对话Bitget AI負責人：AI交易可以無限逼近高分，但無法達到100分

This podcast focuses on Bitget's AI trading product layout. Dr. Bill, the head of AI at Bitget, reviewed his transition from traditional AI research and industry experience into the crypto industry. He systematically introduced Bitget's iterative path for AI trading products over the past year: starting from helping users capture market information and organize news and signals, to creating risk profiles and personalized suggestions based on user history, and then attempting to lower the barrier to AI trading through Agent Hub, Telegram formats, and interactive methods similar to Claude Code.

The interview also discussed the boundaries of AI in trading: it has significantly improved the efficiency of information processing and decision-making assistance for average users, but it is still difficult to completely replace top-tier traders. Future competition will focus not just on model capability but also on security systems, cost control, product smoothness, long-term memory systems, and continuous learning from users' actual trading habits. Both parties concluded that the market remains highly complex, and human nature and black swan events still make it difficult for any single system to completely dominate trading.

Dr. Bill's AI Background and Entry into the Crypto Industry

Mao Di: Welcome to this episode of the "Wu Shuo Bu Jia Mi Bo Ke" podcast. Today, our guest is Dr. Bill, head of AI at Bitget. Please introduce yourself first. How did you enter the crypto industry? Also, I'd like to hear about your experience in AI. I've heard everyone call you Dr. Bill; did you come from an AI background?

Dr. Bill: I graduated with my PhD in 2009. My undergraduate, master's, and doctoral studies were all in AI. During my studies, I also visited many companies and research institutes for exchanges and attended many international conferences.

After graduation, I first worked at an overseas research institute for four years, doing R&D in artificial intelligence. Then, I went to a large domestic company for four years, working on search recommendations and natural language processing, and I was responsible for the Natural Language Processing department. Later, I went to an overseas e-commerce company for four years, leading overall AI R&D, and then went to another large enterprise to oversee global marketing algorithm R&D for three years. In total, I've been doing this for sixteen years.

At the beginning of last year, a headhunter contacted me about this opportunity at Bitget. Although I hadn't worked in the crypto industry before, I have always been interested in finance and have traded US and Hong Kong stocks for many years, so I decided to give it a try.

At that time, I wasn't very familiar with Web3. I had some understanding but hadn't actually done related work, so I was a bit nervous before the interview. But the interview went smoothly, and I got the offer. My position is the head of AI at Bitget, and I've been here for over a year now. Overall, this experience has been quite exciting for me. There are new challenges and projects every day. Although the pressure is high, it's also very fulfilling.

For me, the biggest change has been the cognitive impact. I had only heard of Web3 before, without deep involvement, so after joining, I've basically been learning while working on projects, which has been very fulfilling.

Is the Combination of AI and Trading a Gimmick or Already Practical?

Mao Di: Bitget is a platform primarily focused on trading products. How do you view the "AI + Trading" landscape? Is it genuinely entering a feasible stage, or is it still more about market hype? Not just the crypto industry, but almost all industries are embracing AI. Returning to this topic, do you think it's more about practicality now, or is there also an element of jumping on the bandwagon?

Dr. Bill: For Bitget, I believe this is no longer a gimmick but a necessity. Bitget didn't have a dedicated AI team in its first seven years, and algorithm applications were extremely rare. It's only in the last two years that we've started systematic investment. The core reason is that AI has matured enough to genuinely enter trading scenarios, whether for cost reduction and efficiency improvement or for increasing revenue and trading efficiency. It already has practical value.

Trading itself is very complex. Different users have different knowledge, risk preferences, strategies, and operating methods. So, the key isn't "whether to do AI" but "at which layer of the trading chain AI should be applied."

If we're talking about full automation, like fully autonomous driving, I think it's not yet achievable. But as an assistant, providing assistance in stages and layers, it is already very feasible. Regardless of whether Bitget does it or not, other companies are already working on it and reaping benefits.

For example, some traders primarily look at short-term trends and quantitative signals. Previously, they might have needed to monitor many screens and data points. Now, AI is very suitable for integration and judgment assistance. Others make decisions based on news, financial reports, and social media. Much of this work involves information gathering and organization, where AI can significantly improve efficiency.

Going further, users also hope AI won't just help them find information but will provide more specific strategy suggestions, such as position size, direction, leverage, or even preparing the trading button. At a more advanced level, it can even approach an asset management model.

Our assessment is that AI cannot completely replace the most elite professional traders. However, for average users, achieving a 95% replacement of their work is already in the practical stage today.

Bitget's AI Product Evolution: From Information Organization to Trading Assistance

Mao Di: You mean the first layer is already relatively mature, like helping users understand project backgrounds, organize information, and make judgment calls. Is Bitget's current AI product more focused on early-stage decision support, or has it moved towards specific execution?

