
huge shoutout to @FourPillarsFP for covering how we believe financial intelligence must be built for high stake environments like one of the most active exchange in Asia, @BithumbOfficial. research.4pillars.io/en/rese…
✓cofounder and ceo of @SurfAIHQ (ai for finance), @logitshq (research lab). prev @berkeley_ai

huge shoutout to @FourPillarsFP for covering how we believe financial intelligence must be built for high stake environments like one of the most active exchange in Asia, @BithumbOfficial. research.4pillars.io/en/rese…

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.

Compute should never be the bottleneck. If you’re in the serving track, apply! logits.com/OpenKernelChallen…

.@shiyuSQ breaks down how Surf built Bithumb AI with a clear line between what deterministic code does and what the AI model does, and why that line is what lets @BithumbOfficial trust Surf with their 500,000+ investors on the platform. youtube.com/watch?v=QB2qxgYM…




Logits is the research lab started by the team behind @SurfAIHQ. Read the full engineering note: logits.com/notes/bithumb-ai Grateful to Bithumb for the partnership, and to @the_bunny_chen and the @FireworksAI_HQ team for supporting us since day one.

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.

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.

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.

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.

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…

Bithumb AI: 500K users in 47 days, and 95% of Bithumb's high-volume traders use it. Built by @SurfAIHQ x @BithumbOfficial. Proud of this one.🔥🔥 finance.biggo.com/news/a7977…

When I worked on reCAPTCHA a few years ago, the goal was to let humans in and keep bots out Now we need to let the good bots in, and even let them participate in the economy Great to speak on the Machine Economy panel at @PanteraCapital's Blockchain Summit with @FranklinBi, @ryanli from @SurfAIHQ, and @darylX24 from @b3labs
