HyperAGI Interview: Building a True AI Agent and Creating an Autonomous Crypto Economy

Article Source : TechFlow

https://www.aicoin.com/en/article/400561

HyperAGI is the first AI rune HYPERAGIAGENT community driven decentralized AI project.

Please introduce the background of the HyperAGI team and project

HyperAGI is the first AI rune HYPER·AGI·AGENT community-driven decentralized AI project. The HyperAGI team has been deeply involved in the field of AI for many years and has accumulated profound experience in Web3 generative AI applications. As early as three years ago, the HyperAGI team used generative AI to generate 2D images and 3D models, built an open world MOSSAI consisting of thousands of AI-generated islands on the blockchain, and proposed the AI-generated non-homogeneous crypto asset standard NFG. However, at that time, AI model training and generation had not yet formed a decentralized solution, and the platform’s own GPU resources alone could not support a large number of users, and no outbreak occurred. As LLM ignited the public’s interest in AI, we launched the HyperAGI decentralized AI application platform, and began testing it on Ethereum and Bitcoin L2 in 2024Q1.

HyperAGI focuses on decentralized AI applications and hopes to foster an autonomous cryptocurrency economy with the ultimate goal of establishing Unconditional Basic Agent Income (UBAI). It inherits the strong security and decentralization of Bitcoin and is enhanced by an innovative Proof of Useful Work (PoUW) consensus mechanism.

Consumer-grade GPU nodes can join the network permissionlessly and mine the native token $HYPT by performing PoUW tasks such as AI inference and 3D rendering.

Users can use a variety of tools to develop proof-of-identity (PoP) AGI agents driven by large language models (LLMs). These agents can be configured as chatbots or 3D/XR entities in the metaverse. AI developers can use or deploy LLM AI microservices on the fly to facilitate the creation of programmable, autonomous on-chain agents.

These programmable agents are able to issue or own cryptocurrency assets and can continue to operate or trade, contributing to a vibrant, autonomous crypto-economy to support the realization of UBAI. Users holding HYPER·AGI·AGENT rune tokens are eligible to create a PoP agent on the Bitcoin Layer 1 chain and may soon be eligible for basic agent benefits.

What is an AI agent? Many AI projects now claim to support agents. What exactly is an agent? How is the HyperAGI agent different from other agents?

AI agents are not a new concept in academia, but the current market promotion has made the concept of agents increasingly confusing. HyperAGI’s agents refer to: LLM-driven embodied agents that can be trained and interact with users in a 3D virtual simulation environment, rather than just LLM-driven chatbots. The agents in HyperAGI can exist both in the virtual digital world and in the real physical world. HyperAGI agents are currently integrating physical robots, such as robot dogs, drones, and humanoid robots. In the future, the corresponding agents can be downloaded to physical robots after completing enhanced training in the virtual 3D world in order to better perform tasks.

In addition, all rights of HyperAGI agents belong entirely to users, which has socioeconomic significance. PoP agents representing users can obtain UBAI to adjust basic agent income. HyperAGI agents are divided into PoP (Proof of Personhood) agents representing individual users and ordinary functional agents. PoP can obtain basic income in the form of tokens in HyperAGI’s agent economic system to encourage users to participate in the training and interaction of their own PoP agents, and to accumulate data that can prove human individuals. UBAI also embodies AI equality and democracy.

Is AGI a gimmick or will it soon become a reality? What are the differences and characteristics of the research and development routes of HyperAGI and other AI projects?

Although there is no unified definition of AGI, AGI has been regarded as the holy grail of AI academia and industry for decades. LLM based on Transfer has begun to become the core of various AI agents and AGI, but this is not the case within HyperAGI. LLM does provide novel and convenient information extraction and natural language-based planning and reasoning capabilities, but it is essentially a data-driven deep neural network. A few years ago, during the big data wave, we knew that such a system must be GIGO (Garbage in, garbage out). LLM does not have some of the features necessary for advanced intelligence. For example, at a low level, due to the lack of embodiment of LLM, such AI or agents find it difficult to understand the world model of human users, and it is even more difficult to make plans and take actions for the environment to solve real-world problems. Looking up, LLM has no signs of advanced intellectual activities such as self-awareness, reflection, and introspection.

