We talked with Nesa co-founder Harry Yang about the progress of the project and the personal choices he made as an expert in the field of AI in the huge wave of the times.
Interviewer: Harry Yang, Co-founder and CTO of Nesa
Interview and article: Wendy, Foresight News
Now is a golden age for experts in the field of AI (artificial intelligence). AI talents face multiple choices: leaving traditional Internet giants to join large AI companies represented by OpenAI, using Web2 to create new AI application scenarios and operate new startups, or seeking to combine AI with blockchain/Crypto to solve market pain points and create new economic models and ecosystems.
Harry Yang chose a third path.
Nesa, co-founded by him and several other AI talents who graduated from well-known universities such as MIT, Carnegie Mellon University, and Cambridge University, is one of the projects selected for the Binance Labs Season 7 MVB Accelerator Program, and is also one of the star projects emerging in the AI+Crypto track in 2024. On the company's official website, Nesa describes a vision to everyone as "putting AI on the chain for the future."
Harry Yang, who holds a Ph.D. in artificial intelligence from the University of Southern California, has worked in the AI team of Meta for many years. Although he knew "little" about encryption and Web3 before, the more he invested and learned, the more he liked it, because "the concept of decentralized platforms fits well with my vision of future artificial intelligence." This is the biggest feeling of Harry Yang, co-founder and chief technology officer (CTO) of Nesa, since the company was founded.
Foresight News recently spoke with him about the progress of the project and the personal choices he made as an expert in the field of AI in the huge wave of the times.
Foresight News: Chris Dixon, the founding partner of a16z Crypto Fund, recently published a new book about the next generation of the Internet. Bloomberg's comment on the book is very interesting, saying that the problems that Web3, or the so-called "next generation Internet" companies are trying to solve are actually caused by a group of technology giants invested by VCs such as a16z. For example, Meta. What role did your experience working at Meta, a technology giant, play in your entrepreneurial journey of deciding to create a decentralized AI platform?
Harry Yang: Meta is a centralized platform, but they also open-sourced a lot of their models. I am not against centralized platforms. I learned a lot through my personal experience working there. I also got first-hand experience building models like Make-A-Video and Llama. This laid a very good technical foundation for my experience at Nesa.
Regarding centralization and decentralization, I think it is more of a choice than a competition. For some applications, you may prefer a centralized platform. But I believe that in many scenarios, you may prefer to choose a decentralized platform.
I think now is a great time for people to start building and creating more projects independently, because AI really empowers a lot of personal projects and allows you to be an entrepreneur and CEO. In the past, you needed a full team to achieve this, and now the times have changed and the development is getting faster and faster, but I still think there are a lot of applications that we want to see, and we can start to create them now.
Decentralized platforms give everyone the ability and power to do what they want, not just those who work for Meta, OpenAI, or Google. If you don't work for these companies, how can you participate in building the future? I think decentralized platforms give everyone this opportunity.
Foresight News: Now is a golden age for experts in the field of AI. Whether you work for OpenAI or other emerging large AI companies, start a Web2 company focused on AI, or choose Web3+AI to create a new company, there are many opportunities. When the company went public, the founder of Coinbase once said that he missed the golden opportunity of the Internet, but finally found a generational opportunity in blockchain/crypto. Personally, why did you choose this path now? Ten years later, the results of the above different choices may be quite different.
Harry Yang：My background is more in AI. Before founding Nesa, I knew very little about crypto and Web3. Of course, our CEO Patrick Colangelo has a lot of experience (in crypto), and the team also has a lot of experience in developing Web3 applications. This has been a very good learning experience for me, and I have become very passionate about this field in the process. (Entering Web3) does require a learning curve, and I have to spend a lot of time trying to understand different fields. But the more I invest and learn in this area, the more I like it because there is a lot of autonomy and freedom to do things here.
The idea of a decentralized platform fits in very well with my vision for the future of AI. As you said, people can work on different paths, like big companies, small companies, Web2 or Web3. I just think we shouldn't think too much about which path to choose, but if you choose a path, commit to it wholeheartedly and enjoy it.
At the moment, I really like the path I’ve chosen.
Foresight News: So what is the market pain point that Nesa, which you co-founded, is trying to solve? What is the solution?
