OI 007 · · 11 min
From CyberConnect to Surf: Shiyu Zhang on AI That Can't Hallucinate With Money
Co-founder · Surf AI
Post-training a model under a regulated exchange's rules, why a trained model beats a wrapper, and where AI belongs in a trade: the decision, not the execution.
Watch the conversation
7:08Shiyu Zhang“based on the market data, you don't want the AI to hallucinate and make the wrong choice. That will cost a lot.”
Full conversation
- Intro: Shiyu Zhang, Surf AI
- From DLive and CyberConnect to Surf
- Two products: Ask Surf and enterprise AI
- The Bithumb partnership
- Proprietary data and post-training
- Regulation, localization and cost
- Bithumb, Surf's first centralized exchange
- Enterprise demand in Korea
- Execution vs decisions: where AI belongs
- AI and crypto in five to ten years
- What's next: campaigns and NFTs
- Where to find Surf
From this conversation
A closer look
- Guest
- Shiyu ZhangCo-founder, Surf AI (at the time of recording)
- Host
- Kinsa Durst
- Recorded
- TOKEN2049, Singapore, released
- Topics
- Agentic trading
In this conversation
Shiyu Zhang, co-founder of Surf AI, tells Kinsa Durst how a company that started in crypto social (DLive, CyberConnect) now runs AskSurf.ai for traders and researchers and powers the AI inside Bithumb's app, which he says reaches over a million users. He explains why Surf post-trains its own model rather than wrapping an API, why Korean regulation shaped that work, and why he thinks AI belongs in the trading decision rather than the execution.
Questions discussed
What does Surf AI actually sell?
Two things, according to Shiyu: AskSurf.ai, a retail product he describes as a crypto version of ChatGPT built on on-chain, social and market data; and an enterprise business, whose first deal is Bithumb, the Korean exchange, where the AI functions inside the exchange app are run by Surf.
Why train a model instead of using an API?
Because the enterprise work needs it. For Bithumb, Shiyu says Surf gathered the data and post-trained a dedicated model, then spent most of its effort on regulation and localisation, including rules about which words an answer may not contain. He adds that training their own model cut costs compared with the large model APIs and raised accuracy.
Is Bithumb Surf's first exchange customer, and what came after?
Yes. Shiyu says the Bithumb launch brought inbound interest from other Korean enterprises: financial institutions, exchanges and non-crypto companies that want either AI integration or crypto data, which he links to Korea's maturing crypto regulation.
How far should AI go in a trade?
Shiyu separates execution from decision. Execution is automation and speed, not AI; the decision is where AI matters, and it has to be reliable because a hallucination costs real money. He says Surf is still raising model accuracy and will not let it make mistakes at critical moments.
In their words
“based on the market data, you don't want the AI to hallucinate and make the wrong choice. That will cost a lot.”
“the biggest product market fit of crypto is issuing assets and trading assets. I think AI is still going to be playing a major role in this field.”
Quotes follow the on-screen captions of the released edit; filler words (um, uh, like) removed and omissions marked with an ellipsis.
Transcript
Machine-transcribed from the released edit and lightly edited; names, companies and figures were checked against the on-screen captions, but the text has not been reviewed line by line and may contain errors. Speaker labels are automatic. When in doubt, the video is the record. Timestamps link to the video. Download (.vtt)
Read the transcript (1,615 words)
Kinsa Durst Welcome to Panoramic Content Suite Singapore. We have Shiyu, the co-founder of Surf AI, Very exciting project that we have here with us Looking ahead into Token2049 Before that we were in KBW together. But thank you for making time to join us for conversation.
Shiyu Zhang Thanks for having me.
Kinsa Durst Yeah, before we get into Surf, I wanted to maybe quickly ask about yourself And how you got to where you are today.
Shiyu Zhang So I've been the crypto space since 2017. I started from doing DLive, Which is a decentralized live streaming platform. And then later I co-founded CyberConnect And we issued a CYBER token trying to explore The decentralized social graph And the social space in this area. So after that last year, I co-founded Surf, which is a crypto AI company. We sort of do the AI stuff for the companies. Yeah, yeah.
Kinsa Durst AI, crypto, those are hot buzz words That I think everyone's trying to figure out How that's going to turn out. I think in the beginning, We were like AI agents are going to trade for us. It's going to take over the on-chain ecosystem. But yeah, what's your approach To AI and crypto with your product?
Shiyu Zhang So we have two products. One is to the retail user. So we have a retail product called AskSurf.ai And it's a basically crypto version of ChatGPT. So we provide the professional data, Including the on-chain data, social data, As well as market data. And we combine those together And to provide the professional traders And the researchers the best answer in crypto space. And on the other hand, We also have products for the enterprises, Which is a B2B business. Recently, we just launched a partnership with Bithumb, Which is the top exchange in Korea. So what they have is in their application, They have a whole new AI function And it's all done by Surf.
Shiyu Zhang So basically now Surf AI and Surf Intelligence Is accessed through the Bithumb users Over a million users today.
Kinsa Durst Wow, congratulations. I think Bithumb's like with 40% of the market share As a Korean exchange. And Korea obviously is a huge, huge crypto market ecosystem. So big congratulations. Thanks. I can't wait to go back and try it myself As I'm based in Seoul. But nice, that's really cool. Because in terms of the LLM, My understanding of AI is that There's big LLM providers like OpenAI, But kind of the moat that startups and projects have built Is kind of like proprietary data, open weights, Or some sort of like data weights that they use To fine tune kind of the model for specific purposes.
Kinsa Durst And you guys are very much focusing on trading, Providing the best data, best possible guidelines And advice for users, for your users To make the best decisions And ideally make the most money Besides just using any other tool out there. Is that the correct understanding?