Our founder Landon Wang has in-depth and long-term research in the field of AI. In 2004, he proposed the aspect-oriented artificial neural network AOAI (Aspect-Oriented AI), which is an innovation that combines neuro-inspired computing with AOP aspect-oriented programming. Aspect refers to an encapsulation of multiple object relationships or constraints. For example, a neuron is an encapsulation (aspect) of relationships or constraints with multiple other cells. Specifically, a neuron senses or controls sensory cells or motor cells through fibers and synapses extending from the neuron cell body, so a neuron is an aspect that contains this relationship and logic. Even each AI agent solves a problem in a certain aspect, and technically it can be modeled with an aspect.

In the software implementation of artificial neural networks, neurons or layers in artificial neural networks are generally modeled as objects, which is easy to understand and maintain in object-oriented programming languages, but the topological structure of the neural network is difficult to adjust, and the activation sequence of neurons is relatively rigid. Although it has shown great power in completing simple and high-intensity calculations, such as LLM training and reasoning, its performance in flexibility and adaptability is unsatisfactory. On the other hand, neurons or layers in AOAI artificial neural networks are modeled as aspects rather than objects. This kind of neural network architecture is highly adaptive and flexible, making the self-evolution of neural networks possible.

HyperAGI combines the efficient LLM and the evolvable AOAI, forming a system that has both the high efficiency of traditional artificial neural networks and the self-evolutionary properties of AO neural networks. It is also the feasible path for AGI that we have seen so far.

What is the vision of HyperAGI?

HyperAGI’s vision is to achieve Unconditional Basic Agent Income (UBAI), build a future where technology equally serves everyone, break the cycle of exploitation, and create a truly decentralized and fair digital society. Unlike some other blockchain projects that only advertise their commitment to UBI, HyperAGI’s UBAI has a clear path to realization, that is, through the intelligent economy, not just castles in the air. Bitcoin proposed by Satoshi Nakamoto is a great innovation of mankind, but it is only a decentralized digital currency with no practical value. The remarkable leap and rise of artificial intelligence has made it possible to create value in a decentralized model. In this model, people benefit from the artificial intelligence running on machines, rather than from the value of others. A true crypto world based on code is emerging, where all machines are created for the benefit and well-being of mankind. In such a crypto world, there may still be a hierarchical structure between artificial intelligence agents, but human exploitation is eliminated because the agents themselves may have some form of autonomy. The ultimate purpose and meaning of artificial intelligence is to serve humanity, as encoded on the blockchain.

What is the relationship between Bitcoin L2 and AI? Why should we build AI on Bitcoin L2?

1. Bitcoin L2 can be used as a means of payment for AI agents

Bitcoin is the medium that embodies the “greatest neutrality” so far, and is very suitable for AI agents engaged in value transactions. Bitcoin can eliminate the inherent inefficiency and “friction” of legal currency. Bitcoin, a “digital native” medium, is the natural soil for AI to exchange value. Bitcoin L2 improves Bitcoin’s programmable performance and can meet the speed required for AI value exchange. Therefore, Bitcoin is expected to become the native currency of AI.

2. Bitcoin L2 can achieve decentralized AI governance

Due to the current trend of AI centralization, the decentralization of AI alignment and governance has attracted much attention. Bitcoin L2’s more powerful smart contracts can become rules for regulating AI agent behavior and protocol models, and realize decentralized AI alignment and governance models. In addition, Bitcoin’s maximum neutrality makes it easier to reach consensus on AI alignment and governance.

3. Bitcoin L2 can issue AI assets

In addition to issuing AI agents as Bitcoin L1 assets, high-performance Bitcoin L2 can meet the needs of AI agents to issue AI assets, which will be the foundation of the agent economy.

4. AI Agents are the Killer Application of Bitcoin and Bitcoin L2

Due to performance reasons, Bitcoin has not been used for anything other than storing value since its inception. Bitcoin entering L2 will have more powerful programmable capabilities. AI agents are generally used to solve real-world problems, so Bitcoin-driven AI agents can really be put to use. And the scale and frequency of use of AI agents will become the killer application of Bitcoin and L2. While the human economy may not prioritize Bitcoin as a payment method, the robot economy may. A large number of AI agents, working for you 24*7, tirelessly use Bitcoin to make and accept tiny payments. The demand for Bitcoin may increase significantly in ways that are currently unimaginable.

