←
ryan 🌊2,301 posts
ryan 🌊 avatar✓

ryan 🌊 ✓

@ryanli

cofounder and ceo of @SurfAIHQ (ai for finance), @logitshq (research lab). prev @berkeley_ai

◷ Joined December 2013▧ 141 Photos and videos
JK 🛡️ avatar
⟳ ryan 🌊 retweeted
JK 🛡️@jonathankingvc·19h

New @cbventures market map alert! 🚨 AI agents are moving from talking about markets to trading in them. But as agents get easier to build, where will lasting value accrue? My new post maps the agentic trading stack through that lens. Here’s where I’m focused: • Agentic trading harnesses and terminals that turn a user’s goal into an ongoing strategy, with real retention. • Market reasoning models and data assets that give trading agents a real edge. • Identity, audit trails, and performance verification infra built for non-human, financial decision makers. I care less about shiny demos and more about products that get harder to copy as they grow. If you’re building here, I’d love to hear from you.

Post attachment
ryan 🌊 avatar
ryan 🌊@ryanli·Nov 3

built this to track gpu prices across all marketplaces. dm if you need access

Post attachment
Open Interest avatar
⟳ ryan 🌊 retweeted
Open Interest@openinterest_hq·Nov 2

From CyberConnect to Surf. Shiyu Zhang, co-founder of Surf AI, on putting AI inside Bithumb's app, training a model that can't hallucinate when money is on the line, and why execution is automation but the decision is AI. With @notkinsa. Supported by @krakenfx and @trycodex.

Post attachment
Watcher Guru avatar
⟳ ryan 🌊 retweeted
Watcher Guru@WatcherGuru·Oct 30

JUST IN: 🇺🇸 US stock market turned into a crypto memecoin market today. • Intel $INTC: -31% • Reddit $RDDT: -30% • Roblox $RBLX: -20% • Illumina $ILMN: -19% • SanDisk $SNDK: -15% • Arm $ARM: -12.7% • Dell $DELL: -12.5%

Post attachmentPost attachmentPost attachmentPost attachment
ryan 🌊 avatar
ryan 🌊@ryanli·Oct 30

How did we fix it? We let an agent loop train and score 100+ variants over about a dozen iterations. SFT got the model writing in the right report format, but it kept inventing numbers. So the last step was DPO on its own drafts, each paired with the smallest fix, and that cut those mistakes by more than half.

Post attachment
ryan 🌊 avatar
ryan 🌊@ryanli·Oct 30

The hard part? The eval. This BTC draft reads fine, yet it breaks five of Bithumb's rules: it names a venue that isn't allowed, gets the flow direction wrong, cites a time window the data never had, uses the wrong Korean register and mixes up two different metrics. We spent 30-40% of our engineering effort catching mistakes like these.

Post attachment
ryan 🌊 avatar
ryan 🌊@ryanli·Oct 30

Bithumb AI now runs on a model it owns: Qwen3.8-27B, post-trained by @logitshq and served on @FireworksAI_HQ. It writes research reports for over 500k traders, scores higher than GPT-5.6 Sol, costs 96.5% less and runs 4x faster. Here's what we learned.

Post attachment
Open Interest avatar
⟳ ryan 🌊 retweeted
Open Interest@openinterest_hq·Oct 23

What does it take to build AI for @BithumbOfficial? @shiyuSQ, co-founder of @SurfAIHQ, explains how Surf adapts its models for Korea’s regulatory requirements while improving accuracy. Recorded in Singapore during TOKEN2049.

1:20
Surf AI avatar
⟳ ryan 🌊 retweeted
Surf AI@SurfAIHQ·Oct 22

Bithumb AI, produced by Surf team, is now being used by 80,000 users every day, and 95% of high-volume traders on @BithumbOfficial. To mark the milestone, all eligible Bithumb users will be gifted ~100,000 won in virtual asset. Read our latest coverage: finance.biggo.com/news/a7977…

ryan 🌊 avatar
ryan 🌊@ryanli·Oct 16

testing our new mlx inference engine on a new 256g m5 ultra mac studio

Post attachment