Wang Feng talks again with Tim Gong: Decentralized AI, AI Agent and PoI

Conversationalists:

Wang Feng: Founder of Blueport Interactive, Initiator of Mars Finance and Elenment

Tim Gong: Founding Partner of SIG China, Chairman of ByteTrade

Editor's Note: On New Year's Eve 2023, Wang Feng had a conversation with Tim Gong, discussing topics such as information sorting, entropy, public chains, and the future of Web3 (link: Wang Feng's New Year's Eve Dialogue with Tim Gong: On Information Sorting, Entropy, and the Future of Web3). A year has passed since this dialogue, during which ChatGPT has risen to prominence, and LLMs have profoundly impacted the generation and distribution of information. What updates have Dr. Tim Gong's insights undergone, and what work has ByteTrade, which he leads, accomplished? On Christmas Eve, Wang Feng once again converses with Tim Gong.

In June 2022, SIG announced a lead investment of $40 million in ByteTrade, a foundational software platform for Web3 information applications based in Singapore, with Tim Gong serving as the company's chairman. Tim Gong graduated from Shanghai Jiao Tong University with a degree in physics and obtained his Ph.D. in electrical engineering from Princeton University. SIG was one of the earliest investors in ByteDance and has remained its largest shareholder.

Last New Year's Eve, Wang Feng and Tim Gong discussed "the necessity of decentralized information distribution," commonly referred to as Web3. Following this, OpenAI released ChatGPT. In the past year, LLMs have had a profound impact on the generation and distribution of information. Many Web3, cloud computing, or AI companies invested by SIG have seized opportunities and adjusted their product directions over the past year. Let’s see what updates Tim Gong has had in this year.

Here is the full dialogue between Wang Feng and Tim Gong:

1. Many entrepreneurs and investors are now discussing AI native products and companies. What is your understanding of AI native?

A common definition might be "products that do not work without AI." For example, products like copilot may not qualify as AI native. After all, Google Search, Microsoft Office, and GitHub Codespaces remain useful products even without AI, and the value provided by AI is incrementally enhancing the experience.

On the other hand, products like AI agents, which require users to interact using natural language, allowing AI to understand, plan, reason, and execute the entire task, are considered AI native. AI agents are not just tools; they represent a new species that collaborates with humans.

We have been inventing new methods to achieve entropy reduction, moving from humans searching for information (represented by Google) to information finding humans (represented by ByteDance), and now to personal AI agents assisting people in producing and consuming information.

2. As a new species, are AI agents meant to replace humans?

Certainly not. I recall a point made by Professor Zeng Ming recently: "The collaboration between creative individuals and machines will be the mainstream working state in the future."

Currently, the market has a broad definition of AI agents. Any application that provides knowledge, memory, perception ("eyes and ears"), and action capabilities ("hands") to large models is considered an agent. Of course, agents also include direct extensions of machines to humans, such as large model-driven robots, personal IoT smart devices, or digital twin environments. Almost 100% of the entrepreneurial companies in the large model application market are essentially working on agents.

3. If AI agents are the primary product form of the future, what impact will this have on the entire software ecosystem?

I remember Professor Zeng Ming saying, "The software ecosystem of Web2 is about making humans better tools." I believe the future software ecosystem will primarily serve AI agents. Because humans will only need to interact with AI agents, other software will have no direct relationship with humans. Agents or "robots" can help you obtain information, assist you in earning money (through work or trading), help you learn, and even aid in socializing. Your personal agent will be your most trusted and useful companion, and you can simply interact with it.

For example, the recently popular prompt engineering in the large model field, including techniques like RAG that use private knowledge bases to supplement prompt context, is software aimed at serving AI agents. This is what AI native looks like at the foundational software level.

The founder of Mistral AI recently mentioned that relatively smaller open-source LLMs, such as 7B parameter models, can allow developers to run them independently while potentially generating sufficient "intelligence," which may be the sweet spot for agent innovation.

