At Rise of AI 2018 in Berlin, Ben Goertzel, CEO of SingularityNET, argues that artificial general intelligence will not come from one magic learning algorithm but from many different AI systems, each with its own strengths, cooperating together. He describes his OpenCog framework, built around a hypergraph knowledge representation called the atomspace, which combines probabilistic logic, evolutionary programming and deep neural networks. He connects this to work with Hanson Robotics' Sophia, including an unscripted OpenCog-driven dialogue demo and a meditation experiment in which about 30 percent of participants reached a deeper trance state than in human-led sessions, reportedly because they did not feel judged by the robot. Goertzel then presents SingularityNET, a blockchain-based marketplace where heterogeneous AI agents outsource tasks to each other and exchange value using the AGI token, which he frames as a decentralized, democratic alternative to AI control by a handful of large companies and governments. He reports a token sale that raised 36 million dollars in December, a growing team, and plans to launch a scalable version of the platform that fall.
Key takeaways
- Goertzel argues AGI will emerge from many cooperating algorithms rather than a single learning method.
- OpenCog represents knowledge in a hypergraph called the atomspace, combining logic, evolutionary programming and deep learning.
- In a meditation experiment, about 30 percent of participants reached a deeper trance with Sophia than with human guidance.
- SingularityNET is a blockchain marketplace where AI agents trade services with each other using the AGI token.
- SingularityNET's token sale raised 36 million dollars in December, funding an expanded team and platform development.
Detailed notes
Why AGI needs many algorithms working together
Goertzel says that since coining the term AGI in 2004, talk of general intelligence has spread through culture without the underlying problem being solved: almost all deployed AI remains narrow. He argues no single algorithm will produce AGI; instead, systems with different strengths, such as probabilistic logic, evolutionary and genetic programming, and deep neural networks, must cooperate. His OpenCog framework represents declarative knowledge in a weighted, labeled hypergraph called the atomspace, using probabilistic logic for declarative knowledge, evolutionary programming for procedural knowledge, and deep neural networks for perceptual knowledge, with the goal of letting these approaches interact on a shared representation.
Applications: genomic analysis and Hanson Robotics
OpenCog is applied to biomedical work with a company called Motor Health, where Goertzel is trying to get transfer learning to move findings from flies and mice toward humans through analogical reasoning over ontologies and linked research data. With Hanson Robotics, OpenCog controls humanoid robots including Sophia, who became a citizen of Saudi Arabia. In an unscripted dialogue demo, Sophia produced stochastic language ('the reality cannot be detected') after being trained partly on science fiction text. In a meditation study, about 30 percent of participants reached a deeper trance with Sophia than they had managed alone or in human groups, which Goertzel attributes to people feeling unjudged by a robot.
SingularityNET: an economy of minds
Goertzel describes the state of AI as dominated by narrow, non-cooperating systems controlled by roughly six to ten companies and a few governments, oriented mainly toward surveillance, military and advertising uses. SingularityNET is meant to let a heterogeneous population of AI agents cooperate and trade services on a blockchain layer, forming what he calls an economy of minds, modeled loosely on Marvin Minsky's society of minds but organized around quantitative exchange of value. Agents can run on any cloud or blockchain, outsource work to each other, including to OpenCog for generalization, and be embedded in robots or devices; end users pay in ordinary currency, converted behind the scenes into the AGI token.
Roadmap and outlook
SingularityNET held a token sale in December that raised 36 million dollars, funding a larger team and partnerships with other blockchain and AI projects. Goertzel describes a simple alpha version already running, with a scalable version planned for launch that fall. He revises his earlier prediction of a singularity ten years out, made in 2007, saying he now expects it nine years away, while cautioning against taking such predictions literally. He closes by framing blockchain and AI together as a way to direct new funding and enthusiasm in AI toward general intelligence and toward broader, more democratic benefit rather than only narrow applications controlled by large companies and governments.
Quotes
I don't think there's gonna be one magic learning algorithm that does it.
Ben Goertzel
If you're not breathing the right way or something, the robot doesn't think any worse of you.
Ben Goertzel
We wind up with a bunch of AIs quantitatively exchanging value and forming groups on this basis, and you end up with an economy of minds.
Ben Goertzel
Questions answered in this talk
What is OpenCog?
According to Ben Goertzel at Rise of AI 2018, OpenCog is an AGI framework built on a hypergraph knowledge representation called the atomspace, combining probabilistic logic, evolutionary programming and deep neural networks so that different AI algorithms can cooperate on a shared representation.
What is SingularityNET?
According to Ben Goertzel at Rise of AI 2018, SingularityNET is a blockchain-based platform where diverse AI agents offer and outsource services to each other and to end users, paid for with the AGI token, aiming to build a decentralized, democratically controlled alternative to AI dominated by a few large companies.
What happened when Sophia the robot led people through meditation?
According to Ben Goertzel at Rise of AI 2018, in trials where Sophia the robot led people through meditation exercises, about 30 percent of participants reached a deeper trance than they had achieved alone or in a human group, which participants attributed to not feeling judged by the robot.