Artificial Intelligence researcher Ben Goertzel has spent the past 3 decades working in the field of Artificial Intelligence. He coined the term Artificial General Intelligence (AGI) to differentiate a sentient, cognizant computer from a machine capable of narrowly intelligent behavior in confined areas. Dr. Goertzel has founded the OPenCOG foundation designed to promote open AGI development, and he currently is the Chief Scientist of Hanson Robotics as well as the author of numerous books and Journal articles. During the past several months Goertzel has concentrated his efforts on launching an open marketplace for AI programs and services. This marketplace, called SingularityNET, is designed to offer AI services to customers and to encourage the development of beneficial AGI. Alongside SingularityNET, a cryptocurrency called AGI Coin will be unveiled, with the Initial Coin Offering (ICO)taking place on December 19, 2017. In an interview with Sander Olson for Next Big Future, Goertzel discusses SingularityNET, AGI Coin, and why he believes that toddler-level AI will be achieved by 2030.
Question 1: I first interviewed you for Next Big Future in 2009. How has the AI field evolved during the past 8 years?
In the last 10 years we’ve seen a “narrow AI revolution.” Now AI is used across a huge variety of vertical markets, carrying out all sorts of different user-facing and back-end tasks. For decades AI was important in a few specific markets, e.g. finance and supply chain. But now it’s highly valuable almost across the board.
We’ve also seen an attitudinal revolution — now you can say in polite company that you’re working on building AGI or human-level or superhuman intelligence or whatever, and you are fairly likely to be taken seriously. In 2006 when we started the AGI conference series, AGI was not something you talked about among serious academic or industry researchers, it was something you whispered about in the hallways with your fellow futurist freaks … or discussed at the bar after too many beers.
On the other hand, the basic AI algorithms in play now were all invented in the 50s, 60s and 70s (multilayer neural nets, evolutionary algorithms, logic provers, semantic networks, etc.). Though many of the details of the algorithms have been improved, and the hardware and data situations are much improved. And we still don’t have anything near to a human-level AGI, in terms of practically deployed and demonstrable systems.
Question 2: You have accomplished quite a bit since 2009, including founding the OpenCOG foundation. What is the OpenCOG foundation, and how is it funded?
OpenCog Foundation is a small US nonprofit organization, which holds copyright on a bunch of OpenCog open source software code. So far it hasn’t done a lot else, but in future it may become more active. At the moment most of the work on the OpenCog Artificial General Intelligence framework is funded by other organizations, including nonprofits like Proteus Foundation and Humanity+, and for-profits like Hanson Robotics, Novamente LLC and SingularityNET.
The OpenCog AGI framework is, in my view, the most promising and most fully fleshed out design for AGI at the human level and beyond, that exists on the planet today. It’s an evolving design, with a powerful partial implementation that is getting further advanced by the way, and SingularityNET is likely to lead to a substantial acceleration in OpenCog development progress, via creating software tools that OpenCog systems can leverage, and via entraining OpenCog AI Agents with other sorts of AIs in mutually valuable ways.
Question 3: Although the field of AI is white-hot now, much of the excitement is due to the fact that GPUs/DSPs/ASICs have dramatically increased Deep Learning capabilities. Have there been any important AI advances during the past decade unrelated to Deep Learning?
Many areas of AI have been advancing, indeed. Automated theorem proving can do more and more each year, as you can see from the ever-improving results at the CASC conferences, http://www.cs.miami.edu/~tptp/CASC/ . Evolutionary learning now can learn complex things like control of simulated or physical robots (http://ieeexplore.ieee.org/document/7161918/), due to new algorithms like CMA-ES …. Boolean Satisfaction (SAT) solvers are the tool of choice for model checking (checking correctness of hardware and software) and also are getting better and better at planning and other applications (https://courses.cs.washington.edu/courses/csep573/11wi/lectures/ashish-satsolvers.pdf is an old reference, there are newer…)
Deep NNs have proved especially good at handling perceptual data like images and sounds, so their feats have been easy for the general public to understand, but other AI techniques have also been advancing quite rapidly behind the scenes…
Understanding of human cognitive architecture and its embodiment in AI systems has also been advancing, as seen in OpenCog’s design, but also in other efforts like the SOAR/Sigma/ACT-R community which held a workshop on “A Standard Model of Mind” earlier this year. This sort of research has not yet lead to extremely showy or economically valuable practical results, but I think it will.