Dr. Bill: Let's start from last year. One month after I joined, we launched the Agent direction. At that time, Agent was a very new concept, and everyone was exploring. Initially, we made a small attempt called "Meme Hunter" because Meme coins were very popular then, market signals were fast and chaotic, and it was hard for users to catch trading opportunities in time.

We worked on this product for two months. The results were decent, but its capability was relatively singular, mainly focused on capturing Meme-related signals. Later, we upgraded it to GetAgent. The initial goal was to solve the first layer of demand, which is information collection and organization. Since this part is essentially manual labor, fine-tuning the process and models can significantly improve efficiency.

So, initially, we focused on information-side capabilities, including customizing important news sources for the crypto space. We then fed these high-quality sources to the model for analysis, rather than simply letting the model search the entire internet on its own. This approach significantly improved the accuracy of information collection and analysis, leading to high user satisfaction.

But later, users started demanding more, wanting not only information but also decision-making advice. For example, whether to go long or short, how much to buy, or what strategy suits their risk level. So, we began creating user profiles based on their historical trading records, analyzing their risk preferences and trading habits, and then providing more personalized suggestions.

The information layer can be relatively universal, but at the trading layer, differences become huge. Different users facing the same problem might get completely different answers. So, GetAgent gradually moved towards personalized matching, and we did a lot of detailed polishing in this area.

We actually went as far as the execution layer. For example, a user could directly say, "Buy 10U worth of Bitcoin for me," and the system would quickly prepare the trade button. The user could then confirm and place the order. Of course, the instruction had to be clear enough, not too vague.

After this feature launched, people did use it, and trading volume increased. However, we later realized that if we continued to push deeper into "directly placing orders for users," users could easily misunderstand the product, thinking it could make money for them. Once losses occurred, there would be a gap between expectations and reality.

So, we adjusted our direction. Instead of continuing to optimize automated order placement, we refocused on information gathering, aggregation analysis, and personalized provision, making these capabilities more robust.

Then, at the beginning of this year, we launched Agent Hub. Unlike GetAgent, which operates as a Q&A within the app returning lengthy content, Agent Hub is more geared towards advanced users. It supports them in calling underlying capabilities through programs and executing trades via command lines.

This direction gained some attention at the time, but the barrier to entry was still high. Few people can program or execute trades via command line. The vast majority of users are ordinary traders who need simpler, more direct product forms.

So, later, we moved the entry point to Telegram. Users just need to open a link, log into their Bitget account, and they can complete trades using an Agent-like method. The overall experience is smoother.

Mao Di: How do you address security concerns?

Dr. Bill: For security, we implemented sandboxing, four-factor authentication, and independent environments. The core is to ensure user asset security. Also, we try to lower the barrier for average users. Many similar products require users to connect models themselves, manage token costs, and select service plans, which is too complex for most people. We aim to abstract away this underlying complexity, making it easier for users to get started.

Underlying Logic and User Experience Design of Bitget's AI Trading Products

Mao Di: Which large model do you use?

Dr. Bill: We use multiple large models and intelligently allocate tasks based on requirements. The core is to balance cost and performance. Simple tasks shouldn't always use the most expensive model, and complex tasks can't rely solely on cheap models. So, we're more like doing a system-level optimization.

In product design, we aimed to lower the barrier from the start. For example, we offer users a certain amount of free credits initially, and they pay after using them up. This makes it easier to get started. Users don't need to buy tokens or choose models themselves; they can directly use the underlying capabilities we've already polished.

Later, we migrated many capabilities to Telegram, including information acquisition, analysis processing, and some basic trading strategies. The product on Telegram is called GetClaw. This allows users to interact with the system as if they were chatting, making the experience smoother. Previously, when it was in the app, many users couldn't even find the entry point. But on Telegram, the path is more direct.

After streamlining this experience, GetClaw quickly took off. We also ran trading competitions alongside it, offering trial funds and rewards. The core goal was to help users naturally adapt to this Agent-based trading model.

But we always emphasize that no matter how good the tool, trading cannot completely detach from human judgment. Knowing when to enter and when to exit is still crucial. Relying entirely on models is infeasible, and ignoring them is also unwise. What we want is not to replace users but to make the tools good enough while helping users improve their understanding. This is why, from the outset of our AI work, we set a goal: "Empower 100 million users to be on par with Wall Street." Essentially, we want to help them become better traders.

Our goal is actually to make trading simpler and more personalized. For instance, the system can gradually learn your trading habits, risk preferences, and operational style. It condenses the complex analysis process upfront and presents you with a few clear decision options. This way, you feel more informed and confident when operating.

Therefore, the core of this product model has two points: First, long-term memory and personalized adaptation, where the system continuously learns from the user. Second, it needs to be safe and effective, with continuously evolving underlying tools. Over the past year, GetAgent has polished many underlying capabilities, and GetClaw was built on this foundation. Of course, it's not perfect yet, and further iteration is ongoing.

Mao Di: Have you calculated the trading volume generated by AI trades? Approximately how much is it?

