Dialogue with the head of Bitget AI: AI trading can get infinitely close to a high score, but cannot reach 100 points

This episode of the podcast focuses on Bitget's AI trading product layout. Dr. Bill, the head of Bitget AI, reviewed his transition from traditional AI research and industry experience to the cryptocurrency sector, and systematically introduced Bitget's iterative path for AI trading products over the past year: from initially helping users capture market information, organizing news and signals, to combining user historical behavior for risk profiling and personalized suggestions, and then attempting to lower the barriers to AI trading through Agent Hub, Telegram formats, and interactions similar to Claude Code.

The interview also discussed the boundaries of AI in trading: it has significantly improved the information processing and decision-making efficiency for ordinary users, but it still struggles to completely replace top traders; the future competition will focus not only on model capabilities but also on security systems, cost control, product smoothness, long-term memory systems, and continuous learning of users' real trading habits. The two sides also explored whether AI trading will be "winner takes all," and whether strategies will quickly become ineffective, concluding that the market remains highly complex, and human nature and black swan factors still make trading difficult to be completely dominated by a single system.

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

Cat Brother: Welcome everyone to this episode of "Wu Says No to Cryptocurrency Podcast." Today's interviewee is Dr. Bill, the head of Bitget AI. Please introduce yourself and how you entered the cryptocurrency industry. Also, I want to talk about your experience in AI. I heard everyone calls you Dr. Bill; did you come from an AI background?

Dr. Bill: I graduated with a PhD in 2009, and my undergraduate, master's, and doctoral studies have always been 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 for four years in an overseas research institute doing AI research and development. Later, I went to a large domestic company, where I worked for four years in search recommendation and natural language processing, and was responsible for the natural language processing department. After that, I went to another overseas e-commerce company for four years, responsible for overall AI research and development, and then went to another large enterprise to oversee global marketing algorithms for three years. In total, I have worked for sixteen years.

At the beginning of last year, a headhunter contacted me, saying there was an opportunity at Bitget. Although I had never worked in the cryptocurrency sector 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 couldn't say I was familiar with Web3; I only had some understanding and had never really done related work, so I was a bit nervous before the interview. However, I passed the interview quickly and received an offer. My position is the head of AI at Bitget, and it has been 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 is also very fulfilling.

For me, the biggest change has been the cognitive shock. I had only heard about Web3 before and had not participated deeply, so after joining, I was basically learning while doing projects, which has been very enriching.

Is the Combination of AI and Trading a Gimmick or Has It Entered a Practical Stage?

Cat Brother: Bitget is primarily a trading product platform. What do you think about "AI + Trading"? Is it really in a feasible stage now, or is it more about market hype? Because now, not just in the cryptocurrency industry, almost all industries are embracing AI. Returning to this topic, do you think it is mainly practical now, or is there still an element of riding the trend?

Dr. Bill: I think for Bitget, this is no longer a gimmick but a necessity. Because Bitget did not have a dedicated AI team for the first seven years, and there were very few algorithm applications, it was only in the past two years that we began systematic investment, the core reason being that AI has matured enough to truly enter trading scenarios, whether for cost reduction and efficiency improvement or for enhancing revenue and trading efficiency, it already has real value.

Trading itself is very complex; different users have different cognitions, risk preferences, strategies, and operating methods, so the key is not whether to "do AI," but at which level AI enters the trading chain.

If we talk about full automation, like fully autonomous driving, I think we still can't achieve that; but if it's about assistance, and assisting in segments and layers, that is already very feasible. In fact, regardless of whether Bitget does it or not, other companies are also doing it and have already reaped many benefits.

For example, some traders mainly look at short-term trends and quantitative signals; previously, they might have to monitor many screens and data, but now AI is very suitable for integration and auxiliary judgment. There are also some people who make decisions based on news, financial reports, and social media; much of this work is essentially information collection and organization, and AI can significantly improve efficiency.

Furthermore, users will also hope that AI not only helps them find information but also provides more specific strategy suggestions, such as position size, direction, leverage, and even preparing the trading button. At a higher level, it may even approach asset management models.