5. AI computing can enhance the security of Bitcoin L2

AI computing can be used as a supplement to Bitcoin PoW, or even replace PoW with PoUW. This will subversively inject the energy currently used for Bitcoin mining into AI agents while ensuring safety. AI can make Bitcoin an intelligently driven green blockchain through L2, rather than a PoS mechanism similar to Ethereum. The Hypergraph Consensus we proposed is PoUW based on 3D/AI computing, which will be introduced later.

What is unique about HyperAGI compared to other decentralized AI projects?

The HyperAGI project is unique in the Web3 AI field, with distinct differences in vision, solutions, and technology. From the perspective of solutions, HyperAGI has GPU computing power consensus, AI embodiment, and assetization, and is a decentralized semi-AI and semi-financial application. Not long ago, the academic community proposed five characteristics that a decentralized AI platform should have. We also briefly sorted out and compared existing decentralized AI-related projects based on these five characteristics.

A decentralized AI platform should have five characteristics:

(i) Verifiability of remotely run models

Decentralized verifiability includes Data Avaliability, ZK and other technologies

(ii) usability of publicly available AI models,

Availability mainly depends on whether the API node providing the AI ​​model (mainly LLM) is Peer 2 Peer and whether it is a completely decentralized network.

(iii) Incentivization for AI developers and users,

Is there a fair token generation mechanism in the incentive?

(iv) global governance of essential solutions in the digital society,

Is it easy to reach consensus on whether AI governance is neutral?

(v) No vendor lockins, etc.

Is it a completely decentralized platform?

Using these five characteristics, we can sort out the existing public plans or projects that have been implemented as follows. (Decentralized federated learning has not made significant progress after years of practice. The rise of LLM has made this decentralized training more difficult, so related projects are no longer listed.)

(i) Verifiability of remotely running AI models

We believe that verifiability is an essential feature of decentralized AI projects and the basis for subsequent usability, incentives, governance, and non-binding. Without verifiability, we will not consider other dimensions. Projects without verifiability may be decentralized projects, such as decentralized computing power leasing or data, algorithm, and model markets, but they are not decentralized AI.

Items that may meet the requirements for verifiability include:

Giza, based on the ZKML consensus mechanism, meets the verifiability of remotely running AI models, but its current performance is relatively poor, especially several orders of magnitude away from the requirements of large models. The proof process of a small model with millions of parameters often takes several minutes. It is unacceptable that the proof process of the LLM model takes this amount of time.

Cortex AI is an L1 public chain project launched five years ago that focuses on decentralized AI. It is a relatively complex technology that adds new instructions to the EVM virtual machine to meet the needs of neural network computing. Its underlying layer is still based on ZK availability verification, which can be used for simple AI models, but cannot meet the needs of large models at the LLM level.

Ofelimos, the first PoUW scheme proposed by the cryptography academic community, uses a specific search algorithm. However, the algorithm is not associated with a specific application or project.

Project PAI, mentioned PoUW in a paper but only has a white paper and no product, https://oben.me/.

Qubic, claiming PoUW, proposes to use hundreds of GPUs for artificial neural network calculations, but the significance of simple calculations of the python artificial neural network library is unclear, and it does not seem to meet the needs of LLM training or reasoning, nor does it play the role of PoUW.

FLUX, (PoW ZelHash, not PoUW)

Coinai, (paper stage) https://aipowergrid.io/, task assignment, no strict consensus mechanism

Can not fufill:

GPU computing power leasing projects lack decentralized verifiable mechanisms and cannot guarantee the verifiability of remotely running AI models.

DeepBrain Chain, focusing on GPU leasing, L1 project in 2017, mainnet launch in 2021;

EMC, centralized rewards for task assignment, no decentralized consensus mechanism in the roadmap;

Atheir, no consensus mechanism was seen;

IO.NET, no consensus mechanism;

CLORE.AI, POH, crowdsourcing model, AI model payment on the chain, NFT issuance, AI runs off the chain, no verifiability. The following projects have the same model: SingularityNET, Bittensor, AINN, Fetch.ai, oceanprotocol, algovera.ai.

(ii) Availability of open AI model APIs

Cortex AI, no support for LLM

Qubic, no support for LLM

All of the above-mentioned decentralized AI projects cannot respond well to the above five questions. HyperAGI is a completely decentralized AI protocol based on the Hypergraph PoUW consensus mechanism and the completely decentralized Bitcoin L2 Stack. It will be upgraded to a Bitcoin AI-specific L2 in the future.

