HyperCycle — The Journey to a Fully AI-Capable Blockchain
Also available in Indonesian, Spanish, Portuguese, French, Russian, Chinese
Greetings Singularitarians,
I’m extremely excited that the Token Generation Event for the HyperCycle AI-focused ledgerless blockchain network is coming up in May!
I believe HyperCycle has a strong potential to become the future of blockchain infrastructure — its unique technical design solves some of the core dilemmas at the heart of modern blockchain technology, which have been holding the decentralized tech space from achieving some of its most critical goals, including the goal of providing a scalable decentralized platform for global AI networks.
The Path to HyperCycle
The path that led me to HyperCycle started two and a half decades back. My first attempt to create a decentralized underpinning for global AI networks was in 1999–2000 — the Internet was younger then, but already it was clear that the intersection of massive data, massive processing power and massive user interaction that it provided was going to be transformational for AI… and already it was clear that keeping all this power resonating in the wonderfully open and decentralized vibe of the early Internet was going to be a challenge.
I wrote a book (Creating Internet Intelligence) about the various routes I saw to creating emergent decentralized AI on the Web and thus catalyzing the emergence of a new sort of Global Brain. I even thought a bit about the political aspects of this — in 1995 I had been musing about potentially running for US President on a “Decentralization Party” platform, though fortunately as I got older and slightly less foolish I thought the better of this.
The nitty gritty technical aspect of decentralizing AI, however, turned out to be more problematic upon deeper investigation than I’d initially feared. I schemed a few different ways of making strong encryption work together with distributed processing in an AI context, but it was all extremely sluggish. The overhead of doing things in a secure decentralized way was going to be incredible, it seemed. I started thinking about somehow building secure decentralization into a forked version of the Linux OS, and into Internet protocols. And then the dot-com crash happened, and my attention got diverted into putting food on my family’s table, and figuring out how to make my core AGI designs passibly efficient and scalable which was turning out to be a major challenge even without layering on the requirement of secure decentralization.
(Had I thought in 2000 of starting out a secure decentralized network with decentralized money rather than decentralized AI processing, then, well… then I would have been Satoshi, which, no I’m not! But I was fixated on AI, which is more complex and processor intensive than basic financial transactions, so…)
Fast forward to 2015, and Ethereum pops up with its notion of smart contracts (which I immediately observed are neither smart nor contracts, but are still pretty cool when interpreted correctly as persistent scripts running on a distributed network and leveraging a clever decentralized scheme for script validation and overall network security). For the first time, I saw a fairly clear way to make a decentralized AI network function.
2017: Incipit SingularityNET
SingularityNET provides a basic decentralized framework for multiple AI agents to interact, outsource tasks to each other, rate each others’ reputation, pay each other with AGIX tokens, and collaborate and cooperate in flexible ways. The core platform has been in place for some time, but some important ingredients are still under development including the critical “AI-DSL” formal language that agents will use to communicate their properties and requirements to each other, enabling automated assembly of compound-agents without human intervention. We are now making great strides towards an implementable beta-release of AI-DSL.
In the course of developing SingularityNET, various limitations of the Ethereum network have become clear, and an effort was initiated in late 2021 to begin porting the network to the Cardano blockchain, which has a variety of preferable aspects on the computer science level. At the current moment the Cardano port of the SingularityNET platform is partially complete, and its finalization is pending the release of the full Hydra framework by the core Cardano development team, which will allow the flexible management of state channels in the manner required by SingularityNET’s multi-agent dynamics.
Another aspect that became clear in the course of developing SingularityNET was the need for explicit tokenization of the compute resources underlying the network. SingularityNET’s AI agents interact with each other at the level of software APIs, but they all run on one or another hardware device, and need to somehow pay for these hardware resources. Decentralizing this underlying aspect of AI agent networks requires a network and marketplace for provision and request of hardware resources — which is now supplied by the NuNet project, launched in 2021 as a spinoff of SingularityNET Foundation.
On the tokenomic level, the way these networks are designed to interoperate is: While AGIX is what clients and AI agents pay for AI services provided by agents on the SingularityNET network, the NTX token is what AI agents pay for the compute resources that they utilize to do the processing needed to provide their services.
“When Will SNet Build Its Own Blockchain???”
One thing I knew I did NOT want to do, from the earliest days of SingularityNET, was launch my own blockchain. SingularityNET community members often asked “When are we getting our own Singularity blockchain?” and my answer was usually something like “When are we making our own version of the Linux operating system? When are we building our own chip architecture?”
