Our goal is not to monopolize the entire machine learning ecosystem, but to establish Gensyn as a protocol for optimizing computing resource utilization, only located above electricity, to significantly enhance human ability to effectively utilize computing resources. -- Ben Fielding, Gensyn Co-Founder
Interview: Sunny and Min, Deep Tide TechFlow
Guest: Ben Fielding, Gensyn Co-Founder
In January 2024, OpenAI CEO Sam stated that the future's two important "currencies" will be computing power and energy.
However, as the power currency of the AI era, computing power is often monopolized by large companies, especially in the AGI large model field. Where there is a monopoly, there is also the power of anti-monopoly, and thus decentralized artificial intelligence (Decentralized AI) has emerged.
"The blockchain, an unlicensed component, can create a market for the buying and selling of computing power (or any other type of digital resource, such as data or algorithms), without the need for intermediaries to conduct transactions globally," a well-known investment institution a16z once described the blockchain path of AI computing power in an article, and the project described is Gensyn.
Gensyn is a decentralized deep learning computing protocol, aiming to become the foundational layer for machine learning computing, which can facilitate task allocation and rewards for machine learning through smart contracts, to quickly achieve the learning ability of AI models and reduce the cost of deep learning training.
Gensyn connects developers (anyone capable of training machine learning models) with solvers (anyone who wants to use their machine to train machine learning models). By utilizing idle computing devices with machine learning capabilities around the world (such as small data centers, personal gaming computers), Gensyn can increase the available computing power for machine learning by 10-100 times.
In summary, the core goal of Gensyn is to achieve AI democratization through blockchain planning.
In June 2023, Gensyn announced the completion of a $43 million Series A financing round, led by a16z, with CoinFund, Canonical Crypto, Protocol Labs, Eden Block, and others participating.
Gensyn was founded in 2020 by senior computer science and machine learning researchers Ben Fielding and Harry Grieve. Harry Grieve studied at Brown University and the University of Aberdeen and is a data scientist and entrepreneur; Ben Fielding graduated from Northumbria University and was the co-founder of the SaaS platform Fair Custodian and a director of Research Analytics.
Deep Tide TechFlow interviewed Gensyn co-founder Ben Fielding to understand his journey in encrypted AI and Gensyn's AI arsenal.
Viewing Gensyn's value proposition from the founder's perspective
TechFlow: What inspired you to establish Gensyn?
Ben:
My initial background was in academia, where I worked as a machine learning researcher, focusing on the field of Neural Architecture Search. This field involves optimizing the structure of deep neural networks, especially for computer vision applications.
My work involved developing algorithms to evolve neural network structures in a collective manner. This process involved training numerous candidate model architectures simultaneously and gradually evolving them into single-element models optimized for specific tasks.
During this time, I encountered significant challenges related to computing resources. As a doctoral student, I had access to several high-performance GPUs, which were housed in large workstations under my desk, which I managed to purchase.
At the same time, companies like Google were conducting similar research, but they were using thousands of GPUs and TPUs in data centers, running continuously for weeks. This difference made me realize that despite having all the necessary resources in addition to ample computing power, people around the world were facing the same limitations, hindering the pace of research and societal progress. I was dissatisfied with this situation, which ultimately led to the creation of Gensyn.
Before fully committing to Gensyn, I spent two years co-founding a data privacy startup. This company focused on managing consumer data flows and consent-based user data access, aiming to improve the interaction between individuals and businesses in terms of data.
This experience taught me valuable lessons, including common pitfalls in entrepreneurship, and reinforced my cautious attitude towards personal data flows and consent-based access.
Four years ago, I closed my startup and joined an accelerator, Entrepreneur First, in London, where I met my partner Harry Grieve. It was there that we launched Gensyn with the aim of addressing the global computing resource challenge. Our initial strategy involved distributing computing tasks across private data islands of individual organizations (federated learning), which was very interesting. We quickly realized the broader potential of expanding this approach globally. To address this expanded vision, we had to tackle the fundamental trust issues associated with computing sources.
Since then, Gensyn has been dedicated to ensuring the accuracy of machine learning tasks processed on devices through a combination of proofs, game theory incentive mechanisms, and probabilistic checks. While the specifics may be quite technical, Gensyn is committed to developing a system that allows anyone in the world to use any computing device to train machine learning models.
TechFlow: Sam Altman needs $7 trillion to operate an AI chip factory to address the global chip shortage issue. Is his plan realistic in terms of scaling chip supply? Meanwhile, what AI problems is Gensyn solving that are different from Altman's solution?
