Mid November last year, NuNet team has participated in Odyssey Momentum 2020 online hackathon within Nature 2.0 challenge. As a result, the team got the opportunity to work further within Nature 2.0 and Odyssey ecosystem towards a pilot project with other four Nature 2.0 teams and three large Dutch ecosystem providers. However, this post is not about that cooperation — we will inform about its fruits when they come about. Instead, I want to characterize the place we are in our research and development trajectory, perspectives that open from here and our short term plans to harness them. The MVP that we have put together during Odyssey Momentum seems to be a good way to do that.
Results from Odyssey Momentum 2020
During the hackathon, we hacked on a new application running on top of NuNet platform prototype and, using that use case, experimented with the actual blockchain integration at the platform level. We wanted to showcase a radically decentralized application composed of independently developed ML models and executed on decentralized hardware resources, potentially owned by independent individuals, but combined into a single workflow with the help of NuNet APIs and the tokenomic model.
The immediate result of the hachathon was a Minimal Viable Product of a fake-news-detection application, composed of chrome browser extension and backend, powered by SingularityNET infrastructure and NuNet platform prototype (for more about concepts behind NuNet platform and their place in SingularityNET ecosystem see our earlier blogposts here and here). Open source algorithms and ML models which constituted pieces of fake-news-detector application logic were onboarded into SingularityNET infrastructure by NuNet team during the hachathon, but originally developed by third parties during a grassroot event — Fake News Challenge — a few years ago (we used the two winning models).
Here is a short live demo of the fake-news-detection MVP:
While this application is merely a prototype, it showcases several important aspects of NuNet platform:
- decentralized hardware — the ability to dynamically assemble an ad-hoc pool of community sourced computational resources needed for performing a required computational task is he essence of NuNet platform; in this MVP we go a few steps further into that direction.
- decentralized software — one of the premises and goals of SingularityNET infrastructure is the ability to source AI and other algorithms from many independent developers and provide tokenomic compensation for their usage. The real potential of this infrastructure will be revealed when such independently developed pieces of code (which can be referred to as “agents”, “functions” or “actors”, depending on one’s point of view and theoretical inclinations) will be able to assemble into more complex computational workflows — think of a functional program with independently developed functions which are dynamically assembled into an actual program of arbitrary complexity. Furthermore, each developer gets compensated according to the actual usage of their code. Even further, in such architecture, any “agent”, “function” or “actor” of the workflow can be changed into a more efficient, cheaper or better algorithm as soon as one is eveloped (e.g. when the next state-of-the art image recognition or NLP approach comes about). In this direction, our fake-news-detection MVP shows the assemblage of a few independently developed OSS algorithms into the single workflow.
- mapping (decentralized) software w/ hardware—in the world of AI, there exists a tight relation between hardware and software in the sense, that not any software runs, or, at best, efficiently runs, on any hardware (think of ASICs, GPUs, etc.). NuNet is being designed to provide best suited execution environments for the decentralized software assemblages as explained above, so that each “agent”/ “function”/”actor” runs on appropriate, dynamically sourced hardware resource, which also could be independently owned and compensated via a tokenomic mechanism. For NuNet, this fuse of hardware and software means close coordination with SingularityNET’s roadmap and protocols. Providing execution environment for a dynamically scale-able decentralized software architecture as explained above requires some sort of ‘hardware & software mesh’ orchestration environment. With the fake-news-detection MVP we build the basis of such environment¹;
- data exchange — obviously, for a decentralized assemblage of software modules running on decentralized hardware in a single computational workflow to work, there needs to be an efficient way to exchange data between these modules and, furthermore, a way to intelligently place software on geographically or topologically close hardware resources depending on the amount of data exchange between each module. Therefore, along the line of ‘mapping decentralized software w/ hardware’, we have experimented with this in the latest MVP.
