Dialogue with Bitget AI Lead: AI Trading Can Approach High Scores Infinitely, But Cannot Reach 100 Points

In this episode of the podcast, we delve into Bitget's AI-powered trading product strategy. Dr. Bill, the head of AI at Bitget, recounts his journey from traditional AI research and industry experience into the crypto space. He systematically introduces Bitget's iterative path for AI trading products over the past year: from initially helping users capture market information and organize news and signals, to combining user historical behavior for risk profiling and personalized advice, and then attempting to lower the barrier to using AI trading through methods like Agent Hub, Telegram interfaces, and interaction styles similar to Claude Code.

The interview also discusses the boundaries of AI in trading: it can now significantly enhance information processing and decision-making efficiency for average users, but still struggles to completely replace top-tier traders. Future competition will focus not just on model capabilities, but also on security systems, cost control, product smoothness, long-term memory systems, and continuous learning from users' actual trading habits. Finally, both parties explore whether AI trading will lead to a "winner-takes-all" scenario or if strategies will quickly become obsolete. The conclusion is that the market remains highly complex, and human nature and black swan events will prevent trading from being completely dominated by a single system.

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

Mao Di: Welcome everyone to this episode of "Wu Says Crypto Podcast." Today, our guest is Dr. Bill, the head of AI at Bitget. First, could you please introduce yourself and tell us how you entered the crypto industry? Also, we'd love to hear about your experience in AI. Since everyone calls you Dr. Bill, did you come from an AI background?

Dr. Bill: I received my Ph.D. in 2009, and my major was AI throughout my undergraduate, master's, and doctoral studies. During my studies, I also went to many companies and research institutes for exchanges and attended many international conferences.

After graduation, I first worked at an overseas research institute for 4 years, focusing on AI R&D. Later, I moved to a large domestic company, spending 4 years in search, recommendation, and natural language processing, and leading the NLP department. Then, I went to an overseas e-commerce company for 4 years, responsible for overall AI R&D. After that, I joined another large enterprise for 3 years to lead global marketing algorithm R&D. In total, I've done this for sixteen years.

At the beginning of last year, a headhunter contacted me about an opportunity at Bitget. Although I had never worked in the crypto space 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 only had some understanding but no real relevant work experience, so I was a bit nervous before the interview. However, the interview went smoothly, and I received an offer. The position was Head of AI at Bitget, and I've been here for over a year now. Overall, this experience has been quite exciting for me. Every day brings new challenges and projects. Although the pressure is high, it's also very rewarding.

For me, the biggest change has been a cognitive shock. I had only heard of Web3 before but hadn't deeply participated, so after joining, I basically learned while working on projects. It's been very fulfilling.

Is the Integration of AI and Trading a Gimmick, or is it Already in a Practical Stage?

Mao Di: Bitget is a platform primarily focused on trading products. What's your take on "AI + Trading"? Is it truly in a viable stage now, or is it still more about market hype? Currently, not just the crypto industry, but almost all industries are embracing AI. Returning to this topic, do you think it's more about practicality, or is there also an element of riding the trend?

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

Trading itself is very complex. Different users have different levels of cognition, risk appetites, strategies, and operation methods. So, the key isn't "whether to do AI," but at which layer of the trading chain AI should be integrated.

If we talk about full automation, like fully autonomous driving, I think it's not achievable yet. But being an assistant, providing support in different stages and layers, is already very feasible. Regardless of whether Bitget does it, other companies are already doing it and reaping significant benefits.

For example, some traders mainly focus on short-term trends and quantitative signals. They might have had to watch many screens and data sets before. Now, AI is very suitable for integration and assisted judgment. Others make decisions based on news, financial reports, and social media. A lot of the work here is already information gathering and organization, and AI can significantly improve efficiency.

Going further, users will also want AI to not just help them find information but to provide more specific strategy suggestions, such as position size, direction, leverage, and even prepare the trade button. At a higher level, it could even approach an asset management model.

So, our assessment is that AI cannot completely replace top professional traders, but for average users, replacing up to 95% of the work is already in a practical stage today.

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

Mao Di: So, do you mean the first layer is already quite mature, like helping users understand project backgrounds, organizing information, and assisting judgment? Is Bitget's current AI product leaning more towards early-stage decision support, or has it already moved towards specific execution?

Dr. Bill: This goes back to last year. A month after I joined, we launched our Agent initiative. At that time, Agent was a very new concept, and everyone was figuring it out. We started with a small experiment called "Meme Catcher" because Meme coins were extremely popular, and market signals were fast and noisy, making it hard for users to seize trading opportunities in time.