Harry Yang: What makes Nesa different is that we want to solve the problem of model inference. Many companies are doing this now, such as inference on Llama, Mistral, audio generation, and image generation under stable diffusion. We host various types of models, including images, text, audio, and video, allowing you to perform various types of inference on our platform.
The pain point we want to solve is how to do it easily and at a more affordable price. In terms of how it works, we share all the GPU resources of the community in a distributed way. By optimizing the model to utilize high-end GPUs and CPUs, everyone can contribute and get tokens in return. Our prices are cheaper than other platforms, even AWS or GCP (Note: Amazon and Google's cloud platforms, respectively).
Another pain point (we are trying to solve) is privacy and security, because there are some things you might not want ChatGPT to do. For example, if you want a chatbot that talks like you as your personal assistant, then you need to upload all your chats from WhatsApp or WeChat for ChatGPT to learn from them. But you might not want to do that.
On our platform, all data is encrypted, and all training and reasoning are encrypted. You can trust us with your data and personal information because they will not be visible to anyone. And you can see the whole process through technologies such as TEE (Trusted Execution Environment) and zero-knowledge proof to ensure that everything is safe.
So I think the pain points we want to solve are price, security, and privacy - that's what Nesa is about.
Foresight News: For the general public, can you briefly explain what model inference is, how expensive it is in Web2 scenarios like OpenAI or other big companies, and how you are reducing the price?
Harry Yang: If you run model inference on our platform, you need to pay gas fees. We rent computing resources from providers such as Io.net, GCP or AWS, but we also use computing nodes from personal computers or even Raspberry Pi. These are all provided for free and can be cheaper than cloud service providers.
We crowdsource computing resources. If you sign up (on our platform), you can contribute your laptop or home desktop to this network of computing nodes. In return, you get NES tokens. This is how we keep prices down.
Foresight News: There are many AI+Crypto projects now. You just mentioned some of the aspects that make Nesa unique, but there are also projects on the market that provide similar solutions. So what is your competitive strategy?
Harry Yang: What is unique about Nesa is that in the future our vision is to host private models. Right now our platform hosts models like Llama, OpenAI's ChatGPT, Mistral, StableDiffusion, and DALL-E, but you may also be able to use them elsewhere. While our inference is cheaper and more secure, these models are not limited to our platform.
We are currently developing some models that can be used in applications, such as video generation models, image generation models, personal chatbots, and personalized conversational chatbots developed based on user information. Because our model training and reasoning are secure and private enough, we can develop models that meet personalized needs based on user information. In the future, you will also see more and more proprietary models developed by us on the platform.
In addition, you can run Character AI or DALL-E on Nesa, which is better and more confidential. Another unique thing about us is that it is very easy to use. Once you enter the platform, you will see a page with a list of all the models, and you can just click to chat with Llama or Mistral. It runs very fast and is end-to-end encrypted.
I think this is a unique platform, there is no other like it.
Foresight News: For most models, they tend to require a lot of data, the more the better. Providing data through distributed nodes or individuals may mean that the number of people contributing data to the platform will be relatively limited. So what kind of customers generally build their models or applications on a platform like Nesa?
Harry Yang: Our buyers may be enterprises and financial institutions. We will work with banks and insurance companies. They trust us so much that they will use their own data, such as social security information or personal privacy information that is inconvenient to share with other platforms, and then run the model on our platform to make decisions.
For individual users, we have many killer models on our platform that can be used directly. It can replace ChatGPT, because it is secure and encrypted, so in addition to answering questions and chatting, you can also share your personal information. These models not only understand you better than ChatGPT, but also perform multiple tasks such as image generation, video generation, and chatbots.
Therefore, both enterprises and individuals are users of our services.
Foresight News: You talked about some specific use cases or scenarios, that is, some customers who are very sensitive to data privacy and security. Some of the key technologies used here are also published on your website, one of which is zero-knowledge proof (ZK). Ideally, financial institutions or the healthcare industry can adopt this technology to solve the problems they face in data privacy and security. However, we have written many articles before about the development of ZK technology, and it seems that there are some problems with the large-scale adoption of this technology. So, what do you think about this issue?
Harry Yang: In addition to ZK, there are other solutions (for data security and privacy), such as our use of TEE (Trusted Execution Environment), a mature technology for data and model encryption. Some nodes are equipped with TEE, but because machines with TEE are more expensive, the overhead is also greater. This is one of the solutions to the hardware problem.