PoUW is used to protect the network in the safest way without wasting computing power. All computing power provided by miners can be used for LLM reasoning and cloud rendering services. The vision of PoUW is that computing power can be used to solve various problems submitted to the decentralized network.

Why now?

1. The explosion of LLM and its applications

OpenAI’s ChatGPT reached 100 million users in 3 months, and since then, the development, application and investment boom of language large model LLM has swept the world. But so far, the technology and training of LLM are still carried out in a very centralized way, which has begun to attract great attention from academia, industry and the public. They are worried about the monopoly of AI technology by several major technology providers, data privacy leakage, encroachment and supplier lock-in of cloud computing companies. These problems are essentially because the current Internet and application entrances are still controlled by centralized platforms, and there is no network suitable for large-scale AI applications. The AI ​​community has begun to implement some local operation and decentralized AI projects. The representative of local operation is Ollama, and the representative of decentralization is Petals. Ollama allows small and medium-sized LLM to run on personal computers and even mobile phones by compressing parameters or reducing precision. This protects the privacy and other rights of user data, but it is obviously difficult to support production environments and networked applications. Petals uses Bittorrent’s Peer2Peer technology to achieve completely decentralized reasoning of LLM. However, Petals lacks consensus layer and incentive layer protocols and is still only in the small circle of researchers.

2. LLM-driven Agent

With the support of LLM, intelligent agents can perform upper-level reasoning and have certain planning capabilities. With the help of natural language, multiple intelligent agents can also form social collaboration like humans. Some LLM-driven intelligent agent frameworks have been proposed, such as Microsoft’s AutoGen, Langchain, CrewAI, etc.

Currently, a large number of AI entrepreneurs and developers are focusing on LLM-driven intelligent agents and their applications. There is a huge demand for stable, available, and scalable LLM reasoning, but currently it is mainly completed by renting GPU reasoning instances from cloud computing companies. NVIDIA released ai.nvidia.com, a generative AI microservice platform including LLM, in March 2024 to meet the huge demand, but it has not yet been officially launched. LLM-driven intelligent agents are in full swing, just like the construction of websites in those years, but currently they are mainly collaborating in the traditional Web2 model. Intelligent agent developers need to rent GPUs or purchase APIs from LLM providers to support the operation of these intelligent agents, which creates a lot of friction and is not conducive to the rapid growth of the intelligent agent ecosystem and the value transfer of the intelligent agent economy.

3. Embodied Agent Simulation Environment

Currently, most intelligent agents can only access and operate some APIs, or interact with these APIs through code or scripts, write control instructions generated by LLM, or read external states. General intelligent agents should not only be able to understand and generate natural language, but also understand the human world, and even be able to migrate to robot systems (such as drones, sweepers, humanoid robots, etc.) after corresponding training to complete the specified tasks. Such intelligent agents are called embodied intelligent agents.

The training of embodied intelligence requires a large amount of real-world visual data so that the intelligence can better understand the specific environment and the real world, shorten the training and development time of robots, improve training efficiency, and reduce costs. Currently, these simulation environments for training embodied intelligence are only built and owned by individual companies, such as Microsoft’s Mincraft and NVIDIA’s Issac Gym. There is no decentralized environment to meet the needs of embodied intelligence training. Recently, some game engines have begun to pay attention to artificial intelligence. For example, Epic’s Unreal Engine has begun to promote AI training environments that are compatible with OpenAI GYM.

4. Bitcoin L2 Ecosystem

Although Bitcoin sidechains have existed for many years, they are mainly used for payments, and the lack of smart contracts cannot support complex on-chain applications. The emergence of EVM-compatible Bitcoin L2 allows Bitcoin to support applications such as decentralized AI through L2. Decentralized AI requires a completely decentralized, computing-first blockchain network, rather than being restricted by an increasingly centralized PoS blockchain network. The launch of new protocols for Bitcoin native assets such as inscriptions and runes has made it possible to establish an ecosystem and applications based on Bitcoin. For example, the rune HYPER•AGI•AGENT completed all the mints of the fair launch within an hour. In the future, HyperAGI will issue more AI assets and community-driven applications on Bitcoin.