Ben:
Regarding the AI field and the challenges it faces, Gensyn is addressing problems similar to those faced by Altman. Essentially, there are two ways to address the computing access problem. Machine learning is becoming increasingly pervasive and may be integrated into every technology we use, transitioning from imperative code to probabilistic models. These models require a significant amount of computing power. When you compare the computing demand with the world's chip manufacturing capacity, you will notice a significant gap; demand is rapidly increasing while chip production is only growing gradually.
The solution lies in (1) manufacturing more chips to meet demand or (2) improving the efficiency of existing chip usage.
Ultimately, both of these strategies are necessary means to address the continuously growing demand for computing resources.
I believe Altman is effectively addressing this issue. The issue lies in the chip supply chain, which is a very complex system. Certain parts of this supply chain are particularly challenging, and only a few companies have the capability to manage this complexity. Currently, many governments are beginning to view it as a geopolitical issue, investing in domestic repositioning of semiconductor manufacturing and addressing some bottlenecks in the supply chain.
In my view, what Altman has proposed is a $7 trillion test of the market to gauge the global financial market's level of concern about this issue. This astonishing figure has not been directly rejected, which is very noteworthy. It prompts people to rethink, "This sounds absurd, but is it true?"
This reaction indicates a significant concern and a willingness to allocate a large amount of funds to address the issue. By setting such a high benchmark, Altman has effectively created a reference point for any future chip production efforts. This strategic move demonstrates a firm commitment to addressing the challenges of chip manufacturing, even if the actual cost does not reach $7 trillion.
The approach of Gensyn is different; we aim to optimize the use of existing global chips. Many of these chips, from gaming GPUs to Macbooks equipped with M1, M2, and M3 chips, are not being fully utilized.
These devices are fully capable of supporting AI processing without the need for developing new dedicated AI processors. However, utilizing these existing resources requires a protocol to integrate them into a unified network, similar to how TCP/IP facilitates internet communication.
This protocol will enable these devices to be used on-demand for computing tasks.
The main difference between our protocol and traditional open-source protocols like TCP/IP is financial. While the latter is purely a technical solution, using hardware resources itself involves inherent costs, such as electricity and the physical cost of the hardware.
To address this issue, our protocol incorporates cryptocurrency and decentralized principles to establish a value-coordination network to incentivize contributions from hardware owners.
Therefore, Gensyn represents a dual-natured solution: it is both an open-source protocol for software connectivity and a financial mechanism for resource compensation.
Additionally, the challenges faced by the machine learning market are not only related to computing resources.
- Other factors such as data access, knowledge sharing, and more also play a crucial role. Through decentralized technology, we can promote the attribution of value to these different components, driving a more integrated and efficient ecosystem. Therefore, Gensyn does not operate in isolation; we address a part of the broader challenges, but other solutions are also needed to address the remaining issues. This collaborative effort is crucial for advancing the field of machine learning.
Defining the Dual-Natured Solution of Gensyn
TechFlow: Can you explain Gensyn's dual solution in the simplest terms?
Ben:
In simple terms, Gensyn is a peer-to-peer network built on open-source software that allows your device to participate by simply running this software, and your device must be capable of performing machine learning training tasks. This network consists of multiple nodes, each running the software like your device, and they communicate directly, sharing information about available hardware and pending tasks. The benefit of this approach is that there is no need for a central server, and your device can interact directly with other devices, avoiding the need for a central server.
An important feature of Gensyn is that its communication process has no central authority. For example, if you are using a MacBook, it will directly connect and communicate with other MacBooks, exchanging information about hardware capabilities and available tasks.
One of the main challenges Gensyn faces is verifying off-chain non-deterministic computations, which are too large for the blockchain.
Our solution is to introduce a verification mechanism that allows devices to generate verifiable proofs of computation. These proofs can be checked by other devices to ensure the integrity of the work, while not revealing which parts of the task to be verified, preventing devices from only completing parts of the task that may be checked.
Our system encourages devices to participate in the cryptographic proof process as solution providers and verifiers or selectively rerun work to determine the validity of completed tasks. Essentially, Gensyn aims to achieve interoperability between nodes, mutual verification of work, and consensus on completed tasks. Task payments are executed within this framework, leveraging the trust mechanism of the blockchain. This technical ecosystem mimics the functionality of Ethereum, focusing on mutual verification between nodes to ensure the integrity of tasks.