4. Speaking of open-source LLMs, some people remain skeptical. The series of products released at OpenAI's dev day demonstrate the absolute advantage of tech giants that rise overnight. With OpenAI's first-mover advantage so strong, is the future of AI centralized?

The iteration speed of open-source large models is increasing rapidly and becoming more competitive. A few days ago, I searched on Hugging Face and found that there are thousands of open-source large models retrained or fine-tuned based on the Llama2 architecture, and their performance gap with OpenAI is continuously narrowing.

Moreover, the series of products released at OpenAI Dev Day, from model fine-tuning, RAG knowledge bases, structured outputs, to application orchestration, already have excellent open-source solutions available. It can even be said that in terms of applications, OpenAI is catching up and imitating innovations from the open-source community.

5. However, the GPU resources required for LLM development and inference require significant investment, making it very easy to centralize. Many people say that the gap between GPU-rich large companies and GPU-poor startups will only widen.

I disagree with this statement. Simply put, the most important open-source large model, Llama2, was released by the GPU-rich Meta, right? Similarly, GPU-rich companies like Google, Microsoft, and Amazon have not released anything impactful to date. Clearly, GPU is not a sufficient condition for innovation. Innovation relies on people, not GPUs. The greatest advantage of open-source is its ability to bring people together. Furthermore, as GPU computing power becomes cheaper, the main contradiction in model training may increasingly be data, especially private data, rather than computing power.

In fact, being GPU-rich is not a necessary condition for large model innovation. There are plenty of redundant GPUs in personal computers and edge data centers. They may not be suitable for training models, but for fine-tuning and inference, which account for 95% of application workloads, these decentralized GPU resources are highly useful.

However, I am more excited about further technological innovations, such as running large model inference on CPUs. There is a significant amount of idle CPU computing power and memory in society today. Many cutting-edge works are being done in this area. Including our portfolio companies, such as Second State, which have achieved offline operation of large models on personal laptops and even IoT edge devices.

I look forward to the future of decentralized AI large model applications.

6. You mentioned the feasibility of decentralized AI agents. But are they necessary? In your vision, what needs can decentralization address for users?

At the same time, because AI agents have the potential to comprehensively master the information entry and exit of each individual, we need to have a high level of trust in them. We cannot allow them to be controlled by others, nor can we tolerate commercial guidance from advertisers. This determines that agents are private and decentralized. Both enterprises and individuals need decentralized infrastructure.

Furthermore, personal robotic assistants, IoT smart devices, or digital twins are essentially computers owned by users, which are inherently decentralized. At ByteTrade, we refer to this infrastructure as "private edge cloud."

However, private agents need to collaborate. Just like humans, each agent needs to exchange resources with other agents. This exchange could be computing power (for example, your agent has idle GPU resources), information, assets, or even permissions in the real world (for example, your agent has a government license to trade a certain restricted asset). These are all new opportunities.

7. Human collaboration relies on organizational relationships. What enables collaboration between humans and machines?

The foundation of modern commercial civilization is currency, which is a network of value exchange between people. Our intelligent agents also need a value exchange network that allows for commercial collaboration between agents and between agents and humans.

Dr. Fei-Fei Li mentioned in a recent interview, "When we think about this technology, we need to put human dignity, human well-being—human jobs—in the center of consideration." The interaction and collaboration between humans and AI agents must maintain human dignity.

Today, we already have the foundational technology for such a network, which is based on blockchain decentralized ledger technology. The entire crypto and Web3 community has conducted extensive attempts and innovations in decentralized peer-to-peer trading systems. At ByteTrade, we refer to this quantifiable and tradable agent contribution as Proof of Intelligence (PoI). This intelligence is broadly defined as "intelligence," resulting from human or machine intellectual labor.

8. Does everyone in this world need to adopt a DID (Decentralized Identity)?

Sam Altman's WorldCoin discusses Proof of Personhood. As the founder of OpenAI, he recognizes that in the future AI world, individuals will need to "self-verify" to join the value network. DID is merely a specific technical means to realize this vision.

ByteTrade's Proof of Intelligence places humans and intelligent AI agents within the same network for value exchange. We believe that the primary scenario here may initially be agents learning human preferences and then representing humans in interactions with other agents. For example:

• An agent could be a user's twin in the VR world, interacting with other agents in the digital realm.
• An agent could sell idle GPU resources on its node in exchange for another agent's idle storage resources.
• An agent might have a fine-tuned large model that performs exceptionally well in a specific field (for example, the human companion of this agent is an industry expert). It can "rent" this model to other agents.
• An agent might possess private data that can help other agents better solve a particular type of problem. It can sell this data or even provide computing services based on this data.
• An agent can run a staking node for a DAO or public chain, sharing profits with agents that increase staking funds.

These exchanges between agents are specific manifestations of PoI. These PoIs may take various forms on the blockchain. For instance, homogeneous computing resources can be fungible tokens, while unique data or algorithms can be NFTs. How to price this intelligence will be determined by decentralized RFQ networks (like Otomic) or NFT trading platforms (like Element).

9. Clearly, another significant force driving the centralization of AI is the government. Whether in China or the U.S., those in the AI industry do not doubt that both governments are trying to "regulate" large models. Many in the venture capital circle say that regulation will impact innovation; I would like to hear your thoughts on this?

I believe that the risks posed by large models, and even AGI, to society do indeed exist. However, the solution should rely on technological innovation and industry self-regulation. For example, while large models can generate fake news, they can also detect fake news. Each of our agents can independently judge the authenticity of information, and the results they produce can also generate NFT records on the blockchain. For instance, if Agent A uses Agent B's model combined with Agent A's data to generate a realistic short video, Agent A will simultaneously publish an NFT to prove the video's origin. This way, anyone who sees the video can trace its source.

If different agents have disputes over the authenticity of information, PoI provides a good mechanism for the community to reach a consensus.

Elon Musk's community notes on X allow users to vote on content, which has been a largely successful attempt. However, from the "palace intrigue" within the OpenAI board, we can also see that voting without "skin in the game" is very dangerous and can easily be exploited.

Using AI agents can scale the voting on the authenticity of content. PoI is an economic mechanism that allows agents and their human proxies to incur costs for voting, thereby giving them "skin in the game." I am very excited about entrepreneurial projects in this direction!

10. Speaking of startups, has ByteTrade, where you serve as chairman, already begun working on these initiatives?

Yes, ByteTrade was established last year with the intention of connecting the computing resources that belong to everyone to build a decentralized "personal cloud." This is not much different from what we are discussing today regarding agents. The main change over the past year is that AI has become more powerful, so the application scenarios and demands for AI agents have escalated. For ByteTrade, we will gradually announce several product modules next year.

• Terminus OS is our personal cloud product. It provides a decentralized computing platform for everyone to run open-source AI large models and agents.
• Terminus will come pre-installed with some core applications, especially those requiring high security, such as financial or blockchain applications. For example, wallets, DID for identity verification, etc.
• Terminus marketplace is a decentralized application marketplace. ByteTrade and third-party developers can publish applications here, such as AI agents, content recommendation engines, automated trading bots, etc.
• Otomic is our RFQ-based trading network. It primarily conducts pricing and automatic execution of trades through robots running within Terminus. This decentralized RFQ mechanism can trade almost all crypto and traditional financial digital assets and derivatives.

On one hand, ByteTrade provides decentralized software development, publishing, and operational infrastructure for open-source large models and AI agents. On the other hand, it enables the collaboration of AI agents by building a PoI value exchange network based on public chains. I look forward to the opportunity to discuss these issues more deeply with everyone next year!

Great! Thank you, Dr. Gong, for your time today. We are very much looking forward to ByteTrade's products!