Microsoft × Claude Bloom: Builder Josh

With sponsorship from Microsoft, we held a Claude Bloom event with Builder Josh, who has gone from solo builder to running an AI-native company.

Personal stories are always what stays with people longest. Josh kept saying "I'm nobody" throughout his talk, but after the event, people stopped him at the door. His story of his career, his failures, and starting over hit harder than any framework.

Honestly, Microsoft's security talk was much more fun than I expected. Microsoft reportedly spends about as much on security alone as Naver (Korea's largest internet company) is worth on the stock market. That's because, after the US Department of Defense, Microsoft is the most hacked organization out there. I'm thinking we should bring them in as a speaker at a future event.

A lot of people from big companies came this time: Samsung, Hyundai, LG, Hanwha, McKinsey, Toss (Korea's leading fintech app), and other heavy hitters.

The small groups always go over well. We give people three questions and tell them to network, and now they carry the conversation on their own without any prompting.

Asking each group to present is always a slightly awkward moment. But oddly enough, everyone had taken careful notes, and one person even played back a recording and read it out as is. People speak clearly and well, too. It seems like a lot of people who like to talk, and are good at it, come to these.

Some people said the presentation time was too short and asked others to post their LinkedIn on Discord and request coffee chats. Last time, I heard that one group went out to dinner together after the small groups ended. It feels like the community is growing in a useful, healthy way.

Summer is almost here. Before it arrives, I want to set up something that isn't a seminar: a gathering where people can talk comfortably without watching the clock. Rent a big cafe with a yard, and talk about AI on an early summer night while it's still cool, with a little alcohol. As soon as I posted it on Discord, the emoji reactions blew up. Everyone must have been waiting for it.

Junsu from our company made his debut as a host at this event. He said it was fun. Looks like we can double the number of events now.

1. Why an AI-native community gathered at Microsoft

Claude Bloom started with the goal of creating a space where people struggling with AI adoption can help each other. Since our events are capped at about 100 people, requests for materials kept coming from those who couldn't attend, so we also opened a YouTube channel. In the first three weeks of running it, we reached 1,200 YouTube subscribers and 1,500 Discord members. Considering how hard it is to grow a company account, those are unusual numbers.

Claude Bloom's first event was April 14, 2026. More than 2,000 people applied. I felt the weight of that, but we decided to keep going. Our schedule is already booked through the end of June. There's a collaboration event with Yonsei MBA, and on June 25 the rapper Paloalto is scheduled to give a talk.

The May 21 event was held at Microsoft's office. To keep events free, we'd been looking for companies to sponsor space and snacks, and Microsoft reached out. Jinha Seo, who gave the security briefing that day, said they hesitated at first about joining a community event for a competitor's product. Even so, they came because Claude Bloom had helped them a lot.

Korea has a culture of showing up eagerly for talks and seminars. When you see people with laptops open, taking notes as they listen, the mood is different from events abroad. Overseas, people tend to build relationships by enjoying leisure together first and getting to know each other as people. In Korea, things tend to start with swapping business cards and figuring out what the other person does for work. I think Claude Bloom is building a place somewhere in between.

2. A world where 95% use it but only 1% are safe

Jinha Seo, security manager for small and mid-sized businesses at Microsoft, gave a five-minute talk. It was short, but the point was sharp.

According to research, 95% of companies have already adopted AI. But only 1% of them are ready to use AI safely. Outside of large companies, you could say there are basically none.

Microsoft defines itself as the world's largest security company. There's a reason a company known for PowerPoint, Teams, and Copilot calls itself a security company. Since every company uses Microsoft, it gets hit by more hacking attacks than anyone except the US Department of Defense. It collects hacking signals from 190 countries every second, and invests about as much in security R&D alone as Naver's entire market cap.

The core message Seo stressed was account management. Security covers tools, devices, cloud, and networks, so it feels complicated, but protecting accounts well can prevent 99% of security incidents.

There was a real case two weeks earlier. A developer turned off multi-factor authentication (MFA) for 10 minutes to share a YouTube Premium account. In that window, their Google account was taken over, and 1.2 million won was charged in Gemini API usage. Even before generative AI, password attacks ran at about 7,000 attempts per second. Now, thanks to AI, that's down to milliseconds. Security has become a fight between spear and shield.

The second point Seo stressed was data oversharing. Developers very often paste sensitive information like API tokens straight into prompts, and OpenAI detects tens of thousands of leaked tokens a day. When you connect an AI agent to Slack, sensitive data like HR information can be exposed with a single prompt. That's why you need DLP labeling and a system for separating data.

Security also covers ethical problems, like AI generating copyright-infringing or politically sensitive content without any checks. You can use Microsoft Defender to evaluate the safety of the AI models you're using, and Microsoft also runs a Solution Assessment program that diagnoses a company's security environment for free.

3. A born designer quits

Builder Josh's real name is Seungkwon Kim. He studied design in both high school and college, a designer through and through. He calls himself a "designer to the bone." He started his career in 2012 at an agency with a starting salary of 24 million won.

That wasn't the whole story. He taught himself to code and wrote a book. That got him a job at SK Telecom (Korea's largest mobile carrier), and his salary jumped to 82 million won. His previous salary had been 29 million won, so it nearly tripled. After that, he raised it by about 10 million won a year, finishing 13 years of corporate life at 150 million won.

Around year 13, a problem came up. He didn't know what the next step was. He'd faithfully followed the stable path Korean society lays out, but if the step after team lead was executive, could he actually do that? Back then, he defined an executive as "someone with the leadership to take control of an organization and push things through." So he boxed himself in: "I just need to be good at design." Now he says that was a misunderstanding. Only after becoming a CEO himself did he learn how different hands-on skills and management skills are.

He changed direction. At a big company, design isn't a core function, so there was a ceiling on promotion, and he made it to final interviews at Google and Meta but didn't get in. He started a startup and side projects, and found that starting a business wasn't as hard as he'd thought.

In 2023, when he was subscribed to a lot of newsletters, he came across a business that recruited people for online challenges. Each cohort had 80 people paying 200,000 won each. It was the first time he saw that a newsletter alone could be a business. He tried it himself, and at first it didn't go well. But as he kept repeating the loop of putting things out, learning, and putting them out again, he realized a content business runs on the same principles as running a startup. In January 2024, he left SK Telecom.

4. Trust accumulates

The core of a content business is trust. Seungkwon quoted Chung Ju-yung, the founder of Hyundai: "Business is trust that accumulates." He said that's the line that has resonated with him most since running a business as CEO.

Justin Welsh, the No. 1 creator on LinkedIn worldwide, has a business model that fits on one page. Good content gets picked up by the algorithm so people discover you, and the more deeply people experience you through longer content, the more trust builds. You gather followers on short-form platforms like Threads or LinkedIn, then move them to long-form platforms like newsletters and YouTube. Once you've collected subscribers' emails or phone numbers, that converts through courses, e-books, and conferences.

Seungkwon makes content in a data-driven way. Instead of what he wants to make, he uses data to figure out what the market wants, then learns it and makes it, and he enjoys that process. He focuses on long form. At his busiest he makes three videos a week, 40 minutes to an hour each, and publishes two newsletters a week. He also runs a podcast, How I AI.

He started YouTube properly in November 2025. After some trial and error, he's now gaining an average of 10,000 subscribers a month. This month he had so many corporate training sessions that he couldn't post for three weeks, but he expects to end up with 6,000 new subscribers anyway. He dreams of reaching a million.

Short-form video alone, like Instagram Reels or TikTok, makes it hard to build trust. YouTube is the most powerful platform because it gives you algorithmic discovery and relationship building at the same time. He runs a funnel that brings clients from his YouTube channel to his AX agency, Joshua.

5. How AI-native companies work

Making content led him to report on AI-native organizations. His favorite example is GPTers, a company that runs AI study groups with tens of thousands of people. They built and use an AI agent called Bbojjak, based on OpenClaw. Bbojjak handles the whole process from recruiting to wrap-up. It sends texts from the database, sorts students by status, sends Zoom links, gathers and summarizes recordings, and comes up with improvements.

The key was the persona. They gave it a precise name and tone of voice and fed it context every day. Bbojjak reads yesterday's conversations, meeting notes, and every conversation from the 30 to 40 study groups running at the same time. They learned that it had to feel familiar, not like a stiff AI, for people to use it without pushback. He described it as raising it like a Pokémon.

Delta Society is a small company, but it adopted GitHub and Claude Code as company-wide tools. Every employee, non-developers included, loads data into a single source of truth (SSOT) and uses it for decisions and product development. The SOP is updated every day. What they learned yesterday goes into today's standard guidelines.

SpaceY, an AX consulting firm, works in an unusual way in client meetings. Partway through a 45-minute meeting, they pause and tell the AI to build it right then. By the time the meeting wraps up, the prototype is done, and they show it on the spot. He said clients get flustered.

Samyang Foods (the maker of Buldak spicy noodles) didn't change its existing organization. It set up a separate new AI-native team. By standardizing on a small number of tools and working only with those, it quickly became an AI-native company. The case shows how hard it is to transform an existing organization, and how realistic it is to start a new team instead.

AI-native companies have two things in common. They work in new organizations without legacy, and they have high vibe-coding literacy. Being able to use Codex and Claude Code is becoming a core organizational skill.

6. How Joshua works

Here's Joshua's core tool stack right now. Claude Code and Codex act as the operating system, and the Hermes agent handles the hands-on work. They used OpenClaw, but after errors became frequent they switched to Hermes, and performance improved. GitHub serves as the company-wide SSOT, and Slack is the hub where people and AI work together. Obsidian is for visualizing and managing meeting notes and shared files. Plaud Note is a device that costs about 300,000 won. With one button, it automatically transcribes meetings, and the transcripts pile up in Slack and the database.

The pipeline goes like this. Record a meeting with Plaud Note, a Zapier webhook drops it into Obsidian, it goes through GitHub, and Hermes reads it. As meeting notes keep piling up, the agent learns the context.

The agent that works in Slack is named Jena. Right now it drafts newsletters, issues tax invoices and quotes, sends text messages, and drafts B2B emails through the Gmail API. They also built their own internal ERP, which they run with no subscription fees.

Every morning at 9, Hermes suggests the day's work. Based on the context built up so far, it tells the PM, the assistant, and the PD what each of them has to do that day. It's a reference, nothing they have to follow, but he said it's cut down the time spent in the morning figuring out what to do today.

The team SOP is updated once a day. What they learned yesterday goes into today's AI guide. Under the name One Day, One Automation, they take on one automation task every day. The short-term goal is to have AI take over 20% of all work. Some people claim "AI can do 80%," but he said honestly he doesn't know. Start with 20%.

He didn't hide the current limits either. Automating important external communication is still hard. AI writes the drafts, but final decisions and review take longer than you'd expect. There's also an AI literacy gap between team members. When he goes out to do corporate training, no matter how much people are encouraged to learn, the "I don't want to, this is killing me" mood is obvious.

Even so, the biggest change is that they record 100% of their meetings. Documentation got faster, and so did decision-making. He closed his talk with this: "Going AI native isn't a matter of changing things. It means starting fresh from zero. Start recording, from your meeting notes, from your conversations."

7. The tables speak

After the talk, the roundtables begin. Each group discusses the questions they're given, then a representative from each group presents. That day, 10 groups presented.

Taeyoung Park from Lotte Innovate (Lotte Group's IT company) brought the idea of "greenfield and brownfield." In an environment with no legacy, you can automate the whole process from design to delivery, but companies trying to put AI on top of existing systems get stuck on limits like token caps and move slowly. One line stuck: "If something doesn't work with Claude, that doesn't mean it can't be done. It means you haven't found the skill for it yet."

Hunseok Choi from Hanwha Vision talked about comfort. Everyone is struggling on their own, running agents, setting up harnesses, adjusting permissions, and just confirming that you're not the only one who's lost made this meetup worthwhile. Old hands or newbies, high skill or low, everyone was fumbling.

Byeonguk So from Fujifilm raised process and responsibility. Plenty of teams are comfortable enough with the technology but haven't sorted out "How are we actually working right now?" Just as a first-year employee and a division head have different scopes of responsibility, how much you hand over to AI depends on the scope of that person's authority. Defining the process comes first.

Chaeun Jeon, COO of Devel, did something different. They recorded the conversation during the small group, had AI summarize it, and read that out as is. It was exactly what Seungkwon had talked about, put into practice right there. Their group also shared that Samsung Electronics, after adopting AWS Bedrock, started a Claude Code service this week. The rollout was finished about a month and a half after the PoC began. There was advice to drop OpenClaw and switch to Hermes. And someone said that to make an AI deal happen at a big company, "getting the report approved so it actually runs comes before building a useful service."

A developer from Toss talked about the AI gap. On one side, there are teams where one person handles backend, frontend, and app development alone. On the other, there are companies where having a Gemini subscription is something people brag about internally. If it's hard to push on your own in an organization that doesn't use AI company-wide, changing jobs is one option, the group said. They also described setting up a Mac Mini for the team with only read permissions, configured so it can open only specific PRs.

The group led by a sales ops person, also from Toss, talked about nothing but sales automation the whole time. They ran out of time and barely got to the other questions. Their conclusion was that AI is useful for finding overseas B2B customers. Wanting to keep talking, they asked people to post their LinkedIn on Discord after the presentations and request coffee chats.

Jinwoo Kim, from an industrial safety IoT startup, came face to face with the gap between startups and big companies. At startups, CS automation and building ontologies are already underway, and one example was a 100-page proposal for a government grant program, written with an agent, that actually got selected. Meanwhile, someone from a big company said they can't even use Claude Code at work. Their workaround was to make heavy use of outside activities to bring harness engineering and agent-building methods back to their teammates.

Suhyun Lee of Agile Coach Squared pointed out the reality of AX. The literacy gap barely narrows. Even when a company runs an internal contest and allows Claude Code, coordinating across teams is hard. There's clearly meaningful change in individual productivity, but it spreads slowly across the organization.

Minhyuk Kim, from a genetic testing startup, talked about two extreme paradigms. There are startups that want to hand everything to AI. Meanwhile, one person had an executive at a big company push them with "It's just one click, why can't you do it?" and they nearly ended up presenting a video they'd ground out themselves as "made by AI." Somewhere in between, everyone was building their own workflow.

8. Piling up data is where it all starts

Seungkwon's final message is the one that stays longest. Going AI native means starting fresh from zero, and fixing the old way won't get you there. The starting point is meeting notes.

When big companies try to connect internal data to AI, the most common blocker is the burden on the security team. Building isolated environments and blocking external leaks are not easy. So the starting point he suggests is modest: press the record button right now. Recognize meeting notes as data and start piling them up. Every conversation and every bit of tacit knowledge becomes material for AI.

Ten groups from all kinds of organizations told their own stories, and they sounded a lot alike. The technology is already here. Process, responsibility, literacy, data, and organizational culture are just slow to catch up with it.

Claude Bloom is a place to close that gap.

Originally written in Korean on Brunch, May 22, 2026.