For software solutions, as you mentioned, we have zero-knowledge proofs. We have some very good experts on our team in this area. They have come up with solutions on how to use zero-knowledge proofs with large language models. Our technology, using a Llama model like Llama 8B or Llama 3, takes about 40 minutes to get the proof and verify, so it's still relatively slow.
Speed is a pain point for current zero-knowledge solutions, but if you don’t need instant or real-time results, I think that’s acceptable, as it ensures your data and models are secure and private.
Of course, in addition to zero-knowledge proofs, we are also working on a wide range of solutions such as distributed learning, digital watermarking, and homomorphic encryption. All of these combined will make our system more secure, and some of them actually run very fast.
Foresight News: Speaking of hardware, NVIDIA has become a big winner in the current AI wave. Both Web2 and Web3 have generated new demands for chips. Many companies are also competing for hardware and chip resources. Does this bring pressure to you? For you, will decentralized hardware and computing power solutions further increase your pressure, or can they alleviate some of it?
Harry Yang: Indeed, it is our vision to use decentralized computing power to utilize and centralize all GPUs and CPUs. This is because we cannot rely on giants such as Nvidia or AMD. We just need to utilize all other computing nodes outside. Of course, this is just a vision. How to achieve it specifically, I think it will take more time to see whether it can significantly provide a large amount of computing power to the community, or whether there are new challenges.
This may sound a bit centralized, and it will take time to prove. But so far, I think it works pretty well. As more people join, we can scale the system more. Because when we build our own models, we also need a lot of GPUs.
We can get it from GCP or Google, but they have quotas and it's hard to get it even if you pay. So as our platform gets bigger and better, we can just use our own crowdsourced GPUs and not rely on those big companies. That's our plan.
Foresight News: In your opinion, what are the biggest challenges facing Nesa, and the broader AI+Crypto platform?
Harry Yang：I think the technology is mature. We have ChatGPT and various large models. This is why AI is so popular now.
For Nesa, we were one of the first teams to do this, so there were a lot of technical challenges in terms of scaling, load balancing, and how to host a large number of models, and ensure that we could support a large amount of query traffic.
Scaling these queries and models on the platform is a challenge for us. We are working hard to make our system more robust and stable so that it can run 24/7 and handle multiple queries simultaneously. We are working towards this goal.
Foresight News: What is the current stage of Nesa’s development? What can users and the community expect in the short term?
Harry Yang：We have a testnet that is open to a subset of users and whitelisted users. Many people are already testing our platform. The overall feedback is very positive because it is easy to operate and can be registered through the API or Web UI. Users can also call our models or register as miners to contribute their computing power.
There are three main things to look forward to in the short term: the first thing is of course the mainnet that is open to the public and everyone will be able to use it; the second thing is that we are adding more security and privacy features - such as zero-knowledge proofs - to more models so that people can choose to enable or disable it according to their needs; the third thing is that we will build or host more models on top of Nesa. These models should be better and run faster than the existing open source models. Users can come and use our own models for image generation, chatbots, and video generation. These are all things we can expect Nesa to achieve next.
Foresight News: Is there a specific timeline for the mainnet?
Harry Yang: The plan is about a month later.
Foresight News: I noticed some changes in your founding team. A few months ago, there was another partner who also graduated from Cambridge University. Of course, at the same time, I also saw new members joining. For example, Yue Zhang, who is an assistant professor at the University of Southern California. Do you have any information you can reveal about the changes in the founding team?
Harry Yang: We had a member who provided some help in the early stage of the project. I think his departure was more due to personal reasons. He was working in another company at the time and had less time to devote to the project. It was not suitable.
The Nesa team has been working together for a long time. Most of our team members have known each other for years, many of them for as long as 8 years. As a result, our team is very close-knit, which is a big reason why we have been able to grow so fast as a company and a project.
Foresight News: As a key creative member of the team, what belief do you most want to convey to everyone?
Harry Yang: I personally look forward to the future development, not only the development of the Nesa platform, but also the development of AI. Frankly speaking, I feel that my work efficiency has increased 10 times compared to before ChatGPT (appeared).
I have a lot of ideas and a lot of things I want to build, and our team has a lot of things we want to achieve. So, we are recruiting the best talents in AI, Web3, crypto and different fields.
Come join us in doing something really exciting and important.