We worked on this product for two months, and the results were decent, but its capability was relatively singular, mainly focused on catching Meme-related signals. Later, we upgraded it to GetAgent. The initial goal was to solve the first-layer need: information collection and organization. This part is essentially grunt work; as long as you tune the process and model, you can significantly improve efficiency.

So initially, our main focus was on the information side. This includes customizing important crypto news sources and providing this high-quality information to the model for analysis, rather than just letting the model search the entire web independently. By doing this, the accuracy of information collection and analysis improved a lot, and user satisfaction was relatively high.

But later, users started asking for more. They didn't just want to see information; they wanted decision-making advice. For example, whether to go long or short, how much to buy, or what risk level strategy is suitable. So, we began combining user historical trading records to create profiles, analyzing their risk preferences and trading habits, and providing more personalized suggestions.

This is because the information layer can be relatively universal, but at the trading layer, the differences are huge. Facing the same problem, different users might get completely different answers. So, GetAgent gradually moved towards personalized matching, and this part required a lot of fine-tuning.

At one point, we even reached the execution layer. For example, a user could directly say, "Help me buy $10 worth of Bitcoin," and the system would quickly prepare the trade button. The user could place the order after confirming. Of course, the premise was that the instruction had to be very clear, not too vague.

After this feature went live, some people did use it, and trading volume was increasing. But later, we found that if we continued to push deeper into "directly helping users place orders," users could easily misunderstand, thinking the product could make money for them. Once a loss occurred, there would be a gap between expectations and reality.

So, we later adjusted our direction. Instead of continuing to optimize automated order placement, we shifted our focus back to information gathering, aggregated analysis, and personalized supply, making these capabilities more solid.

By the beginning of this year, we launched Agent Hub. Unlike GetAgent, which operates in the app with a Q&A format returning long-form content, Agent Hub is more geared towards advanced users. It supports them in calling underlying capabilities programmatically and executing trades via command lines and similar methods.

This direction gained some attention at the time, but the barrier to entry was still high. Very few people can actually write programs and use command lines for trading. The vast majority of users are average traders who need a simpler and more direct product form.

So, we later moved the entry point to Telegram. Users just need to open a link and log in to their Bitget account to complete trades using a similar agent-like interface. The overall experience is smoother.

Mao Di: How did you address security?

Dr. Bill: For security, we implemented sandbox isolation, four-layer identity verification, and an independent environment. The core is to ensure the safety of user assets. Additionally, we try our best to lower the barrier for average users. Many similar products require users to connect models, manage token costs, and choose service plans themselves, which is too complex for most. We aim to hide 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 language model are you using?

Dr. Bill: We use multiple large language models and dynamically allocate them based on different tasks. The core is to balance cost and effectiveness simultaneously. Simple tasks shouldn't always use the most expensive model, and complex tasks can't rely solely on cheap models. So, we operate more like an overall optimization system.

In product design, we aimed to lower the barrier from the start. For example, we first give users a certain amount of free credits, and they only pay after using them up. This makes it easier to start. Users don't have to buy tokens or select models themselves; they can directly use the underlying capabilities we've already refined.

Later, we migrated many capabilities to Telegram, including information access, analysis processing, and some basic trading strategies. The product on Telegram is called GetClaw. This allows users to interact with the system directly, like chatting, making the experience smoother. In the app, many users couldn't even find the entry point, but on Telegram, the path is more direct.

Once this experience was streamlined, GetClaw took off quickly. We also complemented it with a trading competition, providing users with trial funds and rewards, essentially helping them naturally adapt to this agent-based trading model.

However, we always emphasize that no matter how good the tool is, trading cannot be completely separated from human judgment. Knowing when to enter and exit is still critical. Relying entirely on the model won't work, and not using the model at all won't either. So, our goal isn't to replace users but to make the tools good enough while helping them improve their understanding. This is why, since we started working on AI, we proposed a goal: "to make 100 million users comparable to Wall Street's elite," essentially making them better traders.

Our goal is to make trading simpler and more personalized. For instance, the system can gradually understand 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 have better support and confidence in your operations.

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, security and effectiveness, with underlying tools constantly evolving. GetAgent has refined many underlying capabilities over the past year, and GetClaw was built on this foundation. Of course, it's not perfect yet, and we will continue to iterate.

Mao Di: Have you tracked the trading volume driven by AI trading?

