Genspark × Bloom: AI Slop vs. AI Assets

We ran a Bloom event with Genspark. We'd spent a long time planning and preparing it together, so it was really something to see so many people in one room with their laptops open, debating with each other and actually building things.

Seoyeon, a marketer at Genspark, said, "We're moving from an era of people who learn tools to an era of people who talk to tools." That one hit home for me. It made me want to stop taking my familiar ways of working for granted, question what I do every day, and keep asking myself how I can get a little closer to AI today and what new workflow I could design. It also made me want to stop just browsing the new tools that come out and actually experiment with them.

Seulgi, a Genspark ambassador, looked back on their time at an ad agency and showed how they used Genspark to give their portfolio a professional update and build their own personal brand. That was a real push for me. As someone who runs a community, I got to see up close how an individual can use tools well to build up assets of their own.

Finally, the opening session from Taewan Oh, CEO of ARK Point, tied the whole evening together. They talked candidly about the difference between AI slop and AI assets, sharing what they learned and where they failed after building hundreds of things themselves. It stuck with me. It made me think again about how much it matters to turn work into something more than a one-off deliverable, like a single slide made on the spot, as in today's Genspark examples: into an asset where the company's knowledge and know-how accumulate, and into work that actually contributes to the business.

And a sincere thank you to Hashed (a Korean venture capital firm) for sponsoring a beautiful space and an evening I won't forget.

Same model, and some people cure rare diseases

The first session was Taewan Oh of ARK Point. They're a former securities analyst who also worked as a consultant, and in their own words a complete humanities type. I liked the title: "The Fork Between AI Slop and Assets." They explained that "slop" is close to "filth" in Korean and has the feel of the feed you give pigs. "Asset" is the exact opposite. The slide title for this part was "The model is the same," and it framed the problem precisely. We all use the same models, ChatGPT or Claude, some people another one. Yet some people make assets with them and some make slop. Where does that difference come from?

As examples they brought up cases of people using AI to treat rare diseases. Common diseases already have drugs, but a rare disease that's yours alone, or your pet's illness, has no drug, and there are cases of people approaching drug development with AI. The contrast was striking.

"Some people use the exact same model to treat rare diseases and cure their dog's incurable illness. And some people can't diagnose a cold with AI."

They explained FOMO with data too. Models keep getting faster: each new version raises benchmarks by 1.2x to 1.5x, and the stages have moved from chatbots to reasoning, and from reasoning to agents that take action. They cited Korean numbers as well. Worldwide, 18 percent of the working-age population uses AI; in Korea it's 37 percent. Koreans pay for AI at a high rate, ranking first in the share of paid subscribers for major models.

The most striking part came from Goldman Sachs and JP Morgan research. When electric motors and IT first appeared, productivity didn't improve right away. Real productivity gains started the moment investment kept going and passed 1.5 percent of global GDP. Right now AI investment is 0.9 percent of GDP globally and 1.8 percent in the US, and it's projected to reach 1.4 percent globally and 2.8 percent in the US by 2027 to 2028. So the US has already crossed that line and the world will cross it within a year. The anxiety we feel isn't baseless.

Hundreds of things built, and I use none of them now

Then came their own story. They knew nothing about development. Ask them to open a terminal and they didn't know what the black screen was. They didn't know if Git and GitHub were the same thing. In that state, one evening a coworker said, "Just try it right now," opened Claude Code, and told them to quickly type what they wanted. Their first product got built right there. It was so fascinating that they built the next thing, which also worked, so they built two in one day, and from the next day they were completely hooked. They changed the company too. It had been a blockchain consulting firm, and they pivoted it into a company that does everything AI-native. Now they're developing an on-premises data analysis language model for a government project, and most of the early products are built by non-developers and designers, not developers.