So our judgment is that AI cannot completely replace the top professional traders, but for ordinary users, achieving a 95% replacement of work has already entered a practical stage today.

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

Cat Brother: So you mean the first layer is already quite mature, such as helping users understand project backgrounds, organizing information, and assisting in judgment. Is Bitget's current AI product more focused on early 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 the Agent direction. At that time, Agent was still a very new concept, and everyone was exploring. Initially, we did a small experiment called "Meme Catcher," because Meme coins were particularly popular at that time, and the market signals were fast and mixed, making it difficult for users to seize trading opportunities in time.

This product was developed for two months and performed well, but its capabilities were relatively single, mainly capturing Meme-related signals. Later, we upgraded it to GetAgent, with the initial goal of addressing the first layer of demand, which is information collection and organization. This part is essentially manual work; as long as we optimize the process and model, we can significantly improve efficiency.

So initially, we focused on information capabilities, including customizing important news sources in the cryptocurrency space, and providing this high-quality information for the model to analyze, rather than simply letting the model search the entire network by itself. This approach significantly improved the accuracy of information collection and analysis, and user satisfaction was relatively high.

However, later users began to propose further demands; they not only wanted to see information but also hoped to receive decision-making suggestions. For example, whether to go long or short, how much to buy, and what risk level is suitable for the strategy. So we began to combine user historical trading records to create profiles, analyze their risk preferences and trading habits, and then provide more personalized suggestions.

Because the information layer can be relatively generic, but at the trading layer, the differences are significant. Different users may have completely different answers to the same question. So later, GetAgent gradually moved towards personalized matching, and we refined many details in this part.

At that time, we even reached the execution layer. For example, users could directly say, "Help me buy 10 USDT of Bitcoin," and the system would quickly prepare the trading button, allowing users to place an order after confirmation. Of course, the premise is that the instructions must be clear enough and not too vague.

After this feature went live, some people indeed used it, and trading volume increased. But later we found that if we continued to push towards "directly helping users place orders," users could easily misunderstand and think this product could make money for them. Once losses occurred, there would be a problem between expectations and reality.

So later we adjusted our direction, no longer focusing on optimizing automatic ordering, but instead returning our focus to information collection, aggregation analysis, and personalized supply, making these capabilities more solid.

Then at the beginning of this year, we launched Agent Hub. It is not like GetAgent, which is a question-and-answer format in the app returning long content, but is more geared towards advanced users, supporting them to call underlying capabilities through programming, completing transactions via command lines and other methods.

This direction gained some attention at the time, but the usage threshold was still relatively high. Because very few people can actually write programs and use command lines for trading, the vast majority of users are still ordinary traders, and they need a simpler, more direct product form.

So later we moved the entry to Telegram. Users just need to open the link and log in to their Bitget account to complete transactions in a manner similar to Agent, providing a smoother overall experience.

Cat Brother: How do you ensure security?

Dr. Bill: For security, we have implemented sandbox isolation, four-layer identity verification, and independent environments, with the core goal being to ensure user asset safety. Additionally, we also try to lower the usage threshold for ordinary users. Because many similar products require users to connect models, manage token costs, and choose service plans, which is too complicated for most people. We hope to hide this underlying complexity, making it easier for users to get started.

The Underlying Logic of Bitget AI Trading Products and User Experience Design

Cat Brother: Which large model are you using?

Dr. Bill: We use multiple large models and intelligently allocate them based on different tasks, with the core being to balance cost and effectiveness. Simple tasks cannot always use the most expensive model, and complex tasks cannot rely solely on cheap models, so we are more like doing a comprehensive optimization.

In product design, we initially aimed to lower the threshold. For example, we first give users a certain amount of free credits; once the credits are used up, they need to pay, making it easier to get started. Users do not need to buy tokens or choose models themselves; they can directly use the underlying capabilities we have refined.

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 way, users can interact with the system as if they were chatting, making it smoother to use. Because when it was in the app, many users actually couldn't even find the entry point, but placing it on Telegram made the path more direct.