Question 4: You are about to issue a new cryptocurrency, called AGI Token. How is this digital currency unique?
The AGI Token is the native currency of the SingularityNET, which is a decentralized autonomous organization of AI Agents. It is specially designed to be highly effective as a tool for AIs to use to pay other AIs for services. This has to do mainly with the design of the overall “token economy” in which the AGI Token lives — for instance in the system of “AGI token rewards” according to which a certain number of AGI tokens is released each month, and given to AI Agents that are decided by the community of AI Agents to be delivering particular value (either delivering great AI services generally, or delivering accurate and impactful ratings of other AI agents, or doing AI that is especially beneficial to the common good)….
Question 5: Many cryptocurrencies use Proof-of-work schemes to verify the validity of the transactions. Is the blockchain verification system employed by AGI Coin sufficient without a Proof-of-Work verification system?
Yes, Proof of Work is somewhat retro by now. Many blockchains and similar technologies, such as NEM for one example, don’t need it.
The current SingularityNET prototype uses Ethereum, which does use PoW though it’s transitioning toward Proof of Stake. However, during early 2018 we will shift SingularityNET to another blockchain or blockchain-like technology that is faster than Ethereum, and in the course we will likely get away from PoW …
Question 6: AGI Coin will be the native currency for the SingularityNET marketplace. How exactly will SingularityNET work?
Like any complex system, SingularityNET can be viewed from multiple perspectives, each one shedding different insights …
Firstly, from a customer view, SingularityNET is simply: an easier, broader and more powerful way for a business to obtain AI services. It is not a consumer service, though it will be used at the back end of multiple consumer services, created by businesses whose software and hardware uses SingularityNET at the back end. SingularityNET will provide businesses with AI services that are superior to those of competing platforms, in terms of the breadth of applications addressed, the quality and cost of results delivered; and, last but not least, the ethical nature of the societal, aesthetic and environmental impact of the organization’s activities. SingularityNET is currently in an early stage of development, with a limited alpha version available now and a scalable version planned for mid-2018. Out of the box it is applicable to any vertical market, but in the early stages the functionality will be richer in certain niches than in others. The areas of initial strength will likely be robotics, biomedical analytics and cybersecurity. The strength of the platform in additional areas will be increased progressively.
Second, from an AI technology point of view, SingularityNET is a heterogeneous and multi-paradigm framework. The AI services delivered by the SingularityNET will include processes based on every known type of AI, including evolutionary learning, probabilistic reasoning, neural nets (deep and shallow), statistical pattern recognition, graph analysis, and many others. A given service provided to a given customer may rely on the back end, on multiple AI processes using different types of AI algorithms. This heterogeneity differentiates SingularityNET from alternate AI services that are more narrowly restricted to particular AI paradigms such as deep neural networks, expert systems or genetic algorithms. This heterogeneity positions SingularityNET particularly well to serve as a breeding ground for the transition from the “narrow AIs” characterizing the AI business world today, to more powerful AGI — AI that is able to learn and generalize beyond the specific contexts, situations and tasks for which it was programmed or trained.
Third, SingularityNET is also innovative from a business structure perspective. Once its initial development phase is completed, it will be a Decentralized Self-Organizing Cooperative — a new type of socioeconomic structure, different from familiar structures like for-profit-organizations, charities, school, clubs and so forth. Specifically it will be a Decentralized Self-Organizing Cooperative with a purpose of providing businesses with superior Artificial Intelligence services, and ongoingly increasing its own intelligence so that it can do so. A DSoC differs from a typical corporation above all in its decentralized and open aspect: anyone can contribute AI code they have developed to the SingularityNET, and the pricing model is intended to support the widest variety of applications possible. Furthermore the governance of a DSoC is fundamentally democratic, so that once the SingularityNET has been launched in its mature phase, it is the developers and users of the network’s AI services who will make the key decisions regarding the network’s coordination and direction.