- telemetry and ‘computational reflection’ — it comes without explanation that in order to compensate hardware resources for a performed computational work, one needs to estimate the amount of this work in the first place; when orchestrating a dynamic hardware & software mesh, such estimation has to be done automatically, as uniquely required by the execution of each computational workflow. In other words, ‘agents’/’functions’/’actors’ running on the platform have to be computationally reflective — i.e. able to access information about resource usage and properties of hardware on which they execute themselves². This is what NuNet Telemetry API is designed for; with fake-news-detection MVP we have tested and implemented basic principles of the Telemetry API.
- blockhain integration for tokenomic incentives. Last but surely not least, we have moved forward with NuNet Tokenomy API, which takes care of blockchain transactions related to value exchange between software components and hardware components of the architecture — i.e. makes sure that owners of hardware devices that run computational workloads via NuNet network get compensated in tokens. NuNet Tokenomy API has two sides: (1) blockhain side which features a set of smart contracts and (2) a platform side, which features function calls to smart contracts as the tokenomic logic of each computational workflows requires. For the fake-news-detector MVP we have written a set of smart contracts in Solidity for running on Ethereum Ropsten testnet. However, the platform side of Tokenomics API is aimed to be blockchain agnostic — we should be able to work with more than one blockchain, potentially even in a single computational workflow.
Further steps
There’s much to be done. After the hackathon, we have spent few long weeks planning steps needed for advancing the above aspects towards the first alpha version of the NuNet platform. Importantly (but not limited to), these steps include:
- Further developing and rehashing the self-contained and properly abstracted nunet-adapter — the main building block and software element of the NuNet platform. The adapter will be installed on each hardware device of the NuNet network and provide secured communication with other components of the platform, enabling such tasks as deployment of computational workflows, data input and output a and a ‘firewall’ between individual hardware devices and other distributed resources. All NuNet APIs will technically be implemented and exposed to other components of the platform through the nunet-adapter; cannot over-stress the importance of this component;
- Tokenomics API and Telemetry API — the two key aforementioned NuNet APIs and the first ones to be implemented on the nunet-adapter — will be pushed to the next level by making them abstract enough for handling general (i.e. application agnostic) computational workflows. We will also start developing other APIs on a level that is needed for the alpha platform and its exemplar use-case. Note, that all NuNet APIs will be open and eventually subject to modification by community members. The main goal of NuNet is to build the foundation of the platform in a way that enables overall forward-compatibility of the its API of APIs and communication protocols.
- While NuNet relies on blockchain for enabling tokenomic mechanisms and micro-payments between components, it is designed to be blockchain — agnostic and will eventually be able to support cross-chain transactions as well as computational frameworks build to work withdifferent chains. As mentioned, current MVP features ERC-20 tokens and Solidity smart contracts on testnet. Our goal for the first alpha is to add the integration to Cardano Native Assets along the recently launched SingularityNET and IOHK collaboration to explore Cardano/AGI synergies. Note however, that Cardano Multi-Asset token technology is currently in the pre-production stage and is expected to be officially launched within Q1 2021. We are looking forward to close collaboration with IOHK, Cardano developer community (and, of course, continue it with SingularityNET devs) over the next few months. Adding Cardano support will open new horizons for running applications on NuNet and we are eager to explore them.
- For the hardware & software mesh orchestration we will be implementing a protocol- level peer-to-peer communication between decentralized hardware resources over public internet and private subnets behind NAT devices and routers. Along with the nunet-adapter, which will have versions for different operating systems, this aspect of the NuNet platform is device-type dependent. For the first alpha of the platform we will most probably be covering at least PCs and single board computers running Linux and will attempt to onboard other OSes and devices as much as possible;
As described in NuNet’s whitepaper and previous blogposts, our long term development strategy is to cover the roadmap in an iterative use-case based manner. We are carefully selecting real-world use-cases which we use to push the development of the platform while at the same time making these use-cases operational and user-facing from the day one. Our currently selected use-case for the platform development is the fake-news-detector application as described above and we will build the first alpha around it. Stay tuned!
More of NuNet
Notes
[1] The hardware & software mesh aspect of the current fake-news-detection MVP is build using Hashicorp’s open source technologies, including Nomad and Consul.
[2] Computational reflection and other computational and functional aspects of NuNet platform are described in NuNet Whitepaper, p. 13.