Shiyu Zhang Yeah, I would say, yeah, that's correct. So what we do, we're based on our proprietary data And also we provide the most professional AI models In this field. Like for example, for our B2B deals, We will post-training the model To serve our customers' needs. So for example, in the Bithumb deal, What we do is basically two things. One is we gather the data, Two is post-training the model, dedicate serving for Bithumb. And I think the hard part is we have to go Through all the regulations and compliance, Especially given the fact That Bithumb is a regulated financial institution in Korea.
Shiyu Zhang So based on the model, We have to do a bunch of localization And also a lot of rules you cannot contain Which words in the answer. So that's where we spend the most effort on. And we also, by training our own model, We significantly bring down cost compared To charge APT and Gemini and Claude. And also significantly increase accuracy in this case.
Kinsa Durst Well, you guys are a true AI company, Not just a LLM wrapper.
Shiyu Zhang Yeah, trying to do some real thing here.
Kinsa Durst Nice, nice, super sick. I know you guys, I was curious, With your partnership with Bithumb, Is this the first centralized exchange You guys are working with?
Kinsa Durst Oh, wow. Previously, you guys, I'm assuming would use More on-chain data analytics to provide data And solutions and answers to your users, Both B2C and B2B. But with Bithumb partnership, It's kind of opening up a new space For enterprise partnerships.
Shiyu Zhang Yeah, exactly. So we started from the crypto space. After launching with Bithumb, We actually have a lot of inbound interest From the other enterprises in Korea as well, Including the financial institution And also crypto exchanges, As well as other non-crypto, non-financial enterprises. It's because all of those customers Are basically either they're interested In the AI integration or our crypto data. So we can see at the regulation In crypto regulation in Korea matures More and more financial institutions That are increasingly interested In the crypto data and other stuff. So that's where our proprietary data comes into play. And also we are AI company.
Shiyu Zhang We also give the customer solution, AI solution to the customer. So that's where we are increasingly popular In the like enterprise marketing in Korea.
Kinsa Durst Well, that's cool. I'll be curious, There's a lot of recommendations that AI can give, But how far in terms of execution can AI go? Can they actually make some trades for you? Does it differ between on-chain trades And centralized exchange trades? Can they manage your assets Or maybe just move different Between different like liquidity LP pools for yield? Curious, yeah, how far you guys go?
Shiyu Zhang I think that's a very good question. So I would separate those into two parts. One is AI execution, But I won't use the execution part, Does not have many thing to do with the AI part Because the execution is all about automation. It's not about AI, but it's just automation, How fast it can do. And the other part is how to make the decision by AI. I think that's the most important part. You want the AI to be reliable And based on the accurate data, I think that's because in the financial industry, You have the highest standard for AI for accuracy, For reliability, Because based on the market data, You don't want the AI to hallucinate And make the wrong choice, That will cost a lot.
Shiyu Zhang So I think that's still something We're exploring right now. We're increasing the accuracy of our model And based on the existing market data, We won't allow the model to make any mistakes To like hallucinate in the critical moment. So that's what we're doing right now.
Kinsa Durst Right, right. So building towards the ideal, This is the current stage, But speaking of the ideal, Do you have a five, I know AI moves really fast right now, So it's hard to kind of predict, Make predictions, But do you have maybe five to 10 year later, What type of AI trading or AI and crypto ecosystem Do you envision happening?
Shiyu Zhang Yeah, I would say the biggest product market fit of crypto Is issuing assets and trading assets. So I think AI is still going to be playing A major role in this field.
Kinsa Durst Is issuing as well?
Shiyu Zhang Yeah, that's right. I think like the trading part, Not the issuing part. I think the issuing is mainly still controlled By human beings. How about the trading part? I think we will rely on the reliable AI more and more, And as the intelligence will become abundant in the future, I think everyone who has the access To the reliable AI in the future And customize to your investment appetite. So we'll have some customer solution for everyone. Right, right.
Kinsa Durst Reliability is really, really important, Especially when it comes to money. Yeah, yeah. I think it's one thing to chat with LLM And get information from it, But to actually execute on that information When there's money involved makes, It's a whole nother level, But I'm really excited to see How these both centralized and decentralized AI plays out. Yeah, given that you guys have a really successful KBW And Bithumb partnership, I'm assuming there's a lot of, As you mentioned, there's a lot of momentum
Shiyu Zhang That's building up.
Kinsa Durst Is there anything that you're looking forward to here This week during Token2049, Any more exciting announcements that you might want to share? Is it kind of just building on what's going on right now?
Shiyu Zhang Yeah, we can see the momentum Accumulating on the retail side. Recently, we also have several campaigns, Including the mystery box campaigns on the retail products. So our user loves it. So we'll see. Probably we're going to have some major news Announcing in the near future. Also, we are also launching our NFTs. So basically, the NFT is a loyalty reward For the early users of Surf. And we started issuing the NFT one year ago about, So right now we're actually distributing the NFT And reviewing the NFT. So I think that's a big moment for the Surf community.
Kinsa Durst I'll have to hop on the NFT train.
Kinsa Durst And if the audience wants to find out more, Where can they do that?
Shiyu Zhang Just go to the AskSurf.ai. Okay.
Kinsa Durst Ask AskSurf.ai. And they can also follow you on social media channels.
Kinsa Durst And will our listeners also get some special referral deals? Yeah, we can do that.
Kinsa Durst We'll add that in the link below.
Open Interest Season 1 · Singapore — co-sponsored by Kraken and Codex.
Get the next conversation.
One email per episode.
One email per episode. Unsubscribe anytime.