Our primary goal is to achieve consensus on task completion with minimal computational effort, ensuring the integrity of the system while accommodating large-scale machine learning tasks.
In summary, Gensyn can be divided into two main parts.
The first part is the blockchain aspect, including the state machine I mentioned earlier. This is where the shared computation among participants occurs.
The other half of Gensyn involves communication infrastructure, focusing on how nodes interact and handle machine learning tasks.
This setup allows any node to perform any computation, provided it can later verify the work on the blockchain side.
We are building a communication infrastructure that covers all nodes to facilitate information sharing, model segmentation when necessary, and extensive data processing. This setup supports various model training methods, such as data parallelism, model parallelism, and pipeline partitioning, without the immediate need for trust coordination.
Dual-Natured Solution = State Machine + Machine Learning Task Communication
Gensyn State Machine
TechFlow: How does the Gensyn chain function in a specific peer-to-peer machine learning network?
Ben:
Initially, we assume that all participants are fulfilling tasks according to their roles and generating corresponding proofs. Then, we shift our focus to the blockchain aspect, where we maintain a shared state similar to other blockchains, including hashed transactions and operations, as well as the hash of the previous block, forming a complete chain.
The consensus among participants is that if the computation in a block matches and produces the same hash value, the work is considered completed correctly, allowing us to move to the next link in the chain.
Gensyn operates using a POS mechanism, rewarding contributions that verify block generation.
Creating a block involves hashing the operations required for (1) machine learning verification work and (2) recording transactions occurring within that block.
While our approach is similar to systems like Ethereum, our unique contribution primarily lies in the communication aspect, particularly in how nodes manage and collaborate to process machine learning tasks.
TechFlow: How does the Gensyn chain differ from Ethereum? How is a POS chain designed to meet specific use cases for machine learning if the core infrastructure is not novel?
Ben:
Our blockchain's core structure is not novel, except for a novel data availability layer. The significant difference is our ability to handle larger computational tasks, making our operations more efficient than what is typically possible on Ethereum.
This is particularly relevant for convolution operations, which are fundamental components of many machine learning models.
Efficiently executing these operations using Solidity in the Ethereum Virtual Machine (EVM) is challenging.
The Gensyn chain provides more flexibility, allowing us to process these computations more effectively without being limited by the scope of EVM operations.
The real challenge lies in achieving generalizability of models: this means the model can accurately predict the location of entirely new samples, even if they have not been seen before, because it has a broad enough understanding of the space.
This training process requires a significant amount of computing resources as it needs to pass data through the model repeatedly.
The task of Gensyn's machine learning runtime is to acquire a graphical representation of the model and place it within a framework to generate a proof of completion for each operation during computation.
There is an important issue here, which is determinism and reproducibility.
Ideally, in the mathematical world, repeating an operation should produce the same result. However, in the physical world of computing hardware, unpredictable variables may cause slight variations in computation results.
So far, a certain degree of randomness in machine learning is acceptable and even beneficial, as it helps prevent overfitting and promotes better generalization.
However, for Gensyn, both generalizability and reproducibility are crucial.
Changes in computation results may lead to entirely different hash values, which could risk our verification system incorrectly marking work as incomplete, leading to financial losses. To address this, our runtime ensures that operations are deterministic and reproducible across devices, which is a complex but necessary solution.
This approach is somewhat similar to using machine learning frameworks such as PyTorch, TensorFlow, or JAX. Users can define models and initiate training within these frameworks. We are adapting these frameworks and underlying libraries, such as the Compute Unified Device Architecture (CUDA), to ensure that model execution can be reproducible in an accurate manner on any device.
This ensures that the hash processing of operation results on one device produces the same hash on another device, highlighting the importance of machine learning execution in our system.
Gensyn Decentralizes Cloud Services through an Open-Source Blockchain Communication Protocol to Support Decentralized Machine Learning
TechFlow: So, how does this specific blockchain communication infrastructure for machine learning networks function on top of the Gensyn chain?
Ben:
The purpose of the communication infrastructure is to facilitate mutual communication between devices. Its primary function is to allow one device to verify the work and proofs generated by another device.
Essentially, communication between devices is used for mutual verification of work, a process that needs to be done through the blockchain as the blockchain acts as the central arbiter in any dispute. The blockchain is the only trusted source in our system, and without it, there is no reliable way to verify the identity of participants, as anyone could claim they have verified the work.
