth, but it’s almost like this is what you’ve been working towards your entire life, and you have this moment in time right now where, with the funding for Zarqa and what we know about LLMs now, you can take this much further, much faster, than any other group in the world.
Dr Ben Goertzel: Yeah. In SingularityNET we actually have a number of different AI teams in our ecosystem. We’ve had one team working on OpenCog Hyperon, which is a sort of multi-paradigm AI system oriented toward general intelligence. We’ve had a separate team working just on large-scale transformer neural networks, and building supercomputers for large-scale transformer neural networks, and then bit by bit bleeding stuff from the OpenCog team into the transformer neural net side. So we really have been building this as an area of expertise within SingularityNET.
Dr Ben Goertzel: And that group, along with a bunch of my time and others’, is extremely well prepared. That team has been doing pretty much nothing for four years but training transformer neural net models: for natural language, for music, for speech, for audio, for financial data, for biological data. So they’re super primed to take their transformer neural networks and scale them up big time, which is what Zarqa is going to enable.
Brian Rose: Yeah, it’s super exciting. The company is going to be based out of the Middle East, and I know you’re going to be co-CEO with Sergey, one of your other AI scientists. If people want more information, there’s a link below: londonreal.tv/zarqa, Z-A-R-Q-A.
0:14:12 Zarqa’s next leap beyond ChatGPT
Brian Rose: Your final thoughts: are you excited about this project and what’s going to happen? Are you excited to show the world what you believe is the next phase of AI?
Dr Ben Goertzel: Yeah. I think there are going to be a few steps to get from where we are now in AI to human-level general intelligence, and then of course after human-level general intelligence there’s a bunch of other steps: you’re into superintelligence and the Singularity. But I think Zarqa is poised to take the next significant leap beyond systems like ChatGPT and GPT-4, via infusing large language models with logical reasoning through a neurosymbolic architecture, and other elements like evolutionary learning. And I’m really psyched to be doing this on decentralised infrastructure, which promotes a democratic and broadly beneficial methodology for rolling out all this fantastic AI technology.
0:15:17 Why AGI has to be open and decentralised
Brian Rose: That’s one of the points maybe we’ll finish with, which is really important. It’s in the Zarqa pitch deck as well, where you talk about what you’re ultimately about. SingularityNET has been on the blockchain for four years. It’s always been about democratisation; it’s always been about an open marketplace. It’s something we just don’t see in these AI plays. They’re usually surrounded by large tech giants, they’re usually closed off to the public, they’re usually trained in ways we’re not sure of. Maybe speak on how important it is to keep this open source, to make sure that this technology stays the way we want it to stay.
Dr Ben Goertzel: Yeah. I think if AGI is going to create a beneficial technological singularity, if it’s going to be good for all humanity as it gets smarter and smarter, and equals and then exceeds human-level general intelligence, it can’t be siloed and beholden only to the narrow goals of a very small group of people. Then it’s almost by definition not serving the needs of all humanity. So you want the code to be open, you want the deployment to be decentralised, and you want a very diverse, broad group of people ongoingly participating and giving input to the AGI as it develops.
Dr Ben Goertzel: I think this is the most likely way to get to an AGI that’s really broadly beneficial to human beings. It’s also probably the best way to make sure AGI is rolled out in every country in the world and across all different vertical markets, because no siloed-off corporate initiative can actually do everything. There’s just too much out there in the world.
Dr Ben Goertzel: We do need a concentrated group of experts to bear down and focus on training this next generation of LLMs, absolutely. But rolling this out has got to be done in a way that’s democratic and open to everyone, so that as the system evolves beyond what we initially create, it’s taking in code contributions, it’s talking to software systems put online by other people on the SingularityNET platform, and because of that it’s able to reflect the views and serve the needs of a really wide variety of people.
Dr Ben Goertzel: ChatGPT has a lot of users now, but no one knows what’s going on behind the scenes. It’s feeding users what OpenAI wants to feed them, and it’s filtering out things OpenAI doesn’t want them to see. It’s cool to give that functionality to people. It would be even cooler if there was more of an open, decentralised, participatory aspect to the decision of what this thing is that people are interacting with.
0:18:19 AGI by 2029? Goertzel’s timeline for superintelligence
Brian Rose: Last question: what’s your prediction timeline until we get AGI, and then superhuman intelligence? I know I always ask you this, but your years come down. When you were on the show four years ago, I think it was 2029 and 2045.
Dr Ben Goertzel: I’m sticking with Ray, because he did some curve plotting, right? He’s still with those numbers. He did some reasonable extrapolation arguing that we’ll get to human-level AI by 2029. He then thought we wouldn’t get to massively superhuman AI until 2045, and I never bought that one. My thinking was, once you have a human-level AI, once you have a machine that’s as smart as you and me, then that machine can copy itself over and over. It can introspect on its own code, it can rewrite its own code, it can design new hardware and new factories to build that hardware.
Dr Ben Goertzel: I don’t see why it should be more than a few years between human-level AI and superhuman AI, unless the human-level AI itself chooses to throttle and slow down its own development. 2029, I think, is a fair extrapolation. That’s six years from now. It could be a few years sooner or a few years longer; there are some confidence intervals there.
Brian Rose: Okay, but it’s happening within the decade, and then superhuman soon after.
Dr Ben Goertzel: I would say so. We’re living in interesting times.
Brian Rose: We’re working out, eating healthy. We’ve got to live long enough for the Singularity, right?
Dr Ben Goertzel: Yeah, that’s true.
0:19:45 Should people be afraid of AI?
Brian Rose: All right, last question. For the people out there who are worried about AI, having sleepless nights, not knowing if they’re going to be needed or going to be around: how do you feel about it, and what do you say about that?
Dr Ben Goertzel: We are facing radically new things, and we can’t fully predict what’s going to happen. It’s not the first time this has happened in human history, going back to transitioning from hunter-gatherers to agriculture, to the Industrial Revolution, the computer revolution, air flight. Humanity has been inventing new things that project us into new domains that we don’t understand all along, which is pretty much why we’re not cave people anymore. So there is an irreducible uncertainty about this; we can’t know exactly what’s going to happen.
Dr Ben Goertzel: That said, I think there’s no reason to assume a bad outcome is somehow likely. We’re not creating AI by looking at the script of The Terminator or other science fiction movies. We’re trying to create AI systems that love and respect people and have compassion for people, and we’re working on AI systems that do things like elder care and education, art and music, and science.
Dr Ben Goertzel: So I think if we architect AIs in a way that’s designed to make them compassionate, and we have these AIs rolled out doing beneficial things for humanity in a way that’s governed by a participatory, decentralised infrastructure, there’s no reason to be pessimistic and think the AI is likely to suddenly go and become the Terminator. I think that’s just thinking by reference to Hollywood movies, which rarely makes much sense.
0:21:45 Closing thoughts
Brian Rose: All right, well said. I have been watching T2 a little bit too much lately, just in preparation. Again, if everybody wants more information about Zarqa, it’s londonreal.tv/zarqa, link below. Super excited for this project, super excited to support you. I think you guys have a unique vision, and a really good vision, and you’ve got an amazing team. So thank you, Ben.
Dr Ben Goertzel: Yeah, thanks for all your support with the project.
Brian Rose: My pleasure, man. All right, thank you. We’ll see you on the next one.
