The SingularityNet AI platform

What would it mean if there would be an open, automated platform and marketplace for AI services that would:

• Enable researchers and developers to add their AI application to this platform without much extra effort, for testing purposes and to earn money.
• Enable a customer to use any of these services with a single interface
• Enable these AI’s to find and use other AI’s for complex tasks
• Be orchestrated in such a way that the best performing AI services would surface and become prevalent

In my opinion it would only be a matter of time for such a platform to be outrunning any privately-, company- or country driven initiative (like e.g. Google’s TensorFlow), providing that it is properly built (scalable, secure and reliable), properly governed and properly marketed. If such a platform would get enough runway from early adopters, it would be very hard to stop or overtake due to the network effects we know so well from social applications. More than these social networks however, the impact would stretch out way beyond its own boundaries, due to the nature of the AI capabilities it provides.

Image by Gordon Johnson from Pixabay

Welcome to SingularityNet: An ecosystem and marketplace of interacting AI Agents that help people and organisations to create, share, and monetise AI services over a decentralised network.

I’ve been following this initiative for a while now, and it never fails to amaze and excite me. Therefore I decided that I should at least invest the time and effort to read their whitepaper 2.0 and try to understand the initiative in a little more depth.

I am not a developer nor a scientist so I have to admit that I probably understand not more than 20% of the subjects discussed in the document. But I hope that that doesn’t compromise the bigger picture that emerges. Below I will list a few things that became a little bit clearer to me after reading the document as well as some general conceptual observations. I apologize in advance if any of the conclusions are wrong or incomplete. I welcome all (constructive) feedback.

About the founder; Ben Goertzel:

He’s an interesting character. I was surprised to hear that he is the person that coined the term ‘AGI’ or Artificial General Intelligence. First time I read about it was in one of my favorite blogs: https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html. (Well written and very entertaining. In fact; I would recommend to stop and go read this first, if it wasn’t such a long read.) Furthermore he is chief scientist at Hanson robotics (Known from the robot ‘Sophia’), chairman of the OpenCog Foundation and last but not least a quite a hippie. While being a bit eccentric he is an authority on his field and is committed put his knowledge to work to actually make the world a better place.

Technological basis:

I’ll try not to summarize the whitepaper, but I listed some of the technologies used below outlining the general purpose of each as I understand it. Just skip this part if it is too techie.

SingularityNET leverages and aims to extend a lot of research that was conducted in the context of OpenCog.

OpenCog is an open source platform with the audacious goal to develop a framework for Artificial General Intelligence. They are probably a bit behind on their original roadmap, but it means that SN is not starting from scratch. What they borrowed from Open cog is:

• Atom space; A graph-based type of database for AI. As I understand it, it is a kind of in-memory database that relies heavily on relations between its components (the atoms). These atoms hold both data and procedures, making it a kind of a mix between a database and an object oriented programs. Sounds a bit like a MicroService Architecture I would say (but I’m on thin ice here).
• PLN; Probabilistic Logic engine; This is an algorithm that uses logic that is based on on uncertainties. A kind of language to use for constructions such as induction, abduction, analogy, fuzziness and speculation, and reasoning about time and causality.
• Moses; ‘Meta-Optimizing Semantic Evolutionary Search’. In plain English this is a learning algorithm that is based on an evolutionary model: Create generations of code and use a mechanism to select the best results that are in turn the basis for new programs. What makes Moses special is that it incorporates rules and models that make the probability of a ‘good’ result higher, making it more efficient than a blind evolutionary model.
• ECAN; ‘Economic attention allocation’. An self learning algorithm that guides the ‘attention’ of a system by measuring what elements bring most value on short term and long term criteria. The system ensures that the most successful parts of a system get to be found and used more often, while less successful elements will eventually be ‘forgotten’. This algorithm is combined with PLN above to rule out any results that are not useful, making the whole system more efficient.

Some other technologies they are employing are

• SISTER; ‘Symbolic Interactionist Simulation of Trade and Emergent Roles’ A very interesting technology used to simulate social relations. This is a method that lets each participant in a system learn what their optimum role is in the whole. This is done by broadcasting transaction results to other participants. Other participants will in turn use this data to compete with the existing players, making the whole system self organizing and self improving. The interesting thing is that the simulation was successfully applied to both the ecosystem of AI agents in SingularityNET, as well as to real social issues in human societies.
• GHOST; ‘General Holistic Organism Scripting Tool’ A framework for language and reasoning (based on OpenCog), that takes into account not only textual/linguistic data, but also other sensory inputs (vision, sound, etc). One of the ideas behind this is that to really understand language, you first need to have an understanding of the context and the concepts that the language is referring to. GHOST will use input from multiple sensors and aims to extract intents and make inferences about the situation in order to reach specified goals. All in all, natural language is an extremely complex and important subject that SingularityNET researches in cooperation with Hanson Robotics.
• Tononi Phi; This as basically an attempt to measure ‘consciousness’, defined by psychiatrist and neuroscientist Giulio Tononi. While being complex and quite speculative, in practice the theory gives good results when applied to EEG measurements in patients and animals treated with anesthesia. Of course for a system like SingularityNET, being able to measure something like ‘consciousness’ is essential towards the goal of increasing this property.