Dr. Bill: It's still not very high right now. As a percentage of the company's total trading volume, the portion entirely driven by AI is quite low. Building large-scale user trust in "AI-guided trading" itself requires a cultivation process.

Also, this field changes very rapidly. Large language models are evolving quickly. Often, you don't need to drastically change the product's front-end form; simply switching the back-end model from an older version to a newer one can significantly improve overall effectiveness. This shows that model capabilities and the application layer are starting to decouple. When the underlying model upgrades, the user experience improves accordingly.

So, the current state is that the application layer is iterating rapidly, and the models behind it are constantly advancing. The entire ecosystem changes very quickly. Previously, a requirement might take one or two months to implement; now, it can be deployed in a few days or even a day.

In this environment, what truly matters isn't just development capability, but a deep understanding of the business, especially the nature of trading itself. Because tools and models are evolving, the ultimate value of a product is determined by your understanding of the scenario.

Competitive Advantages and Continuous Optimization Directions of Bitget's AI Products

Mao Di: It's not just Bitget; Binance and OKX are also developing AI-related products. Have you seen the "skills" or products they've released? What advantages do you think Bitget's AI products have over other exchanges? In which areas do you think you perform better?

Dr. Bill: This is a great question. We also closely monitor the latest developments in the industry. In terms of AI, all exchanges started from a similar baseline. We see it as an opportunity for a "curve overtaking" or leapfrogging. At the same time, AI is a field requiring immense investment in both talent and capital, destined to be a playing field for top exchanges. Bitget's investment in this area is substantial.

Since we started GetAgent last year, we've been exploring how to build AI Agents for the crypto space. There were hardly any existing references at the time. We had to look at how it was done in other fields and continuously explore based on our own business. After more than a year of work, we've accumulated solid underlying capabilities and developed a methodology for continuous iteration.

Compared to other exchanges, I think our advantages lie in several key areas.

First, our iteration experience. Since we started developing AI Agents in March last year, we've gone through multiple quarters of continuous iteration. This process has been painful, often feeling like we were starting over, but precisely because of this, the accumulated experience is deeper. I wouldn't say we are definitely number one in the industry, but we were relatively early to start and have gone quite deep.

Second, is security. When these Agent products first appeared, many people rushed to try them, but later withdrew due to security issues. Internally, we have always prioritized security. Even if it affects development speed, security must be guaranteed. After several consecutive quarters of refinement, we haven't had any major security incidents in our AI trading and AI Agent efforts. This is also a crucial advantage.

Third, our speed in adopting new product forms. Whether it was Agent Hub or the later GetClaw, we launched them relatively early. We don't just build the product itself; we also design features around the trading scenario. For example, we previously attempted to integrate AI traders with a copy-trading system. Users could choose to follow an AI trader based on their performance. This represents a further innovation in the trading scenario.

On the surface, anyone can build such products now, quickly putting together a prototype using development tools. But the real differences in smoothness, stability, and reliability are huge once it's built. This isn't just about which model you use, but whether you can simultaneously manage model quality, cost, security, and user experience.

Especially in a C-end scenario, cost control is critical. Without optimization, the cost of these products can easily spiral out of control. So, what we are doing now is not just "which large model to use," but how to deeply combine and optimize multiple capabilities, ensuring a good experience and quality while keeping costs within a reasonable range.

In summary, I believe our advantages are mainly three: first, we started early, iterated for a long time, and have a deep accumulation; second, our security system is relatively solid; third, in the integration of skills and product capabilities, we have formed a certain methodology and foundation.

Of course, if there's anything that needs continuous optimization, I think the most important thing is not to keep staring at competitors, but to learn more from users. Because ultimately, AI trading isn't about who has more features, but who understands users better. What are users' current understanding, habits, and expectations regarding AI trading? We need to keep researching this.

Ultimately, users come to trading platforms to make money. We can't guarantee they will make money, but we hope to make their trading faster, easier, and more comfortable. For example, the system might eventually present you with just a few clear, personalized options and clearly explain the logic behind them, making it easier for you to make judgments and decisions with more confidence than before.

So, this journey is far from over. Our current focus is to continue making the experience smoother, safer, and more personalized, while also continuing to learn from both peers and users.

Will AI Trading Lead to a Winner-Takes-All Scenario? Will Strategies Become Obsolete Quickly?

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

The first question is, the capabilities of models executing AI trades will inevitably vary. Will a "winner-takes-all" situation emerge in the future? For example, those with more capital can buy more powerful large models, possess more computing power, and be faster. Ultimately, a few individuals or entities could defeat the vast majority and capture all