After this experience was streamlined, GetClaw quickly gained traction. We also organized trading competitions, providing users with experience funds and rewards, essentially hoping to help everyone adapt to this Agent trading model more naturally.

But we have always emphasized that no matter how good the tools are, trading cannot completely detach from human judgment. When to enter and when to exit is still very critical. Relying entirely on models is not feasible, and not using models at all is also not feasible, so what we want to do is not to replace users but to make the tools good enough while helping users enhance their cognition. This is also why we set a goal from the beginning of our AI journey, called "Let 100 million users stand shoulder to shoulder with Wall Street," essentially to make them better traders.

Our goal is actually to make trading simpler and more personalized. For example, the system can gradually understand your trading habits, risk preferences, and operating styles, condensing the previously complex analysis process, and finally giving you a few clear decision options. This way, when you operate, you will have more basis and feel more at ease.

So the two core aspects of this product model are: first, long-term memory and personalized adaptation, where the system can continuously learn from users; second, it is safe, effective, and the underlying tools are continuously evolving. GetAgent has refined many underlying capabilities over the past year, and GetClaw was developed on this basis. Of course, it is still not perfect, and we will continue to iterate in the future.

Cat Brother: Have you ever counted how much trading volume AI trading currently has?

Dr. Bill: Currently, it is still not much. Looking at the overall trading volume of the company, the proportion driven entirely by AI is still very low. Because to gain large-scale user trust in "AI-guided trading," it requires a nurturing process.

Additionally, this field changes very quickly. Large models are continuously iterating rapidly, and many times, the previous product forms do not require major changes; simply switching the backend model from an old version to a new version can significantly improve the overall effect. This indicates that the capabilities of models and application layers have begun to decouple; once the underlying model is upgraded, the upper-layer experience will improve accordingly.

So the current state is that the front-end applications are rapidly iterating, and the backend models are also continuously improving, with the entire ecosystem changing very quickly. Previously, a demand might take one or two months to develop, but now it can be launched in a few days or even a day.

In this situation, what is truly important is not just development capability but understanding the business itself, especially the understanding of trading. Because tools and models are evolving, but ultimately, what determines product value is your understanding of the scenario.

Competitive Advantages of Bitget AI Products and Directions for Continuous Optimization

Cat Brother: Now not only Bitget, but Binance and OKX are also working on AI-related products. Have you seen the skills or products they have released? What advantages do you think Bitget's AI products have compared to other exchanges? In what areas will you do better?

Dr. Bill: This is a very good question, and we have been paying close attention to the latest developments in the industry. In terms of AI, all exchanges are at the same starting line, so we see it as an opportunity for "corner overtaking." At the same time, AI is a field where both talent and capital are heavily invested, destined to be a battleground for several leading exchanges, and Bitget's investment in this area is substantial.

In fact, since we started GetAgent last year, we have been exploring how AI Agents should be developed in the cryptocurrency space. At that time, there were almost no ready-made references; we could only look at how other fields were doing while continuously exploring in conjunction with our own business. Now, after more than a year, we have accumulated relatively solid underlying capabilities and formed a method for continuous iteration.

If I were to compare with other exchanges, I think our advantages mainly have several aspects.

First, is iterative experience. Since we started working on AI Agents in March last year, we have gone through multiple quarters of continuous iteration. This process has been quite painful; many times it felt like starting over, but because of this, the accumulated experience is relatively deep. We cannot claim to be the industry leader in this regard, but at least we have done it relatively early and in depth.

Second, is security. When such Agent products first came out, many people rushed to try them, but later withdrew due to security issues. We have always placed a high priority on security internally; even if it affects development efficiency, we must prioritize safety. After several quarters of refinement, we have not encountered any significant issues in AI trading and AI Agents so far, which is also a very important advantage.

