Dev Korea × Genspark × Social Series: AI Content Day

Our Bloom event with Dev Korea, Genspark and Social Series wrapped up, and it was packed. This time I wanted a more international crowd, so we teamed up with an overseas developer community. Thanks to that, we got our usual Bloom regulars plus a lot of people from abroad, which meant a lot to me. It was also a good reminder of how global Korea is becoming.

Thanks again to Chanyoung Ryu for a great talk. Chanyoung studied computer science at KAIST (Korea's top science and engineering university) and worked as a developer, but is now a full-time content creator who runs a marketing agency and grew a global-facing Instagram account (@chandlerintelligence) to almost 90,000 followers in three months. The line that stuck with me most was that tech companies today are all turning into social marketing companies. The slides were so good that I'd love to ask for a Korean version of the talk next time.

Finally, a sincere thank you to Uncommon Gallery for sponsoring the space, and to my partners who put in the work with me: Seulgi Kang, Madhav Mistry, Florian Ludot and K Choi.

AI Content Day

The theme was AI Content Day, and we put one question on the table: what happens to our work after AGI? There's no set answer to that, so we kept it open as a space for everyone to share what they think.

It started as a coffee chat. The idea came up while I was talking with Florian Ludot from Dev Korea and Madhav Mistry from Social Series. Bloom has over 7,000 members, but almost all of them are Korean. Dev Korea is the opposite: foreign developers working in Korea. The two groups live in the same city and rarely meet, and that kept bugging me.

The space spoke first

The venue was Uncommon Gallery, run by Common Computer, a Korean AI company. You might wonder why an AI company runs a gallery. Banseok Kim, who leads the business side, brought that up first in the opening.

In the end, I think the point of contact with customers is going to move through agents.

Right now the big topic is how companies handle their own data. Banseok's view was that the next question moves from data to permissions: what you allow. Once agents act on your behalf, agents end up meeting other agents and exchanging things.

Banseok showed what they've actually built. At Walkerhill (a hotel resort in Seoul), they set up an agent that uses internal data to handle room and restaurant bookings. Today, customers meet a company's agent. The next stage, Banseok said, is the customer's agent meeting the company's agent.

The picture was concrete. You tell Google's AI Mode you want to go to Tokyo in two months, or that you want to stay in Gwangjin-gu (the Seoul district where Walkerhill sits), and hotel agents get called in and talk to each other. Banseok said they had expected the agent commerce market to take off within this year.

For that to work, building agents isn't enough, Banseok added. First you have to clean up the company's internal data and connect the pipelines. You need that front layer, ontology or whatever you call it, and they've been doing exactly that work with a number of companies, as B2B AX (AI transformation) projects across about five sectors.

In the project room at the back of the gallery, they've built agents for artists and are running an experiment where those agents buy and sell work from each other in a village of agents. Banseok thinks this could actually happen around next year. I was curious about that room, so I went back in after the talk.

"The biggest change isn't AI"

The first speaker, Chanyoung Ryu, opened like this. (Slides were shared at the event.)

The biggest change happening today isn't AI. It's social media.

Social media used to be network-based: you opened Instagram and saw posts from the people you followed. Over the past year it shifted to interest-based. The algorithm shows you what it thinks you'll care about, no matter who you follow.

Follower count went from real influence to just a number.

An account with a million followers might get 6,000 views on its next post, and an account with 6,000 followers might get a million. When Chanyoung started a new account and posted the first video, it hit 500,000 views in 12 hours.

Chanyoung gave a quick background: studied computer science, worked as a software engineer, then ran product at one of Korea's leading influencer marketing platforms. Earlier this year Chanyoung founded Feed Media, which helps global tech companies enter and grow in the Korean market through social media. For the past few months they've been working with Higgsfield (a generative video and image platform) on making it the number one generative video and image platform in Korea.

Sixteen videos in two months brought the account to 90,000 followers, and along the way came brand deals with companies like SpaceX.

But Chanyoung didn't spend long on the follower count.

This account is my marketing agency's lead generation machine. Every week I get dozens of emails from tech companies asking about brand deals. I don't have to do any sales.

Chanyoung quoted Peter Thiel: the number one cause of failure isn't the product, it's poor distribution. Now that AI has made building products easy, that imbalance has only grown.

At one point Cursor was running 6,000 ads. They could test 6,000 because they worked with 6,000 creators. They ran everything those creators made, picked the ones that performed, and put those back out as paid ads. It wasn't random, either. They chose who to work with and wrote separate briefs, and that work was done by creative staff inside the tech company.

Chanyoung also laid out what product companies can do. If your product can be shown in a few dozen seconds, the fastest move is to just show people using it. One Korean startup started its own channel, posted ten videos and picked up 20,000 followers. There's the mascot approach, where you turn the product into a character and build a world around it, like the Duolingo owl. And there are skits that feature the people behind the product. ClickUp and Cluely use their employees and company culture as content, selling the product and the team at the same time.

Even if you grow your own channel and get millions of views, that's a drop in the ocean. The really big distribution comes from established channels. Those channels spent years building an audience and trust, so when they recommend a product, their audience buys it. That was the point of the Cursor example.

Chanyoung closed with a hiring pitch. The agency is just Chanyoung and a co-founder, they have more work than they can handle, and they're looking for people. Ideally someone good at both AI and content, but that combo is rare, so someone strong in one and eager to learn the other works too.

A few questions came after the talk. Someone who builds B2B company websites on Shopify asked how much of the video editing Chanyoung does personally. The answer: Higgsfield is for content creation, and editing comes after that. Chanyoung shoots their own videos and adds whatever B-roll is needed, sometimes generated with Higgsfield. The editing itself is still done by hand, and Chanyoung would like to outsource that part or hand it to an AI editing tool.

Someone with zero video editing experience followed up and asked whether Chanyoung had tried tools like HyperFrames. Chanyoung hadn't used HyperFrames or Remotion, and mentioned there are YC companies doing AI video editing but hadn't tried them yet.

A developer and data analyst asked how you pull in potential customers when people just skip boring, cheap-looking ads. The answer unpacked a part of the talk that had gone by quickly. You have creators put out hundreds to thousands of formats and storylines first, then buy the license to the videos that perform and run those as secondary ads. That's the main approach tech companies use right now.

Treat the LLM as your junior

Second up was K Choi from Dev Korea. K works as an AI engineer at a manufacturing startup and runs Free Code Camp Korea. They dropped out of college and built a career self-taught.

K walked through the job in detail, and the list was something. K handles LLM security, usage policy and data leakage for the 40 people across the organization who use Claude, and goes directly into people's settings to make sure nothing is leaking. K keeps telling everyone not to use the share button, because those links somehow get indexed by Google search. K also keeps tabs on the watermarking that goes into Claude's output.

On top of that comes ontology and data integrity work: connecting user inputs and real-time manufacturing data to the LLM. Anyone who has dealt with data pipelines knows what a headache that is. K also coordinates between the sales, hardware and solutions teams in both Korean and English, and works on a government project with PhD partners in Germany.

Free Code Camp Korea has 1,100 members on Meetup and 850 on Discord. They run offline meetups every week with 10 to 50 people.

The talk started with the Transformer in 2017. Every model we use today is built on that architecture, and it's nondeterministic by design. The same prompt can give different answers, and even at temperature 0 you just get the most likely token. Hallucination isn't a bug you can fix, K said. It's the price of generalization.

K pointed out that it was Google that came up with the Transformer, not OpenAI, Anthropic or Meta. Before that, AI could handle small tasks like image recognition or speech, but nothing had passed the test Alan Turing proposed in 1950: actually holding a conversation with a person. That was the wall, and nobody thought chatbots would get there.

ChatGPT 3.5 came out in November 2022 and hit a billion users in June 2026. Since then we've added long memory, learning from human feedback, tool use and step-by-step reasoning, but K said all of it sits on top of the 2017 core.

As models got bigger and faster, costs went up too. Benchmark scores keep rising and new benchmarks come out every week, but nobody actually looks at them. When Claude Opus 4.7 came out, K said, the benchmarks were better and K's database got wiped. Every new model release brings fatigue, not excitement.

What K kept coming back to was time. Claude Code came out in February 2025, so it's been a year and a half. Harness engineering, the hottest keyword right now, dates to February 2026, so six months. When a friend asked how hard it is to build a harness, K said to first define what a harness is. Nobody has managed to define it yet. Nobody can become an expert in that short a time.

Sora came up on the same timeline. It launched in February 2024 and was shut down in April 2026, because video models are too expensive. Even the company that built it couldn't afford it.

At the end of the timeline was the EU AI Act, August 2026, the one with the watermark and copyright rules. Any company running AI in production needs to know about it, K said. K's own company is one of them.

K shared how to keep up. K follows things on X, Discord and Reddit, and for people who don't want to go that technical, recommended a few YouTube channels like ColdFusion, Internet of Bugs and 3Blue1Brown. K has watched and vetted all of them personally, but added that they could be wrong too, so always check and stay skeptical.

K used a case to show that confidence and accuracy are different things. In the US, lawyers used ChatGPT and submitted six nonexistent cases to a court. Getting a model to verify its own output doesn't work either. K tried building a verification agent and gave up, reasoning that if it worked, Anthropic or OpenAI would have done it already. K has seen firsthand how the model that wrote the code and the model that wrote the tests fail in exactly the same way.

Then came the hiring data. From 2020 to 2025, junior hiring dropped 34 percent and senior hiring dropped 19 percent.

Nobody's hiring juniors, and nobody's training them. Why would you need someone to write boilerplate code?

So what do you do? K's answer:

You need to become a senior, fast. Treat the LLM as your junior.

That means setting direction, breaking problems down, owning the architecture, making decisions, writing precise specs, and reviewing the output ruthlessly. K asked the audience whether they trust junior engineers. K doesn't.

K listed what seniors can actually do. With CI/CD, test automation, GitOps, a clean commit history, sandboxes and clean architecture in place, agents multiply your ability. With no tests, no CI, a messy repo, loose permissions and no architecture, the agent gets lost, your database gets wiped, and users and revenue walk out the door.

K also listed what stays with humans: history, stakeholders, people, the decision-maker who signs the checks, past meetings, and the lived memory of tech debt and legacy codebases. That's the context a model can't have.

K disagreed with the claim that LLMs are PhD-level.

People say LLMs are PhDs. They're not. I work with PhDs. I know they're not.

Strong on greenfield, helpless on legacy

K then summed up how to treat an LLM. It gives you tireless first drafts and boilerplate fast, and it has broad memory of many languages and APIs, which makes it great when you need to learn something quickly. It does well on tasks with precise specs. But it needs supervision, and it never takes on risk, because you can't blame the AI.

The weaknesses were more specific. New products and PoCs now take a few hours, but that's all greenfield. Drop it into a legacy codebase with built-up tech debt and contradictory information and it falls apart. K said that's the thing LLMs still can't crack, and if they ever do, K would be out of a job. They haven't yet.

LLMs are also weak on truly new problems. As evidence, K pointed out that LLMs have been around for years and we still don't have flying cars or a cure for cancer. Without a close precedent, the model just produces the nearest pattern, and that can be wrong. Python or JavaScript is fine, but move to less-covered languages like Rust or Zig and it breaks down.

Calibration is poor too. The model doesn't know how confident it is. It can sound completely sure and be completely wrong, and the lawyer case is the example: Mata v. Avianca, where six cases ChatGPT made up ended up in front of a court.

Long plans are another weak spot. When there are several interdependent steps, small mistakes and missed state pile up until the thread snaps. There are ways to compress context, but the context window still exists, and if an agent misses the first step, it skips it at the last step too, K said.

Cognitive debt

This is where the idea of cognitive debt came in. If you keep handing things off, your debugging instincts, design judgment and feel for starting code get dull. K had evidence, not just a hunch: an MIT Media Lab brainwave study found lower brain activity and worse performance when people worked without help afterward.

It's not technical debt, it's cognitive debt. You defer now and pay later.

K shared their own countermeasures. One hour a day, turn off the LLM and code by hand like a caveman. Put your code in front of other developers and get torn apart. Put on music and chase flow state. Learn new frameworks by reading the docs. Explain what you learned in very simple words to find where you get stuck.

New frameworks and languages should be learned through docs and real projects, with no LLM shortcuts. K asked the room when they last read documentation, which got a laugh, and admitted it had been a few months for K too. K said to write test-driven development by hand, and called the simple-words method by its name, the Feynman technique.

Then, at the end, K offered a personal prediction.

We're eating the seed corn. The harvest is rotting.

With no juniors coming in, the pipeline dries up. Seniors stop growing because they're delegating. Teams shrink. Contradictory code piles up with no architecture. Eventually people who can solve problems, debug and design become scarce, and around then someone will trust an agent too much and something big will blow up, whether in military, medical or finance. K called it an AI Chernobyl.

K closed with three principles that are 50 years old. The Pareto principle: if you can't keep up, focus on the 20 percent that produces 80 percent. The Mythical Man-Month: adding people to a late project makes it later. Goodhart's law: when a measure becomes a target, it stops being a good measure. K said models today being optimized to do well on benchmarks is an example of the third one. K added one more, the cobra effect: incentivize people to use lots of agents and you burn millions of dollars in tokens and get bad code.

For Pareto, K spelled out what that 20 percent is: context engineering, verification loops, tool use, MCP research mode and multi-agent loops. Get those five down and you'll have no trouble using coding agents. For the Mythical Man-Month, K said an LLM is an infinite junior: it amplifies a good senior, but put it into a bad project staffed only with juniors and things get worse. For the cobra effect, K used AWS as the example.

Those principles date from 1896 and 1975. More than 50 years old, and K uses them every day. The fundamentals that survived the dot-com bubble and the crypto crash are what AI will end up amplifying, K said, and they're also what agents can't replace.

How do you say this to your boss?

Someone asked what to say when leadership pushes you to use more AI. They understood the rules K applies personally, but how do you say it to the people above you?

I have that conversation every day. But the less you know about AI, and the less you know about the subject, the more impressive AI looks. Because the hallucinations look more convincing.

K's answer was to trust yourself. Don't say AI is better than you. Say: this is my skill, and this is what you pay for. But don't say "go ahead and replace me with AI," K added. They might actually do it.

Generation or slop

Third was Madhav Mistry from Social Series. Madhav started from the observation that social media content all looks the same these days. Similar card news (the Korean carousel-style infographic posts), similar quote cards, similar charts. Patterns work, but if you crank content out that way, nothing about it stays yours.

Madhav gave an example. ElevenLabs' social content and the content from its competitor Fish Audio look almost identical. Fish Audio, the later entrant, copied the pattern directly and is growing on it. The pattern works, but you lose any sense of whose content it is.

Madhav also mentioned that LinkedIn recently added a button to report a post as AI slop.

AI-generated and AI slop are different. Give it context and it's generation. Pull it out with no context and it's slop.

So before you throw a prompt at it, do your research, attach reference material and set constraints. Madhav compared constraints to warning signs that tell the AI where to stop.

Madhav split research into two tracks: the reports consulting firms like McKinsey and BCG put out regularly, and going through Reddit, X and YouTube yourself. Then build what you've collected into a source memory inside the AI.

AI search came up too. Low-effort content with almost no human touch doesn't show up in AI-generated answers at all. So put your effort into building structure.

For the design layer, Madhav showed something Madhav had built: a constraints document for Claude that specifies colors and fonts, lists what to remove, and defines up front what the layout should look like. AI generates a lot of noise based on the memory stacked up on its servers, so you plant warning signs telling it where to stop.

Madhav also set one benchmark: a USP, a "unique stand out point," something that clearly sets you apart when people see your content. And Madhav said we're moving from prompt engineering to context plus test engineering: you need to understand the output first to give better input.

Madhav explained why shareability is the benchmark. Content that doesn't get shared doesn't get pushed beyond your followers by the algorithm. So value comes before follower count or impressions. Madhav built the playbook with Cursor, and it comes with a data lab that shows which part of the funnel needs more effort.

Madhav also walked through five layers that prevent slop: source, thesis, voice, design and quality. Once the first four are in place, you have a system, and with a system you can find where the problem is when the output looks off. The one rule was to make what other people will want to share, not what you like. Madhav built 56,000 LinkedIn followers in a year this way, and agreed to share the playbook on the Bloom Discord.

"I keep forgetting what I've already built"

Last was Seulgi Kang, a Genspark ambassador. Seulgi started as a marketer, then became a data analyst, and said opportunities opened up after starting to post on LinkedIn every day two years ago. These days Seulgi handles community management, content creation, marketing and analytics solo.

The story of how Seulgi got into Genspark was fun. Seulgi started using it because the slides looked good, and messaged the Genspark CEO at the time, who completely ignored it. Now Seulgi is an investor in the company. During the talk Seulgi pointed at the screen and said, "This is the guy who left me on read," and everyone laughed.

Seulgi explained Genspark in layers. At the bottom is Second Brain, which collects information from your apps, chats and email. Above that is Super Agent, which reads Second Brain's knowledge and does what you ask. Above that are tools that make sheets, slides and emails. At the top is Gen Team, where multiple AIs collaborate.

I thought it would be a product demo, but the problem came first.

I keep forgetting how much good content I already have.

Every time Seulgi made lecture materials, Seulgi planned them from scratch, only to find all of it was already in Seulgi's LinkedIn posts. So Seulgi switched to collecting everything in one place and pulling from there. When you ask AI to make slides, it usually scrapes random stuff from online, and that isn't your story.

Seulgi connected the apps they use to Second Brain, including email. That brings in schedules, work and community events. Lecture materials and social posts were dragged in and uploaded directly. The result looked like an Obsidian-style knowledge system.

Slides come from a skill Seulgi built. It makes Korean consulting-style slides, layout-driven with lots of text and graphics. Paste in markdown that Second Brain produced, tell it to use the skill, and the whole thing was done in 10 to 15 minutes. There was nothing to fix, because the slides were already written in Seulgi's own voice.

Design worked the same way. The pages Seulgi makes as an ambassador have to share one design, so Seulgi uploaded a Figma file and registered a design system. It read in the fonts, primary and secondary colors, buttons and mockups, and everything made afterward follows that system. Seulgi asked the obvious question, why not use the same feature in Claude, and answered it: running the same steps there produced AI slop.

Seulgi put several agents into one chat room and asked for a plan to become a global creator in 90 days. One prompt got 53 replies. If a brand strategist was needed, that agent would call in another agent.

Something unexpected happened along the way. Seulgi had connected a meeting recorder, and it picked up a private conversation with a friend in a noisy cafe. Seulgi's personal situation ended up in the plan as supporting evidence. Seulgi said it was more surprising than scary.

The recorder fits in the palm of your hand, records up to 35 hours at a time and stores 7,000 hours.

A question came right away from the audience: aren't you worried about data leaks with all your accounts connected? The answer was honest. Seulgi just trusts it. But Seulgi doesn't use it for conversations with their husband or mother.

Seulgi was just as honest about tools. Seulgi mainly uses Claude Code and switches to something else when the results aren't good. Seulgi recently posted a first YouTube video and tried out a bunch of things for the editing.

Seulgi also runs AfterWork AI, an AI community of about 2,000 people. Everyone at the event got 1,000 Genspark credits.

More questions followed. The first was how Obsidian and Second Brain differ. Seulgi uses an LLM wiki hooked up to Claude Code, but without a deep engineering background, structuring the wiki is tricky and citations often fail. Seulgi tried asking both to write a book based on Seulgi's LinkedIn posts. Genspark's version was fine. Claude's read like AI wrote it. If you can structure a wiki well and wire up citations precisely, Claude is probably already a good choice, Seulgi added.

The US came up too. When Seulgi joined the Genspark community, it turned out engineers in the US use Genspark quite a bit. Seulgi had assumed engineers wouldn't, and asked why. The answer was that vibe coding results are stable, so there's less need to fix the code much. People who want to get to the result fast, even if it costs a bit more, pick it.

Four people pointed at the same spot

After the sessions, we left all the remaining time open for networking. The inner and back rooms of the gallery were all open, so people spread out and talked for a long time. A good number of people stuck around even while we were cleaning up.

The four speakers didn't know each other and prepared separately, but as I listened, they kept pointing at the same spot. When making things gets cheap, the value gets pushed to either end: deciding what to make, and deciding who to send it to. I'm honestly not sure yet whether that's right. But four people coming from four different places and landing on the same spot does stick with me.

If we'd done this on our own, this mix would never have happened. Thank you, Dev Korea, Genspark and Social Series.

Originally written in Korean on Brunch, September 2, 2026.