A sedative for the age of AI FOMO

What was discussed

The first session asked what AI adoption actually looks like inside a conservative organization. Kim Sung-soo of Samsung Electronics' LSI division supports 7,000 people (4,000 of them active users) on 100 GPUs, and built an environment on top of that infrastructure to run Claude Code on their own GPUs. LSI works in RTL, a hardware design language whose code has never been public, so general-purpose models are useless and an in-house tool was the only option. For two or three months he fought executives who wanted the GPUs handed only to "the few who are already good at AI," while he argued that "the people with the domain knowledge are the ones who should be using them." He pushed it through in the end, sending four company-wide emails along the way.

Ryu Seung-in, a grade-8 administrative officer at Gwangjin-gu District Office, took the guerrilla route. After his team leader chewed him out over an AI answer that got a statute wrong, he landed on the idea that "to make AI read the law correctly, you have to plug in a USB called MCP," and built a statutes MCP. He posted it on Threads, lawyers, accountants and public officials started using it, and it was featured as an innovation case by the National AI Committee. The district newsletter, which used to mean hand-editing 50 to 100 Hangul (HWP) files, now takes eight minutes thanks to document parsing. Both men had the same prescription for FOMO: "There is no such thing as an AI expert. Someone with real domain knowledge who gets just a little good at AI is far more powerful."

The second session turned to startups. Hwang Hyun-tae of Space Y said B2B SaaS had left him feeling "conned": his seniors were pulling in revenue in the tens of billions of won doing SI work. He brought up FDE, the Forward Deployed Engineer concept Palantir invented: someone who embeds with the client, hears the problem firsthand, and engineers the fix alongside them. In the AI era, one person does what used to take ten. The harder problem was never the AI technology itself but the team structure and mental hedging that come after it, and he argued that innovation only moves when the push comes from outside the organization rather than inside.

Cho Jun-ho of Cliwant was candid: he rolled AI out across the company and set aside Thursday afternoons for experimentation, and revenue stayed exactly where it was. His conclusion was that "the point isn't using AI well, it's being able to redesign the business around AI." A startup exists to break an existing industry with an innovative approach, and that doesn't have to mean a product. When a customer asked him to "handle everything from the proposal to winning the bid," he reframed it as a model that radically compresses the bidding process with AI. "If you don't have a picture of the endgame," he said, "no amount of AI skill gets you to the next stage."

The third session belonged to the solo builders. K-skill, built by KAIST undergraduate researcher Kim Dong-gyu, is a collection of 90 agent skills made for Koreans: KTX booking, Karrot searches, Public Procurement Service bids, and it has 5.6K stars on GitHub. He started it, he said, "because it annoyed me watching gstack rack up tens of thousands of stars off a handful of markdown files." His strategy shifted along the way: once something reaches 80 points, don't grind it to 90: ship it and fix it with feedback. He was equally frank about why he works in the open: it is fun, guarding everything is a hassle, and it is good personal branding.

Lee Jung-min, the other solo builder, has been offered an ambassador role for OpenAI's Codex. He experiments from the premise that "most software will fall apart," and is now focused on "private equity AI roll-ups," where he can hold control and change things directly. The moat for a solo builder, he said, is the channel, and if you have no idea what to build, build something tonight, a camera app or a fortune-telling app, it doesn't matter. In the closing debate on token maxing, Kim Dong-gyu argued in favor: "you have to taste the ceiling before you can see what to strip away." Lee Jung-min gave it conditional support: "without intent and setup, you can burn all the tokens you want and still build nothing."