Dr. Bill: It's still not much. As a proportion of the company's total trading volume, the portion completely driven by AI is very low. Earning large-scale user trust for "AI-guided trading" itself requires a cultivation process.

Also, this field changes very rapidly. Large models are constantly evolving. Often, you don't need to overhaul the product form; just switching the backend model from an older version to a newer one can significantly improve overall results. This indicates that model capability and the application layer are starting to decouple. When the underlying model upgrades, the user experience improves accordingly.

So, the current state is that front-end applications are iterating rapidly, and back-end models are continuously progressing. The entire ecosystem is changing very quickly. A requirement that once took one or two months can now be launched in days or even a single day.

In this context, what truly matters is not just development capability but the understanding of the business itself, especially trading. Tools and models are evolving, but the ultimate value of the product is determined by your understanding of the scenario.

Bitget AI Product Competitive Advantages and Continuous Improvement Directions

Mao Di: It's not just Bitget; Binance and OKX are also developing AI-related products. Have you seen their skills or products? What advantages do you think Bitget's AI products have compared to other exchanges? Where do you think you will perform better?

Dr. Bill: This is a great question, and we consistently monitor the latest industry developments. In the AI space, all exchanges started from the same baseline, so we see it as a "curve overtaking" opportunity. At the same time, AI is a field requiring massive investment in both talent and capital, destined to be a playing field for a few top exchanges. Bitget's investment in this area is significant.

In fact, since we started GetAgent last year, we've been exploring how to build AI Agents for the crypto space. At that time, there were almost no existing reference points. We could only look at how other fields were doing it while continuously exploring based on our own business. After over a year of work, we have accumulated relatively solid underlying capabilities and formed a set of methods for continuous iteration.

Compared to other exchanges, I believe our advantages are mainly in several areas.

First, iteration experience. From starting AI Agent development last March until now, we've gone through multiple quarters of continuous iteration. This process was painful, often involving starting over from scratch. But precisely because of this, the accumulated experience is deeper. We wouldn't dare claim to be the industry leader, but we are certainly one of the early and more in-depth players.

Second, security. When these Agent products first appeared, many people rushed to try them but later retreated due to security issues. Internally, we have always placed a high priority on security. Even if it affects development efficiency, safety must be ensured first. After continuous polishing over several quarters, we haven't had any major issues with AI trading and AI Agents so far. This is a crucial advantage.

Third, we are relatively fast in following new product forms. Whether it was Agent Hub or later GetClaw, we launched them early. Moreover, we don't just build the product itself; we also design gameplay considering the trading scenario. For example, we've previously tried integrating AI traders with the copy trading system. Users can choose to follow an AI trader based on its performance, which represents a further innovation in the trading scenario.

On the surface, anyone can build such products now with the help of development tools. But the difference in smoothness, stability, and reliability after actual development is substantial. This isn't just about which model you use; it's about how well you integrate model, cost, quality, security, and user experience.

Especially in C-end scenarios, cost control is critical. Without optimization, the costs of such products can easily spiral out of control. So, what we are doing now is not just "which large model to use" but how to make deeper combinations and optimizations of multiple capabilities, controlling costs within a reasonable range while ensuring experience and quality.

To summarize, I think our advantages are threefold: First, we started early, iterated longer, and have deeper accumulated knowledge. Second, our security system is relatively robust. Third, in integrating skills and product capabilities, we have formed a certain methodology and foundation.

Of course, if there is an area for continuous improvement, I think the most important thing is not to keep focusing on competitors but to learn more from users. Ultimately, AI trading isn't about who has more features; it's about who understands users better. What are users' current understanding, habits, and expectations of AI trading? We need to keep researching this.

After all, the ultimate goal for users on a trading platform is to make money. We can't guarantee they will, but we hope to make their trading faster, more convenient, and more comfortable. For instance, the system might ultimately present only a few clear, personalized options, explaining the logic behind them, making it easier for users to judge and make decisions more confidently.

So, this is far from the finish line. Our current focus is to continue making the experience smoother, safer, and more personalized, while also learning from peers and users.

Will AI Trading Be Winner-Take-All? Will Strategies Become Obsolete Quickly?

Mao Di: You just described a relatively ideal "AI + Trading" scenario. Let me ask two more detailed questions.

First, the capabilities of models executing AI trading will inevitably vary. Will a "winner-take-all" situation emerge? For example, those with more capital can buy more powerful models, have more computing power, and be faster, eventually allowing a few to outperform the vast majority and capture all the market profits.

Second, the trading market changes very quickly. A strategy is often only effective for a specific period before being imitated, followed, or even targeted. Does AI trading have this limitation too? Isn't it impossible to maintain a fixed advantage long-term without continuous iteration?

Dr. Bill: These are indeed both very hot topics in the industry.

Let's start with "winner-take-all." I think this is unlikely. Using the stock market as