Third, we are relatively quick to follow new product forms. Whether it's Agent Hub or the later GetClaw, we launched them relatively early, and not only did we develop the products themselves, but we also designed gameplay in conjunction with trading scenarios. For example, we previously tried to combine AI traders with a copy trading system, allowing users to choose to copy based on the performance of AI traders, which is actually further innovation in trading scenarios.

On the surface, it seems that anyone can create such products now; with some development tools, they can quickly put one together. But once it is actually made, whether it is smooth, stable, and reliable varies greatly. This is not just about which model is used, but whether you can integrate models, costs, quality, security, and user experience well together.

Especially in C-end scenarios, cost control is crucial. If optimization is not done, the costs of such products can easily spiral out of control. So what we are doing now is no longer just about "which large model to use," but how to make deeper combinations and adjustments of multiple capabilities while ensuring experience and quality, keeping costs within a reasonable range.

So to summarize, I think our advantages mainly consist of three points: first, we started early, iterated long, and accumulated deeply; second, our security system is relatively solid; third, we have formed a certain methodology and foundation in integrating skills and product capabilities.

Of course, if there are areas that still need continuous optimization, I think the most important thing is not to keep an eye on peers but to learn more from users. Because AI trading ultimately is not about who has more functions, but about who understands the users better. We still need to continuously study what users' understanding, habits, and expectations of AI trading are today.

Ultimately, users come to trading platforms to make money. We cannot guarantee that users will definitely make money, but we hope to make their trading experience faster, more convenient, and more comfortable. For example, in the end, the system only gives you a few clear, personalized options and explains the underlying logic, making it easier for you to judge and more assured in making decisions than before.

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

Will AI Trading Lead to Winner Takes All? Will Strategies Quickly Become Ineffective?

Cat Brother: You just described a relatively ideal "AI + Trading" scenario; I have two more detailed questions.

The first question is that the model capabilities for executing AI trading must vary in strength. Will there be a "winner takes all" situation in the future? For example, those with more funds can buy stronger large models, have more computing power and faster speeds, and in the end, a few people can defeat the vast majority and take all the money in the market.

The second question is that the trading market changes rapidly; a set of strategies is often only effective in a certain phase and will quickly be imitated, followed, or even targeted. Does AI trading also have this limitation? Is it impossible to maintain a fixed advantage in the long term and must continuously iterate?

Dr. Bill: These two questions are indeed very much discussed topics in the industry.

First, regarding "winner takes all," I think it is not very likely to happen. It can be compared to the stock market; the quantitative and fund industries have developed for many years, but to this day, no company has taken all the market profits. Even if leading institutions are strong, there will still be many participants in the market for a long time.

The reason is simple: trading systems are inherently too complex; the results cannot be determined by just a few variables, as there may be thousands of variables behind them, along with various unexpected events and black swans. Therefore, I do not believe anyone can truly dominate the market 100%.

As for the second question, I believe AI trading certainly has a ceiling. Assuming perfect trading is 100 points, today AI might achieve 90 points, and in the future, it may approach 99 points, but it is very difficult to truly reach 100 points.

Cat Brother: Are you saying it is currently at 90 points, or that it can only reach this level in the future?

Dr. Bill: I mean it is currently around 90 points. It will continue to improve in the future, but I think it will always be very difficult to achieve a perfect score. Because the most challenging aspect of financial trading ultimately comes down to human nature. As long as there are people participating and making decisions in the market, there will always be emotions, biases, and irrationality.

Of course, in the future, there may also be a more extreme situation where the market is primarily not people trading, but Agents trading against each other. In that case, the situation would be different because machine execution is certainly more stable than human execution, and it would be a competition of model capabilities, system capabilities, and speed.

But looking at today's cryptocurrency market, it is far from reaching that stage. So overall, this is still a continuously evolving process. As long as there are people involved in trading, it is impossible to completely eliminate uncertainty.

Cat Brother: I completely agree with this answer. Because trading often involves using rationality to overcome emotions. If in the future, it is all AI trading, then the final competition may just be about intelligence levels and speed.

Dr. Bill: Yes, we are still far from that step, so there is still a lot of space in this field, and it is very interesting.