And fourth, from a broader perspective, SingularityNET is hoped to have an impact beyond that of an ordinary business, or even that of an ordinary unusually-structured socieconomic entity. AI will play a critical role in the next stage of growth of intelligence and life on Earth and in the nearby region of space . By providing a decentralized, open platform for people and AIs to work together on the creation of more powerful, more generally intelligent and more ethical AI services, SingularityNET is intended to increase the odds that the coming radical societal and technological transitions have a positive outcome.
Question 7: What are “smart contracts” and how will they be employed to verify Trust and security?
A smart contract is software code that embodies a contractual agreement between one or more entities, and makes the contract executable as code.
For instance, if Agent A wants to pay Agent B 100 AGI tokens for carrying out task X, a smart contract can be written that puts 100 of A’s AGI tokens in escrow until B has done task X — and then when A verifies that the task is done, automatically releases the tokens to B. In cases where B says the task is done but A does not agree, then the smart contract can also specify that C gets to do the mediation.
In SingularityNET, the economic interactions between two AI Agents are regulated in terms of smart contracts.
In the Ethereum framework, smart contracts are normally represented as procedures (programmed in a language called Solidity) that is then run in a virtual machine. In SingularityNET’s full-scale release version, smart contracts will be specified in OpenCog’s Atomese declarative logic language (where their properties can be logically analyzed and their correctness checked), and then compiled into procedures for running on various virtual machines (including Ethereum’s virtual machine as desired).
Question 8: How will SingularityNET encourage the development of Artificial General Intelligence (AGI)?
SingularityNET will make it easier for early stage AGIs to interface with narrow AIs, which will make it easier for early stage AGIs to participate in doing useful things and in making money. SingularityNET will also provide a framework in which different AI Agents with different levels of generality can cooperate and form composite multi-part meta-AI systems. So it may lead to the emergence of AGIs that are composed of multiple AI Agents originally written by different people (or different automated programming software) for different purposes — where the multiple AIs became “federated” together via the dynamics of the SingularityNET.
Question 9: For individuals more concerned with making money than promoting beneficial AGI, would you still recommend investing in AGI Tokens?
Well, I am an AI researcher, not a financial advisor. And my legal advisors have been pretty strict in telling me not to give financial advice to people involving things like “investment.”
Question 10: In order for SingularityNET to succeed, AI research contributors will need to eschew potentially massive profits coming from selling to corporations, in return for relatively modest compensation from the SingularityNET marketplace. Are you confident that there are many AI researchers willing to make such sacrifices?
I don’t think this is an accurate way to look at it….
The percentage of AI developers who make big money by selling out to big corporations, is quite small. Most AI developers who work for big corporations — and give these corporations their ideas — make a decent salary but don’t get rich.
Furthermore, it’s not clear that financial reward from participation in the SingularityNET marketplace will always be so modest. Some sellers on Amazon and Ebay make a lot of money, and some AI Agent providers on SingularityNET may as well.
Though there will also be a long tail of AI Agent providers in the SingularityNET, offering specialized services and making modest amounts of money thus enabling their creators to earn a living supplying valuable AI for a certain niche market, or working on AGI -oriented code that helps out in many application areas.
Question 11: You have recently claimed that toddler-level AGI could come about by 2030. How confident are you of that prediction?
It’s looking more and more likely every year. I’ll be pretty surprised if we don’t have toddler-level AGI in the range 2023-25, actually. And then it will be more than a toddler in the human sense. It will be a toddler savant, with ability to e.g. do science and math and answer questions way beyond the level of a human toddler. And once we’re there, the full-fledged Singularity won’t be far off in my opinion…. SingularityNET has potential to accelerate R&D progress in this and other AGI-ish directions, making it increasingly likely that the advancement of AI proves Kurzweil a pessimist…