The blockchain and its encryption technology make identity verification and work confirmation secure. Devices can prove their identity within this mechanism and securely submit information, allowing other parties to recognize and verify the authenticity of this information.
The ultimate goal of this system is to provide compensation to device owners. If you have hardware capable of executing machine learning tasks, you can rent it out.
However, in traditional systems, this process is complex and costly. For example, purchasing a large number of Nvidia GPUs and renting them out—converting capital expenditure into operational expenditure, similar to cloud service providers—involves numerous challenges. You need to find AI companies interested in your hardware, establish sales channels, develop infrastructure for model transfer and access, and manage legal and operational agreements, including service level agreements (SLAs). SLAs require on-site engineers to ensure agreed-upon uptime, with any downtime leading to contractual liabilities and potential financial risks. This complexity is a significant barrier for individuals or small businesses, and it is one of the reasons why centralized cloud services have become mainstream.
Gensyn offers a more efficient approach, eliminating the human and business costs typically involved in these transactions. You only need to run some software, without relying on legal contracts and engineering to build infrastructure. Legal agreements are replaced by smart contracts, work verification is done through automated systems, checking if tasks are completed correctly. There is no longer a need to manually handle breach claims or seek legal resolution; all of this can be resolved immediately through technology, which is a significant advantage. This means that suppliers can immediately earn revenue from their GPUs by simply running some software, without any additional hassle.
Go to Market
We encourage suppliers to join the Gensyn network by informing them that they can immediately enter the demand market for machine learning computations by running open-source software. This is an unprecedented opportunity that significantly expands the market, allowing newcomers to challenge the dominance of traditional services like AWS. AWS and other companies need to manage complex operations, while we are transforming these operations into code, creating new pathways for value flow.
Traditionally, if you have a machine learning model that needs training and are willing to pay for computation, your money will flow to the major cloud providers that monopolize the supply. They dominate the market because of their effective management. Despite increasing competition from Google Cloud, Azure, and other companies, these providers still maintain high profit margins.
The Purpose of Decentralized Cloud Services: Decentralized Training Vs. Decentralized Inference
TechFlow: Machine learning is roughly divided into training and inference. How does Gensyn's P2P computing resources come into play in which part?
Ben:
Our focus is on training, which involves value refinement.
Training includes learning from scratch to fine-tuning, while inference only involves querying the model with data without changing it, essentially seeing what the model predicts based on input.
Training requires a significant amount of computing resources and is usually asynchronous, not requiring immediate results.
In contrast, inference needs to be executed quickly to ensure user satisfaction in real-time applications, which is in stark contrast to the computationally intensive nature of training.
Decentralized technology is currently not sufficient to address the latency issues crucial for inference. In the future, for effective inference, models need to be deployed as close to users as possible, minimizing latency by leveraging geographical proximity.
However, launching such a network is challenging, as its value and efficiency grow with the network's scale, in line with the Metcalfe's law, similar to the development dynamics we see in projects like the Helium network.
Therefore, directly addressing the inference challenge with Gensyn is not realistic; this task is better suited for independent entities focused on optimizing latency and network coverage.
We support protocols specialized in single-function optimization, rather than attempting to develop in multiple areas simultaneously, to avoid diluting efficiency. This specialization drives competition and innovation, leading to a range of interoperable protocols, each mastering specific aspects of the ecosystem.
Ideally, in addition to running Gensyn nodes for computing tasks, users would be able to operate other functional nodes, such as inference, data management, and data labeling. The interconnection of these networks will help build a robust ecosystem where machine learning tasks can seamlessly transfer across various platforms. This decentralized future vision heralds a new network hierarchy, each level enhancing the capabilities of machine learning through collective contributions.
Decentralized AI Ecosystem: How to Win-Win with Decentralized Data Protocols?
TechFlow: Considering that computation and data are important inputs for machine learning, how does Gensyn's computing protocol collaborate with data protocols?
Ben:
Computation is just one aspect; data is another important area that can also apply value flow models, although with different verification and incentive mechanisms.
We envision a rich ecosystem, including multiple nodes running on devices like your MacBook. Your device might have a Gensyn computing node, a data node, and even a data labeling node, contributing to data labeling through gamified incentives or direct payments, often without directly realizing the processes behind these models.
This ecosystem paves the way for what we ambitiously call the machine intelligence revolution, marking a new phase or evolution of the internet. The current internet serves as a vast repository of human knowledge in textual form.