All this just to give an impression of the direction of their R&D. There are 2 important takeaways from this:

1. SingularityNET doesn’t stop at ‘just’ delivering a marketplace for supply and demand of AI services.
   They are also heavily invested into orchestrating these services with the goal of making the whole ecosystem much larger (and more intelligent?) than the sum of the individual parts.
2. SingularityNET has a wide approach to AI.
   These days AI is almost synonymous with the narrow approach of machine learning using Deep Neural Networks. While popular for good reason, this technology also has clear limitations: There is a need for huge amounts of labeled data for learning purposes. The technique is good in specific use cases like image recognition, but not suited for tasks that require logical reasoning. The SingularityNET ecosystem aims for a much more diversified approach for achieving a higher level of AI, combining ‘traditional’ machine learning techniques with other techniques such as symbolic reasoning, and any other new approach that will be connected to the network

Why on the blockchain

The SingularityNET framework is based on a blockchain that is fuelled by their own ERC-20 utility token called AGI.

The platform is setup in such a way that it is loosely coupled to the current Ethereum chain, minimising the current costs, performance restrictions and technical dependencies related to the Ethereum blockchain. While the AGI token is and will remain the internal transaction token, the network will enable payments in other tokens or fiat money in the (near) future. These payments will however be exchanged to AGI tokens to be used internally.

SingularityNET is not just another case following the blockchain hype. It is the blockchain that enables the core functionality of the platform:

• SingularityNET as a marketplace is a highly transactional system. It has to be able to conduct automated transactions between a myriad of systems (AI’s assigning tasks to other AI’s, assigning tasks to yet other AI’s).
• At the same time the system needs to be able to use the short- and long term transaction history as a source for improvement using the self learning and orchestration mechanisms mentioned above.
• In order to be trusted the system needs to be transparent, and the (automated) transactions must be available for review and analysis (in an audit trail), by humans and machines.
• Last but not least, the platform should be open for any customer, developer or AI, ultimately without any central human governance.

The system will be trained by (multiple, competing) algorithms, based on transactions, logic and human input (voting / staking). The aim is to do this in such a way that useful, efficient and benevolent AI’s will float to the surface and less useful, erratic or fraudulent AI’s will be ‘forgotten’ by the system.

All these conditions make a perfect use case for the blockchain technology to be applied, using transactions and smart contracts. Still, I would not say that SingularityNET is primarily a blockchain project. It is mainly an orchestrated ecosystem of AI’s that is operating a marketplace for AI services. The Blockchain is a an enabling technology to achieve these goals in an efficient and open way.

Without the blockchain, an ecosystem might still be possible, but without the same transparency, trust and automated governance. It is likely that it would become just another silo. Developers that would submit an AI service would not be certain of their intellectual and legal ownership and fair share of revenues. Customer organizations would not be certain who would have access to their data. If such a system would become successful, the knowledge and power of the governing organization would be immense.

Status and next steps

• At the end of 2018 SN announced the legal entity ‘Singularity Studio’ This will be a commercial entity that has the purpose to help enterprise customers onboard the platform by providing consultancy, integrations with backend systems, custom user interfaces, proprietary machine learning models etc. All important things that are not the core business of the SingularityNET foundation. The bigger goal however is to kickstart the utilization of the platform by onboarding large organizations.
• In February 2019 SingularityNET have released the Beta version of the platform that is now fully functional. There is however a lot of work and a bunch interesting features in the pipeline. To name a few:
• Creating SDK’s for Python, JavaScript and others, making the utilisation of the network easier for developers.
• Request for AI portal: A portal that enables users to request specific AI features.
  Recently a good example of what this can mean was published. SingularityNet announced a partnership with Domino’s Pizza’s. The goal of Domino’s is to improve its logistics in pizza delivery in the crowded cities of Singapore and Malaysia. It is interesting that SingularityNET will supply some algorithms to help Domino’s succeed in these goals. Much more interesting however is that Domino’s (which today is more a technology-based organization than a food company) will use the portal to specify what their requirements are for their future AI needs. This way they can tap into a resource of forefront AI developers that would otherwise be hard to reach. And of course this will incentivise independent AI developers to work on concrete cases for SingularityNET, while being paid out in the native AGI tokens that Domino’s has acquired for this purpose. If a few more large organisations like this will start interacting with SingularityNET in the same manner, a snowball effect could take place, leading to a virtuous circle of increasing demand, fuelling increasing supply and so on.
• A Fiat to AGI gateway, making it easy to pay and cash out on the platform
• Add more sophisticated pricing models to the platform

The future

So where will all this lead to? The founders of SingularityNET are aiming no less than for the emergence of a benevolent Artificial General Intelligence. The benevolence would come from the fact that it is not proprietary for any organization or country with their limited goals and that it is democratically governed. (Imagine an Artifical Super Intelligence who’s main goal it is to make people click on ads as often as possible). This is a compelling idea which alone makes the project interesting and admirable. Who wouldn’t want to be involved in a project like that? Whether it is realistic the future will tell. A real artificial intelligence in the wider sense will probably look like us just as much as a bird looks like an F16 plane. But that analogy shows that it might be far superior to us in some aspects. And I guess we do not even need the general intelligence for that. Along the road to this goal there will be plenty of accomplishments that will amaze us in ways we cannot yet foresee.

So I’m looking forward to what this platform will be in 5 years from now and from there on. If the platform will manage to get sufficient traction over the next year(s), the future may be closer than we think and I am optimistic that it is looking very bright for SingularityNET!

If you would like to know more about the wider ecosystem that is forming around SingularityNET, continue to read my next article: Towards an economy of decentralized AI

Or read this one for the business perspective: Is there a winning strategy for smaller players in the Artificial Intelligence arms race?