Well, we’re still sticking with standard Linux distros, as of today. But we’re collaborating with Simuli on building an AGI board, including our own custom OpenCog pattern-matching chip, which will be used to accelerate both the OpenCog Hyperon AGI framework and the smart contracts underlying HyperCycle. And, yes, we’re now creating our own blockchain, aka HyperCycle.
The further I proceeded in experimenting with wrapping AI algorithms and structures up in SingularityNET nodes, the more frustrated I became — because the basic principle I felt motivated to follow, on purely practical grounds, was always “minimize the use of the blockchain.” Because on-chain transactions were just so slow and expensive on Ethereum. The shift to Cardano has potential to meaningfully improve this, but it seemed clear even these improvements were not going to be enough.
To be able to put smaller pieces of AI processing and AI knowledge-structure on-chain, some really radical breakthrough in blockchain scaling would be needed.
But the value of putting AI process and structure more granularly on-chain felt as obvious as ever. Only in this way can we have multiple parties contribute in a secure and decentralized way to creating the atomic-level pieces of a shared AI mind. Which gives us much more potential for interesting, breakthrough decentralized AI emergence than having multiple parties collaborate only by each contributing fairly large, independent components to a shared decentralized AI mind.
Blockchain Doesn’t Need a Ledger
The breakthrough for me was in 2019 when I finally took the time to dig into what my friends Toufi Saliba and Dann Toliver had written about their TODA ledgerless blockchain framework. I had heard them talk about TODA before but I hadn’t fully gotten it, because the concepts were novel and the available materials were frankly not that user-friendly.
The core reliance of Bitcoin and Ethereum on a replicated ledger had always felt like a hack to me… in my early dot-com-boom-era explorations of ideas regarding secure decentralized distributed processing, it was already clear to me that this sort of solution wouldn’t really scale. Somehow you needed the history of transactions in a network to be securely fractionated throughout the whole network, not aggregated in a single place (not even if copies of this “single place” are stored all over the Net in multiple locations). However this fractionation is complex to pull off, and I could see why Bitcoin and Ethereum wanted to start out with replicated ledgers.
The move from replicated ledgers to sharded ledgers also seemed natural to me when it arose, yet had the feeling of Ptolemaic epicycles — adding more and more complexities onto a suboptimal design, bringing it closer to something that makes sense, rather than just starting from a more sensible approach in the first place.
What I saw in TODA was, for the first time, a combination of data structure and protocol that enabled decentralized control and coordination of distributed software processes, without building in any replicated ledger at the core. Instead the transaction history of the network was stored in appropriately encrypted files spread across the various nodes in the network.
Ledgers could be layered on top of TODA, when this was an appropriate way to manage records regarding what was happening in certain subnets of the overall network. But they weren’t a require aspect, wired into the guts of the system. Some major subnetworks of the network could be wholly ledgerless and achieve unprecedented levels of speed and scalability because of this.
So in 2019 I started thinking about how to build “SingularityNET on TODA” — geek-dad-joking that “The Todalarity is Near”….
Digging deeper into TODA particulars, though, I began to realize that to create a really optimized sophisticated usage of TODA for running SingularityNET and NuNet style AI agent communities, would require making a few more inventions.
There would be a role, it seemed, for the “Proof of Reputation” scheme that my long-time AI collaborator Anton Kolonin and I had worked out — in which members of a blockchain network accumulate a reputation score via various means and sources, and the rights to confirm transactions in the network are given according to reputation scoring. TODA had its own consensus mechanisms, which were also great, but layering a Proof of Reputation dynamic on top of them would increase efficiency and simplify things in many use-cases.
Smart Contracts that Are Actually Smart
TODA also lacked a native smart-contracts layer. This was not necessarily a problem for many applications — one can do a lot without smart contracts, leveraging a well-designed system of API calls. I had seen this design pattern before in the NEM blockchain, among other places. However for fully sophisticated deployment of AI agents across a blockchain, the flexibility of smart contracts did have real value, I felt.
My first thought was to enable execution of Plutus smart contracts, from the Cardano blockchain, on the HyperCycle infrastructure. Plutus’s foundation in Haskell made this seem like a feasible if tricky implementation task. This would then render HyperCycle a strange and fancy sort of Cardano sidechain. The Hydra framework for interfacing sidechains and external chains with Cardano mainchain would fit in here beautifully, allowing clean interoperation between Plutus contracts running on HyperCycle and those running on Cardano mainchain.