Computing is an important aspect, and data is another key area that can also apply value flow models, although with different verification and incentive mechanisms.
We envision a vibrant ecosystem, including multiple nodes running on devices like a MacBook. Your device may not only run a Gensyn computing node, but also include data nodes and data labeling nodes. Through gamified incentives or direct payments, these nodes will contribute to data labeling, often without users being aware of the complex processes behind them.
This ecosystem paves the way for what we call the machine intelligence revolution, marking a new phase or evolution of the internet. The current internet is a vast repository of human knowledge in textual form.
The future internet we envision is one presented through machine learning models rather than text. This means that fragments of machine learning models will be distributed across devices globally, from MacBooks to iPhones to cloud servers, allowing us to query and infer through this distributed network. This model promises a more open ecosystem compared to centralized models controlled by a few cloud providers, benefiting from blockchain technology.
Blockchain not only facilitates resource sharing but also ensures immediate verification of tasks, verifying the correct and unaltered execution of tasks on remote devices.
Gensyn is committed to developing the computing foundation within this framework and encourages others to explore incentive solutions for data networks. Ideally, Gensyn will seamlessly integrate with these networks, improving the efficiency of machine learning training and applications. Our goal is not to monopolize the entire machine learning ecosystem, but to establish Gensyn as a protocol for optimizing the utilization of computing resources, positioned above electricity, significantly enhancing human capacity to effectively utilize computing resources.
Gensyn specifically addresses the challenge of transforming value and data into model parameters. Essentially, if you have a data sample—whether it's an image, book, text, audio, or video—and you want to convert this data into model parameters, Gensyn facilitates this process. This enables the model to make predictions or inferences about similar future data as the parameters evolve. The entire process of refining data into model parameters is Gensyn's expertise, while other aspects of the machine learning stack are managed by other systems.
Additional Topic: Are AI and Crypto Startups Geographically Limited?
TechFlow: Considering your rich experience, could you compare the differences between dealing with setbacks and challenges in computing and technology as an early builder and researcher in the field, and your current experience? Could you share the impact of this transition and the London tech culture on your development and achievements?
Ben:
The tech environment in London and the UK as a whole is significantly different from Silicon Valley. While the UK's tech community is filled with outstanding talent and pioneering work, it tends to be more insular. This sets up barriers for newcomers trying to integrate into these circles.
I believe this difference stems from the contrasting attitudes between the UK and the US. Americans tend to display a more open demeanor, while the British are generally more skeptical and conservative. This cultural nuance means that integrating and adapting to the UK's tech ecosystem requires effort and time. However, once you do, you'll find a vibrant and rich community engaged in fascinating projects. The difference lies in visibility and outreach; unlike Silicon Valley, where achievements are loudly celebrated, innovators in London tend to work more quietly.
Recently, the UK, especially in its transition towards decentralization and artificial intelligence, seems to be carving out a niche for itself. This is partly due to regulatory developments in the US and Europe. For example, recent US regulations, as outlined in executive orders by President Biden, impose some restrictions on the development of artificial intelligence, including mandatory government reporting for projects exceeding specific thresholds. These regulations may dampen the enthusiasm of new developers. In contrast, the UK seems to take a more open approach, leaning towards open-source rather than strict regulation, fostering an environment more conducive to innovation.
San Francisco is known for its strong open-source movement and is facing new challenges echoing California legislation and federal executive orders. These regulations, while aimed at protecting society, inadvertently concentrate the development of artificial intelligence in established entities. These entities have the capability to comply with regulatory requirements, while potential small players with revolutionary ideas are at a disadvantage. The UK recognizes the value of open-source as a means of societal oversight for the development of artificial intelligence, avoiding the need for restrictive government control. Open-source practices naturally promote scrutiny and collaboration, ensuring that artificial intelligence technology remains under check without stifling innovation.
The initial regulations on artificial intelligence in the EU are stricter than what we see in the UK, and the UK has managed to find a balance that encourages open-source development. This strategy aims not only to achieve the same regulatory goals but also to ensure that the market remains dynamic and competitive. The UK's positioning is particularly favorable for nurturing a vibrant and open ecosystem for artificial intelligence and crypto innovation. This is an exciting time for London's tech industry.
For further reading:
https://docs.gensyn.ai/litepaper
https://epicenter.tv/episodes/471
https://www.techflowpost.com/article/detail_14995.html
https://hyperobjects.substack.com/p/understanding-maximum-computing-supply
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
