Decentralizing Intelligence: AI Agents and the New Network of Trust

Wang Feng’s Second Dialogue with Tim Gong: On Decentralized AI, AI Agents, and Proof of Intelligence

In Conversation:

• Wang Feng — Founder of Linekong Interactive, MarsBit, and Element
• Tim Gong — Founding Partner of SIG China, Chairman of ByteTrade

Editor’s Note (from the original publication):

On Chinese New Year’s Eve 2023, Wang Feng and Dr. Tim Gong held an in-depth conversation on information ordering, entropy, public blockchains, and the future of Web3.

A year has passed since that dialogue — a year defined by the rise of ChatGPT and large language models (LLMs), which have profoundly reshaped how information is generated and distributed.

How have Dr. Gong’s views evolved since then? And what progress has ByteTrade, the company he leads, made over the past year?

On the eve of Christmas, Wang Feng once again sat down with him for a new conversation.

In June 2022, SIG led an investment in ByteTrade, a Singapore-based foundational software platform for Web3 information applications. Dr. Gong, who holds a Ph.D. in Electrical Engineering (Applied Physics) from Princeton University and a B.S. in Physics from Shanghai Jiao Tong University, now serves as the company’s Chairman. SIG, notably, was one of ByteDance’s earliest and largest investors.

In their previous New Year’s Eve dialogue, Wang and Gong discussed why decentralized information distribution is necessary — what we now call Web3.

Since then, OpenAI’s release of ChatGPT has transformed how information itself is created and consumed.

Over the past year, many of SIG’s portfolio companies in Web3, cloud computing, and AI have quickly adapted, refining their product strategies to capture these new opportunities.

Now, let’s hear how Tim Gong’s thinking has evolved — and where he believes decentralized AI, AI agents, and Proof of Intelligence are leading us next.

Here is their full conversation.

Wang Feng: Many entrepreneurs and investors are now talking about AI-native products and companies. How do you define “AI native”?

Tim Gong: A common definition is that an AI-native product is one that simply wouldn’t work without AI.

For example, tools like Copilot probably don’t qualify. Without AI, Google Search, Microsoft Office, and GitHub Codespaces are still highly useful products — AI merely enhances their performance incrementally.

But AI Agents are different. They allow users to interact entirely through natural language, while the AI itself handles understanding, planning, reasoning, and execution. That, to me, is what AI native really means. An AI Agent is not just a tool — it’s a new kind of species that collaborates with humans.

If you think about it, we’ve moved through three major phases of information evolution: from people seeking information (Google), to information seeking people (ByteDance), and now to personal AI agents that help humans both produce and consume information.

Each of these transitions represents a new way of reducing entropy in the digital world.

Tim Gong: Of course not. As Professor Ming Zeng recently said, “The mainstream model of work in the future will be the collaboration between creative humans and machines.”

Today, the term AI agent is used quite broadly. Any application that equips a large model with knowledge, memory, perception (its “eyes and ears”), and action capability (its “hands”) can be considered an agent.

Agents also include machines that serve as extensions of humans — such as robots powered by large language models, personal IoT devices, or even digital-twin environments.

In fact, nearly all LLM-based startups today are, in one form or another, building agents.

Wang Feng: If AI agents are to become the dominant product form of the future, how will that reshape the entire software ecosystem?

Tim Gong: Professor Ming Zeng once said, “The Web2 software ecosystem was designed to make humans better tools.” I believe the next-generation software ecosystem will instead be built to serve AI agents.

In this new paradigm, humans will only need to interact with their personal agents — all other software will interact with those agents, not directly with people. These “robots” will help you gather information, earn income (through work or trading), learn new skills, and even manage your social life. Your personal agent will become your most trusted and useful companion. You’ll simply talk to it, and it will handle everything else.

Take prompt engineering, for example — or retrieval-augmented generation (RAG), which uses private knowledge bases to enrich an AI’s context. These technologies are infrastructure built specifically to serve AI agents. That’s what it truly means to be AI native at the foundational software level.

As Mistral AI’s founder recently noted, relatively small open-source LLMs — around 7 billion parameters — may hit the “sweet spot” for agent innovation. They’re intelligent enough to be useful, yet lightweight enough for developers to run independently.

Wang Feng: Some people still doubt the open-source model. After OpenAI’s Dev Day, many felt its new products underscored the overwhelming dominance of a tech giant that seemed to rise overnight. With OpenAI’s first-mover advantage being so strong, is the future of AI destined to be centralized?

Tim Gong: Not necessarily. Open-source large language models are iterating at an astonishing pace and becoming increasingly competitive. I checked Hugging Face the other day — there are now thousands of open-source models retrained or fine-tuned from the Llama 2 architecture alone, and their performance gap with OpenAI’s models is narrowing week by week.

Moreover, many of the features showcased at OpenAI Dev Day — from model fine-tuning and RAG-based knowledge retrieval to structured outputs and app orchestration — already have robust open-source equivalents. In fact, you could even argue that at the application layer, OpenAI is now catching up with the innovations coming out of the open-source community.

Wang Feng: But the GPU resources needed for LLM research and inference require a huge investment, which makes the field prone to centralization. Many argue that the gap between “GPU-rich” tech giants and “GPU-poor” startups will only continue to widen.

Tim Gong: I disagree. Take Llama 2 — the most important open-source model today — it was released by Meta, which is certainly “GPU-rich”. Yet other GPU-rich giants like Google, Microsoft, and Amazon haven’t produced anything comparably influential. Clearly, GPUs are not a sufficient condition for innovation. Innovation depends on people, not hardware. The greatest strength of open source lies in its ability to bring people together.

As GPU capacity becomes cheaper and more widely available, the true bottleneck will shift from compute to data — especially private data. That’s where the real competitive frontier will emerge.

Being “GPU-rich” isn’t even a necessary condition for large-model innovation. There’s a massive amount of untapped GPU capacity across personal computers and edge data centers. These may not be ideal for training, but for fine-tuning and inference — which make up more than 95% of the workloads — decentralized GPU resources can be incredibly valuable.

What excites me even more is the progress in running large-model inference on CPUs. Society already has an immense reserve of idle CPU and memory power. Many frontier projects are exploring this today — including our portfolio company Second State, which has managed to run large models offline on personal laptops and even IoT edge devices.

That’s the decentralized AI future I’m most eager to see unfold.

Wang Feng: You’ve made a strong case for the feasibility of decentralized AI agents. But are they really necessary? In your vision, what specific user needs does decentralization address?

Tim Gong: Precisely because AI agents may one day control both the input and output of all our personal information, they must be deeply trusted. We cannot allow them to be controlled by others — nor manipulated by advertisers. This makes it inevitable that agents must be private and decentralized. Both individuals and enterprises will require such decentralized infrastructure to ensure autonomy and trust.

In fact, personal robotic assistants, IoT devices, and digital twins are already user-owned computers — decentralized by nature. At ByteTrade, we call this foundation the “private edge cloud”. It’s an architecture where users retain full ownership of their data, computation, and digital identity.

Yet private agents must still collaborate. Like humans, every agent will need to exchange resources with others — compute power (say, spare GPU time), information, digital assets, or even real-world permissions (for example, a government license that allows your agent to trade certain restricted assets). Each of these forms of exchange opens up entirely new frontiers of opportunity.

Wang Feng: Human collaboration relies on organizational structures. What enables collaboration between humans and machines?

Tim Gong: The foundation of modern civilization is currency — the network that enables value exchange between people. Our intelligent agents will also need a similar network, one that allows them to transact and collaborate both with each other and with humans. As Dr. Fei-Fei Li recently said, “When we think about this technology, we need to put human dignity, human well-being—human jobs—in the center of consideration.” Collaboration between humans and AI agents must always preserve human dignity.

Fortunately, the foundational technology for such a value-exchange network already exists — the decentralized ledger systems born from blockchain. The crypto and Web3 communities have made enormous progress experimenting with peer-to-peer transaction architectures. At ByteTrade, we call the quantifiable and tradable output of agents Proof of Intelligence (PoI) — where intelligence refers broadly to the productive results of human or machine cognitive labor.

Wang Feng: Does that mean everyone in this world will eventually need a decentralized ID (DID)?

Tim Gong: Sam Altman’s WorldCoin introduced the idea of Proof of Personhood. As the founder of OpenAI, he recognized that in an AI-driven world, humans will need to prove their existence in order to participate in the global value network. DID — Decentralized Identity — is simply one technical path toward realizing that vision.

At ByteTrade, our framework of Proof of Intelligence (PoI) extends this idea further. It places both humans and intelligent AI agents within the same network of value exchange. In the beginning, the main scenario will likely involve agents learning a person’s preferences and acting on their behalf in interactions with other agents.

Here are a few examples:

• An agent could serve as your digital twin in a VR world, engaging and transacting with other agents.
• An agent might rent out idle GPU resources on your personal node, in exchange for another agent’s unused storage.
• An agent fine-tuned on your professional expertise could lease its specialized model to other agents who need that domain intelligence.
• An agent holding valuable private data might sell access or provide computation services based on that data.
• An agent could even run a DAO or public staking node, sharing yields with other agents that provide capital.

All these interactions are manifestations of Proof of Intelligence — quantifiable exchanges of cognitive or computational value. On-chain, PoI may appear in different forms: fungible tokens for standardized compute, or NFTs for unique data or algorithms.

The pricing and exchange of these intelligences will be handled through decentralized RFQ markets (like Otomic) or NFT platforms (like Element). Together, they form the foundation for a truly decentralized economy of intelligence.

Wang Feng: Another major force driving the centralization of AI is the government. Few doubt that both China and the U.S. are seeking to regulate large models. Many in the venture community fear regulation will stifle innovation. What’s your view?

Tim Gong: The risks that large models — and eventually AGI — could harm society are real. But the best solutions lie in technological innovation and industry self-governance, not heavy-handed regulation.

For instance, while large models can generate fake news, they can also detect it. Each personal agent could independently assess the authenticity of information, and record its judgment as an NFT on-chain. Suppose Agent A uses Model B and Dataset A to create a realistic short video. Agent A could simultaneously mint an NFT certifying the video’s provenance — so anyone who views it can verify its origin.

If different agents disagree about truthfulness, Proof of Intelligence (PoI) provides a powerful mechanism for the community to reach consensus.

Elon Musk’s Community Notes feature on X, where users collectively rate content, was an early and largely successful experiment in this direction. But as we saw from the recent boardroom drama at OpenAI, voting systems without skin in the game are fragile — they can be manipulated.

With AI agents, content validation can scale globally, and PoI introduces the missing economic layer. It ensures that both agents and their human counterparts bear a real cost when voting — aligning incentives and integrity.

That, to me, is an exciting frontier for the next generation of startups.

Wang Feng: Speaking of startups — has ByteTrade, where you serve as Chairman, already begun building toward this vision?

Tim Gong: Yes. When ByteTrade was founded last year, our goal was to connect individuals’ computing resources into a decentralized “personal cloud”. That vision is essentially the same foundation for AI agents.

Over the past year, as AI has grown dramatically more powerful, the use cases and demand for agents have advanced to an entirely new level. Looking ahead, we’ll be rolling out several product modules step by step next year:

• Terminus OS — our personal cloud operating system, providing a decentralized computing platform where anyone can run open-source large models and AI agents.
• Terminus Core Apps — pre-installed applications for high-security use cases, especially in finance and blockchain, such as digital wallets and decentralized identity (DID) verification.
• Terminus Marketplace — a decentralized app store where ByteTrade and third-party developers can publish AI agents, content recommendation engines, and automated trading bots.
• Otomic — our decentralized RFQ trading network, where bots running inside Terminus quote prices and execute trades automatically. This peer-to-peer RFQ mechanism can handle nearly all crypto and traditional financial assets and derivatives.

In essence, ByteTrade is building the decentralized infrastructure for developing, distributing, and running open-source AI models and agents — while also creating a blockchain-based PoI value-exchange network that enables intelligent collaboration between them.

I’m very excited for what’s to come, and I look forward to deeper conversations on these topics in the year ahead.

Wang Feng: Excellent. Thank you, Dr. Gong, for your time today. We look forward to seeing ByteTrade’s upcoming products.

Translated from the original interview published by MarsBit News.