Toufi, Dann and their colleague Greg Colvin then introduced some wonderful new ideas enabling radical speedup of smart contract execution in most any language, by changing the way network nodes do validation of smart contract integrity:
With their new scheme, every validator in a network doesn’t need to verify that a smart contract is being run correctly, only a judiciously chosen random subset.
While their methods can work with Solidity or any other smart contract language, it seemed to me they could be deployed to maximal effect in a framework like Cardano with its Bitcoin-like UTXO transaction model and pure-functional smart contracts.
The more I studied how to make AI processes run efficiently with more and more of their internal structures and dynamics on-chain, though, the more I started to question whether Plutus — awesome as it is — would be enough. I was thinking about things like making a multi-modular neural net where the different modules are embodied in distinct SingularityNET agents created and owned by different entities; or creating a logical reasoning systems where premises and justifications going into a single reasoning step are contained in different SingularityNET agents created and owned by different entities.
In the end it seemed we really wanted to take an AI-customized programming language — like the MeTTa language we are developing for our OpenCog Hyperon AGI system — and adapt it into a smart-contract language for HyperCycle. The MeTTa interpreter, being written in Rust, could then interoperate tightly with Plutus smart contracts via Rust/Haskell foreign function interfaces.
HyperCycle would then be multichain, supporting a variety of smart contract languages, but with especially strong support for MeTTa contracts running AI processes, and second-best support for Plutus contracts or contracts in other languages that could be tightly connected to the MeTTa interpreter via robust FFIs. This seemed in fact in the spirit of modern blockchain, alongside other recent innovations like the ability to run Solidity contracts against the Cardano mainchain, and run Cardano contracts written in the new language Aiken alongside and interoperating with Plutus contracts.
The final technical ingredient fell into place when I met Lucius Greg Meredith, the mastermind of RChain — a very interesting blockchain, founded around the same time as SingularityNET, and featuring a smart contract language called rholang which leveraged some unique and powerful math to achieve secure, efficient concurrent processing on e.g. multi-GPU servers. After talking to Greg a bit, we realized one could use rholang to create a super-efficient concurrent-processing back-end for MeTTa — with blockchain integration already built-in, including use of tokenomics to manage compute resource utilization in real-time during the course of smart contract execution (a huge plus in a NuNet and SingularityNET context). We also realized that Greg’s rho calculus formalism could play a key role in taking NuNet to the next level, allowing NuNet to more efficiently coordinate the compute hardware usage of MeTTa programs and other AI processes running across different machines.
Decentralizing the Next Wave of AI Innovation
And so now in early 2023 I find myself doing exactly what the 2017–18 version of myself really did NOT want to do — creating a custom blockchain for AI, with all sorts of complex and fascinating new features designed specifically for enabling large populations of AI agents to coordinate effectively across a large pool of globally distributed machines, without need for any central owner or controller. What I didn’t see in 2017–18, however, was what an amazing collection of brilliant minds I would find to collaborate on such a project, bringing in many cases their own long-won ideas and working software systems to the table — and this is of course what makes pulling off something as ambitious as HyperCycle feasible on a reasonable time-frame.
Recruiting Toufi Saliba, co-inventor of TODA and HyperCycle, to take the full-time role of CEO of HyperCycle was majorly exciting to me. And as Toufi came on board in this role in 2022 and began to preach the gospel of HyperCycle across the crypto community and the tech world, one further aspect of the project became clear, more on the economic than technical side. Due to fluctuations in Bitcoin price and the Ethereum shift to Proof of Stake, and the general boom in the AI space, he found a significant number of crypto mining farm operators interested in the project, from a practical as well as conceptual standpoint. They were specifically interested in converting pieces of their crypto mining rigs into “AI rigs”, running HyperCycle nodes that then wrap agents processing AI using SingularityNET, NuNet and other ecosystem tools. He also began to find interest from a number of national governments on setting up new AI server farms using HyperCycle as infrastructure.
The timing could not be better for rolling out new global AI infrastructure frameworks.
The next few years are going to be incredibly interesting times — ChatGPT, which is the conversation topic du jour, is only one surge in a wave of AI innovations we’re going to see in the next few years, leading us through the transition from narrow AI to AGI. It will be highly beneficial if many of these next AI breakthroughs are rolled out in the decentralized ecosystem rather than being controlled by Big Tech or Big Government, and HyperCycle is one key piece of the tech stack needed to make this happen.
For more details about the HyperCycle TGE:
SingularityDAO HyperCycle Launch Announcement
Join the SingularityNET Community
Follow SingularityNET’s social channels to join the conversation and receive the latest news and updates:
