Claude Bloom × Yonsei MBA: A Sedative for AI FOMO

This was the first time Claude Bloom partnered with a university. The Yonsei University MBA program reached out first and asked me to speak. I felt it was too much to carry alone, and since this was our chance to show Bloom to MBA students, I wanted to show up armed with the best people I know. So I invited six of them: Seongsu Kim from Samsung Electronics; Seungin Ryu, the Gwangjin District Office public servant behind a legal-code MCP; Hyeontae Hwang of SpaceY; Junho Cho of Cliwant; Jeongmin Lee of ABLD; and Donggyu Kim of K-skill.

We usually mix in small-group sessions, but the MBA students told us they were more used to lectures, so we ran close to two hours of straight talks. The lineup spanned a big corporation, a public office, startups and solo builders, so the content was varied and every talk felt fresh. Nobody got bored, thankfully.

Honestly, I get nervous before every event. I never know what will go wrong, and since everyone is busy, we hadn't rehearsed much together, so I worried someone might get caught off guard. They're all pros, though. Even when I asked a clumsy question, they came back with exactly the right answer, which was a huge relief.

None of the speakers took a cent. They came a long way to share honest, useful stories with the Claude Bloom community and the MBA students, and I'm truly grateful. Thank you as well to the community members who signed up in record numbers again. And to the Yonsei MBA organizers who prepared this with us: thank you for all the work.

A sedative for AI FOMO

On Friday, June 12, the Claude Bloom × Yonsei MBA Tech Group event took place in Room 103 of the business school building on Yonsei's Sinchon campus.

It was already our third event that week. We usually hold these in office seminar rooms, but this time it was a university lecture hall. I never went to business school, and here I was at the front of an MBA classroom. What a thrill.

I started with why I began Claude Bloom. Even working at an AI startup, AI trends moved so fast that I kept wondering whether we were running the business right. I started interviewing people around me, and one of them was an older friend from college who works at Meta headquarters. They had done machine learning for 20 years and said using AI lately made them deeply depressed. When their manager handed them a task, they'd throw it at the AI: "Explain this project to me." Let it run for a day or two, then report back. At first it was a lot of fun. Now it just felt empty. I posted that story on my Brunch blog and it went viral in Korea. I posted it on Reddit and people in the US responded too. I wanted to do more interviews, so I held an event. 2,000 people signed up in three days, and it has kept going ever since.

The name Claude Blue comes from "Corona Blue," the Korean term for pandemic-era depression. The idea is that using AI a lot makes you blue: it's exciting at first, and the more you use it, the more you sink into the blue. That blue looks different for everyone. It can be FOMO, job anxiety or a sense of emptiness. Claude Bloom is about not stopping there and moving toward blooming instead, with like-minded early adopters getting together to share what they've learned.

The event title, "A Sedative for AI FOMO," came from one of the speakers' slogans. I picked the theme because I think FOMO comes from the gap between the AI world you see online and the way AI is actually used at work. So we brought six people from four points on the spectrum: big corporations, the public sector, AI startups and solo builders. The space was opened up for us by Kyunghoon Cho of the Yonsei MBA Tech Group, a club for MBA students interested in IT. They had run a Claude seminar earlier in the year, and this was their second. That same day, the Claude Bloom Discord passed 2,000 members.

I always book speakers the same way. I message the people I want to meet directly on Instagram or LinkedIn, and even with no pay, they gladly come. Our planning principle is "build the event we'd want to go to."

The lineup has a certain flavor too. I'm not a developer, so I avoid dev-heavy lineups. We'll book a lawyer, or have the rapper Paloalto come talk about AI music. "Wait, why is that person here?" is a big part of what makes this community what it is.

1. "I didn't break through. I did it in secret."

The first session paired Seongsu Kim from Samsung Electronics with Seungin Ryu from the Gwangjin District Office. I asked how they were bringing AI into conservative organizations.

Seongsu Kim runs 100 GPUs himself in Samsung Electronics' System LSI division. They serve 7,000 people, 4,000 of whom actually use them. Internal estimates say serving all 7,000 properly would take 300 GPUs, and they're making do with 100. On top of that, he built an environment that runs Claude Code on their own GPUs. He changed the endpoint in the Claude Code install binary to an internal domain and put LiteLLM in the middle as a translation layer, converting requests to an OpenAI-compatible format so they talk to the in-house GPUs. Confidential code never leaves the company. He spread it by building a launch script, deploying it daily and handing it off. Since AWS is also a competitor, they can't send confidential code to Bedrock either, so they run two tracks: confidential code goes to their own GPUs and everything else goes to Bedrock.

System LSI designs hardware, so it doesn't use software languages like Python. It uses RTL, a language made specifically for hardware design. This kind of code isn't public anywhere, whether it's AWS's or Google's, so general models like Claude or GPT have no ability to write it. Samsung does have a lot of in-house tools and a huge library of guide documents. Samsung is known as the heaviest user of Atlassian Confluence in the world. They use RAG over that massive body of internal documentation to run their own agents.

Getting there was a fight. When the GPUs came in, one executive wanted to give them only to a small group with strong AI expertise. Seongsu Kim took the opposite view: AI only really helps when people with domain knowledge use it. They fought for two to three months without reaching a conclusion. So he sent a company-wide email four times, and that's when he started making a lot of enemies.

The same thing happened with Claude Code. Executives had set MBO goals at the start of the year to build agents, and Claude Code alone made all of that unnecessary. The people whose performance metrics collapsed became enemies too. So many people call him now that his company phone number no longer shows up in the internal directory search.

At an executive's request, they now track an internal leaderboard of token usage. The top ten people use 80 to 90 percent of all the GPUs. His view was that one person can't possibly account for 80 or 90 percent of the company's capability, so the people further down need to use it evenly to raise the overall level. That view wasn't adopted.

System LSI is a non-memory chip business. Business is tough, and it runs with about half the headcount of comparable semiconductor companies. So people welcome AI far more than they resist it. When they open an internal Claude Code basics class, it fills up in a minute.

Seungin Ryu took a different route. Ryu is a grade 8 administrative civil servant who posts on Threads as "Ryu Who Does Side Stuff."

"To give you the bottom line, I didn't break through. I did things nobody asked me to. In secret."

The legal-code MCP started when Ryu asked an AI about a law, got a wrong answer and got chewed out by their team lead. It was a hallucination problem. Digging into how to make AI read laws accurately led to the idea of MCP. Put simply, to make an AI read laws, you need to plug in a USB stick called MCP. The Korea Law Information Center publishes an open API, so Ryu connected it through vibe coding and built the legal-code MCP. Ryu said they developed it all night after work, high on dopamine.

After Ryu shared it on Threads, lawyers, accountants and public servants started using it, and Korea's National AI Committee featured it as an example of government innovation. The internal reaction was "Why would you build that?" Things changed later, when Ryu hooked the legal-code MCP into an API-based AI service the office runs internally. Before that, colleagues without paid accounts didn't know how to use it. Ryu would ask, "Do you happen to have a paid account?" and they'd say, "Ah, let's talk another time," and hang up. Once the environment was there, Ryu finally heard, "Now I get what you did." The district mayor who offered support resigned a week later to run in an election, then came back. Ryu still pays out of pocket for Claude Code, ChatGPT Codex and Gemini.

When Ryu took over the district newsletter, the predecessor said they hadn't taken a single vacation day in six months. The job meant receiving 50 to 100 Hangul word processor files and editing them one by one. Ryu connected a document-parsing engine so everything came out in the right format in eight minutes. Ryu could start it, go have a cigarette and come back.

Asked how civil servants keep up with AI trends, Ryu said there was nobody inside to talk to. Ryu once told a close friend about something they'd built the day before and got, "I can't understand what you're doing, so stop." So Ryu looked outside and found Agent Korea, a KakaoTalk community of business developers and vibe coders. Seeing "oh, this works, this actually works" there again and again built a mindset: "If I can't do something, it's because I don't know how yet. It doesn't mean it can't be done."

Central government, provincial governments and local district offices all go by "civil servants," but they're really different groups, with different AI adoption speeds and different goals. When a provincial government adopts something, the districts copy it. They don't try to lead. They check first whether the district next door has done it. "If Seongdong adopts something, Gwangjin catches fire." Until Seongdong does it, the feeling is that you don't have to.

Ryu said security in Seoul's local district offices isn't as strict as in central ministries, and quoted Taewoong Park, a well-known Korean tech industry figure, who once asked what information Korean government offices hold that would actually be a problem if North Korea got it. Even North Korea wouldn't bother taking our stuff. Outside of contract-related departments, district office staff often don't know why security is needed. Central ministries are different, with some fully air-gapped networks. When Ryu gave a talk at the Ministry of Economy and Finance, there was no internet, so Ryu turned on a phone hotspot. That didn't work either, and it only barely went through because Ryu's phone number wasn't registered on that network. Fortunately, the National Intelligence Service has announced that it will apply tiered network security under its new framework within the year.

Local governments often outsource their AX (AI transformation), and Ryu said most of those projects fail. Each person is the domain expert in their own job, and outsiders who are only good at programming can't absorb that tacit knowledge. As one professor put it, if you can write it down, it isn't tacit knowledge. The blessing of the AI era is that the person who holds the tacit knowledge can now use the technology to solve their own problems.

The two speakers' advice on beating FOMO pointed the same way. Seongsu Kim said there's no such thing as an AI expert, and a person with domain knowledge who uses AI just a bit well is far stronger. Seungin Ryu said that instead of rushing to outrun FOMO, it matters more to frame your own fundamental problem well.

2. Using AI well and growing the business are separate things

The second session featured Hyeontae Hwang, CEO of SpaceY, and Junho Cho, CEO of Cliwant.

Hyeontae Hwang is four years into his startup. After years of doing B2B SaaS, at some point he felt like he'd been conned. While he was trying to figure out why B2B SaaS startups weren't working, people senior to him in the industry were doing SI (system integration, custom IT contract work), driving Maybachs and pulling in tens of billions of won in revenue. That's when his question changed from how to build great software to how to find a business where AI could multiply what he was capable of.

This is where he brought up FDEs, forward deployed engineers. Palantir created the role 20 years ago on a defense project: engineers who go into the client company, hear the problems directly and engineer the solution together. In Korea, the work is almost the same as traditional SI. The difference is scale. In the AI era, one person does what ten used to. Engineers in their late twenties are skilled enough, but the pressure that used to be split among dozens of people lands on one person, and they burn out mentally.

So SpaceY is moving to teams of two or three with different personalities: a crew chief, an engineer, a communicator. Everyone already uses the AI tools well. The real problem was the team structure afterward and hedging people's mental health.

He said AX projects split into two kinds: By Working, which automates office work, and By Coding, which automates development work. The KPIs differ too. By Coding is measured on deliverables, like SI. By Working contracts are based on the number of tasks automated.

Doing AX taught him that the push has to come from outside. He'd assumed innovation could happen from within, but with new technology or trends, organizations usually only move when someone outside pushes them, and getting consensus on that took time.

He also talked about how senior and junior roles are changing. QA on an AI product now requires someone at team-lead level or above, someone who knows the company's tacit knowledge and guidelines, to judge whether the output is right. Junior staff have a hard time doing it. The roles have flipped. Seniors organize the company's tacit knowledge into a context graph for the AI to learn, and juniors run 16 sessions in parallel and compete on speed through sheer output. Each contributes in a different way.

Junho Cho rolled out AI across his whole company starting early this year. He set aside Thursday afternoons entirely for AI experiments. He switched every tool to an AI-based one. And company revenue stayed exactly the same.

"Using AI well and the company doing well are completely separate things."

So he went back to the definition of a startup. A startup uses innovative technology and methods to break into an existing industry, and it doesn't have to be a product. The question wasn't how to use AI well. It was how to find a business where AI could multiply what they're capable of.

Customers kept telling him, "Just do all of it for us. Write the proposal and win the bid." Success-fee deals are hard, so at first he said no. Recently he changed his mind. If AI can shrink the bidding process so far that you can produce 100 proposals a day, then an AI-maximized consulting business becomes possible, one that finishes in under a week a process that used to take two months. They've pivoted in that direction.

He said the end picture matters. Elon Musk sets a vision of making humanity multiplanetary, so the team follows. Without an end picture, you can't get to the next stage no matter how well you use AI.

He also answered why a company that makes a bidding solution runs a community like this. His colleague Wonjun wrote a blog post at first, people responded, and he told Wonjun, "Do whatever you want." That's how it got here.

Claude Bloom has that end picture now too. Hundreds of white-collar workers gather every week, and plenty of companies want to advertise to them. In the AI era, ad channels have stopped working, and the ads that work best these days are out-of-home ads. In-person AI events could become the most powerful channel for companies. The end picture is to start in Korea and build the hottest AI event in the world: an AI-native company doing what the traditional MICE (meetings and events) industry couldn't.

Someone in the audience asked: big companies now build their own one-person automations internally, so won't startup products become unnecessary? Junho Cho said yes. In an era when companies build whatever they need in-house, competing on a single product has hit its limit. Value comes from going end to end and handling the customer's whole process from start to finish. He also said offline IP will matter more than products. Hyeontae Hwang came at it from another angle. In his startup's third year, revenue stalled at 2.4 billion won and headcount stalled at five, and he realized something: value now comes from an acqui-hire perspective. A smart, skilled FDE team is itself the asset.

3. Playing like you have nothing to lose

The third session featured two solo builders: Donggyu Kim, the developer of K-skill, and Jeongmin Lee.

Donggyu Kim is an undergraduate researcher at KAIST and leads Nomadamas, an AI open-source hacker house. K-skill is a collection of about 90 agent skills built for Koreans, covering things like booking KTX trains, searching Danggeun (Korea's big secondhand marketplace app), real estate auctions, KOSPI corporate filings and Public Procurement Service bids. It has 5.6K GitHub stars. He figures a GitHub star is worth 10 to 50 times a YouTube subscriber, which would make a repo aimed only at the Korean market the equivalent of a channel with hundreds of thousands of subscribers.

How it started is fun. He saw Garry Tan showing off on Twitter about gstack, which got 30,000 to 40,000 GitHub stars for a few markdown files, and it annoyed him. "I work this hard and I'm at 4.8K, and this guy gets that with a few clicks? Fine. If I make a skill collection for Koreans, I'll get a thousand at least." So he built it.

At first there were only six skills. He posted it with a patriotic, flag-waving thumbnail, and it spread once on GeekNews and LinkedIn, then blew up again when someone reposted it on Threads.

He does open source because it's fun, because he has too many ideas and turning them all into products and defending them is a hassle, and because it builds his personal brand.

His strategy has changed lately too. If a few clicks get something to 80 points, he doesn't push it to 90. He ships right away, gets feedback and fixes it. He learned that from younger builders. Ship at 80, and if something's off, talk your way through it.

"The scariest person is the one with nothing to lose."

If it's already open, there's nothing to hack, and people who copy it end up depending on it. Claude is number one as closed source, and Codex is number two as open source. The number two makes plays the number one can't. He thinks that's also why Chinese AI labs are releasing so many open-source models. Ask a Chinese model "Who are you?" and it says "I'm Claude," because, he said, it was built by buying thousands of subscription accounts and distilling Claude. Once people depend on those models, China gets to lead the chip ecosystem too. Now that Claude has pulled ahead again with Fable, Codex is getting anxious and running a promotion that resets your usage limit when you invite a friend. That's a number-two play as well.

Jeongmin Lee worked at the investment firm Bluepoint Partners, ran a startup for four years and has been a solo builder since May this year. OpenAI even offered Lee a Codex ambassador role directly, which Lee turned down for now because of being too busy. Next week, though, Lee is speaking at OpenAI Founder's Day Seoul. Truly the ultimate insider.

Lee works on the assumption that most software is going to collapse, and runs all kinds of experiments to see what still works. Looking back at the last startup, there's one regret: a period when traffic came in through YouTube and revenue jumped. Lee judged it a play that would be hard to repeat and dropped it. The regret is not pushing it harder.

Lee was candid about the limits of being a solo builder. If you're sick, no work gets done that day, and there's only so much you can hand to AI. You can make a presentation with a few clicks, but walking into a report to a C-level executive or a big event with click-generated slides is a completely different matter. That's where the pain points started.

Lee's current focus is AI rollups with private equity. When VCs reach out, Lee mostly doesn't engage, because a 5 to 10 percent stake isn't enough to change a company. Private equity firms acquire a stake that includes control, run the company directly, bring in AI to raise productivity, increase the company's value and sell it. With control, you can make the decisions yourself. That's what's called an AI rollup. There were no cases in Korea, but one has now been confirmed, and once there's one case, word spreads among private equity firms, who all know each other. Lee got into private equity through an introduction from someone who knew the firm's head, and described it as a business where access itself is the moat.

During the event, Lee mentioned a character chat company ranked around 50th or 60th on the App Store that passed 100 million won in daily revenue within a few months of launching. Lee had also heard of a saju (Korean fortune-telling) app built as a side project that reached 10 million won a month. While Lee was grinding on enterprise projects, others were making that kind of money, and it was a reality check. Lee said that instead of feeling it was too late, it made them want to jump in fast.

Clicking sounds could be heard throughout Lee's talk. Lee was building a camera app on the spot and planned to ship it right after the event.

Asked how a solo builder builds a moat, Lee said to build a channel.

"If you can't build a channel, you honestly shouldn't be a solo builder. If your service is hard to explain, explain yourself first. That's easier."

If you don't know what to build, Lee said, build a camera app, a saju app or a character chat app right now. Subscribe to Claude tonight, even the $100 plan, and start with Fable. Lee added that seeing people around them build games with a few clicks has made Lee think a lot about staying humble.

4. On token maxing

In the Q&A, someone said that the bigger the model, the longer its answers get, and the harder it is to keep up. The answer was to look only at the result. Throw the task as big as you can, hand off writing tests and verification too, and just take the report. After an hour or two of running, you get three lines at the end, and you have another model check whether it went well. Fable produced clean results this way.

Toward the end, we had a debate on token maxing. I'd split the two speakers' positions in advance.

Donggyu Kim was for it. Fable had just come out, combining the strengths of 5.5 and Opus 4.6, so right now is a crazy opportunity, and you have to use it as much as you can while it's available on subscription plans through the 22nd. Only by using the best model to the limit can you see where to make things efficient. You have to taste the ceiling to know what you can cut. Without doing a lot of that, you won't know where to trim, and it's cheap right now, so go use it fast.

Jeongmin Lee was for it with conditions: without intent and setup, you can burn tokens and build nothing. A solo builder doing a lot of work hits a cognitive load where you can't even start thinking about what to do. In that state, even with Claude Pro, Codex Pro, Gemini, Grok and DeepSeek all subscribed, you just spend the money and stare at the monitor. Token maxing is right for first learning how powerful a good model is, but running it relentlessly with no direction is a different thing.

In the end, both were saying the same thing. Around me, people are starting to clear their evenings after the event just to build with Fable.

After the event, we held our first-ever afterparty. The vibe was great.

Originally written in Korean on Brunch, June 14, 2026.