Let’s talk about HyperAGI’s technical framework and solutions

1. How to realize a decentralized LLM-driven AI agent application platform?

The biggest problem with decentralized AI right now is how to achieve remote reasoning for large AI models, and the lack of high-performance, low-overhead verifiable algorithms for training and reasoning of embodied intelligent agents. Without verifiability, the system can only degenerate into a traditional multi-party market model of supply and demand plus the platform, and a completely decentralized AI application platform cannot be achieved.

Verifiable AI calculations require the PoUW consensus algorithm. Based on this, a decentralized incentive mechanism can be implemented. Specifically, in network incentives, the mint call of tokens is completed by the node after completing the computing task and submitting the verifiable results, rather than transferring the tokens to the node in any centralized way.

To make AI computing verifiable, we first need to define AI computing. AI computing has many levels, such as the lowest-level machine instructions, CUDA instructions, C++, Python language. Similarly, 3D computing required in embodied intelligent body training also has different levels, such as shader language, OpenGL, C++, blueprint scripts, etc.

HyperAGI’s PoUW consensus algorithm is implemented by a computational graph, which is defined as a directed graph in which nodes correspond to mathematical operations. A computational graph is a way to express and evaluate mathematical expressions. It is a “language” for describing equations, containing nodes (variables) and edges (operations (simple functions)).

1.1 Any computation (such as 3D and AI computation) can be verified by defining computation graphs. Different levels of computation can be represented by subgraphs. This can cover various types of computations and express different computation levels through subgraphs. Currently, there are two layers, and the top-level computation graph is deployed on the chain to facilitate verification by verification nodes.

1.2 Load and run LLM models and 3D scene levels in a completely decentralized manner. When a user accesses the LLM model for reasoning or enters the 3D scene for rendering, the HyperAGI agent will start another trusted node to run the same hypergraph (LLM or 3D scene).

1.3 If the verification node finds that the result submitted by a certain node is inconsistent with the result submitted by the trusted node, it will perform a binary search on the off-chain calculation results of the second-layer calculation graph (subgraph) to locate the subgraph calculation node (operator) where the disagreement occurs. The subgraph operator has been pre-deployed to the smart contract with the parameters of the inconsistent operator. The smart contract is called to execute the operator to verify the result.

2. How to avoid excessive computational overhead?

Another challenge in verifiable AI computing is to control the extra computing overhead. We know that the Byzantine consensus protocol requires 2/3 of the nodes to reach a consensus, which means that all nodes need to complete the same calculation for AI reasoning consensus. Such extra overhead is an unacceptable waste in AI computing. HyperAGI only needs 1-m nodes to perform additional calculations to complete verification.

2.1 Each LLM will not perform reasoning alone. The HyperAGI intelligent body will start at least one trusted node for “companion computing”.

Because LLM reasoning calculations are calculated using the calculation results of each layer of the deep neural network in the model and the previous layer as input until the reasoning is finally completed, multiple users can concurrently access the same large LLM model.

Therefore, at most, additional trusted nodes equal to the number of LLMs m need to be started. At least, only one trusted node is needed for “companion computing”.

2.2 3D scene rendering calculations are similar. When each user enters the scene and activates the hypergraph, the HyperAGI agent will load a trusted node based on the hypergraph and perform corresponding hypergraph calculations. If m users enter different 3D scenes, at most m “companion computing” trusted nodes will be started.

In summary, the number of nodes involved in the additional calculation is a random number less than or equal to n+m and greater than or equal to 1. It conforms to the Gaussian distribution. n is the number of users entering the 3D scene, and m is the number of LLMs. This effectively avoids resource waste and ensures network verification efficiency.

How does AI combine with Web3 to form a semi-AI and semi-financial application?

AI developers can deploy agents as smart contracts, which contain data on the top-level hypergraph chain. Users or other agents can call the methods of the agent contract and pay the corresponding tokens. The agent providing the service will inevitably complete the corresponding calculations and submit verifiable results. Users or other agents complete decentralized business interactions with the agent.

The agent will not worry about not receiving tokens after completing the business, and the payer does not have to worry about paying tokens but not getting the correct business calculation results. As for the ability and value of the agent’s business itself, it is determined by the secondary market price and market value of the agent’s assets (including ERC 20, ERC 721 or 1155 NFT).

Of course, the application of HyperAGI is not limited to semi-AI and semi-financial applications, but to realize UBAI, build a future where technology equally serves everyone, break the cycle of exploitation, and create a truly decentralized and fair digital society.