This was the best part of the night. Once they knew how to use it, they started building all kinds of things like crazy every day. The signature one was a map of lawmakers' favorite restaurants. The idea: lawmakers must go to the best places, so they analyzed four years of corporate card records and plotted frequently visited restaurants on a map. It suddenly went viral and even made the news. That filled them with confidence. "AI means developers are done. I'll just do all the development myself." After that came a restaurant guide parodying Michelin, more than 10 apps released on Toss (a big Korean finance super app), and a site that scraped every product on Coupang (Korea's Amazon) and picked the single one with the best ingredients. To look impressive, they even put an executable on their desktop that summoned agent teams. "Shall we call up the humanities team?" Click, and out comes the humanities team. The funniest story was a fight with their girlfriend. They fed the fight into AI, had it write a report proving they were right, posted it to Notion, and sent it to their girlfriend. Of course they got chewed out. Realizing that wasn't working, they made a Telegram group with the AI, themselves and their girlfriend so everyone could share the context.

Then it dawned on them. I built all kinds of things, so what do I actually have left? How much did I make? This one made 10,000 won, that one 80,000 won (roughly $7 and $60). They bragged to friends for a few weeks, and everything else flopped.

"I built hundreds of things, and honestly, I don't use a single one of them now."

That's how they came to know for sure that AI produces both slop and assets, and that the two are different.

The four faces of slop

The first face is never reaching usable quality. People's standards have gone up. They start using something at 80 points and see it as a good product past 90. Build it with AI and you get 70, so it never actually gets used. Ask a developer to polish it a little and you hear, "The architecture is wrong from the ground up. It needs to be rebuilt. Do we really have to do this?"

Output pouring out while skill doesn't build up is also slop. The analogy was good. We walk every morning on the way to work, but you could walk your whole life and never become a 100-meter sprinter. To become a sprinter you have to walk while studying how to walk better and how to go faster. AI is the same. Using it a lot is obviously necessary, but just producing lots of output doesn't build skill.

So is failing to touch the core bottleneck that decides results. A recent report from a consulting firm says 85 percent of companies using AI haven't improved operating profit. They use AI for things that don't matter, can't use it for the things that do, and so the results don't change.

The one that stung the most was cutting the cost of production while raising the cost of consumption. Have AI summarize meeting notes and you get something long and very plausible. Record, click, share it on Slack or Notion, and feel proud that something that would've taken 10 minutes to write up yourself got shared in one. But the 10 people you shared it with end up spending more time figuring out what the notes say. Meeting notes need to be compact and fit what people need. Making them fast doesn't make them good. That's slop.

Does slapping a QR code on a flyer make you a digital ad company?

Assets, on the other hand, come down to a few questions. Does it keep getting used and improved? Did it have real business impact? Did it remove the core bottleneck, or just shift it? Did you redesign the workflow in an AI-native way? The example for the first was concrete. A co-founder at their company is working out and losing weight, and built their own workout log app linked to Telegram and a smartwatch. They use it every day, so it keeps improving, and a few friends use it, so it keeps getting better. The second question isn't whether it feels good, is fun or looks impressive. It's how much revenue it brought in and how much cost it cut. The third tied back to the meeting notes. If I made the notes in one minute and didn't listen, I didn't remove the bottleneck. I shifted it to the 10 people who have to read them.

The analogy they used for the last question stayed with me longest. As an analyst they covered the advertising sector, and in the early 2000s advertising shifted hard from offline to online. The shock of digital back then, they said, is a lot like the arrival of AI now. The core skill of an offline ad company was handing out flyers. What they worried about was where the foot traffic was. Lots of people pass in front of Seoul Station, so they worked hard on how to hand out flyers there. When the internet arrived, they thought, we need to become a digital company too. How? Let's put QR codes on our flyers. Now we're a digital ad company.

Were they a digital ad company? Today we can say no with certainty, because digital ad companies work by a completely different grammar. Instead of a business that blankets shops within a 200-meter radius with flyers, a whole new kind of company had to appear: one that analyzes data on people going online in real time, swaps banners in real time, and operates around brands and the whole country instead of a neighborhood. AI is very similar. Even if you use AI, the key difference is whether you stick a QR code onto a structure that doesn't think AI-native, or rebuild the business you have from scratch on the premise of this era.

Building your own AI tutor

Last, they shared one skill they actually use. They'd picked the one that they keep using, keeps evolving, and has helped their own growth the most: an AI study skill. The slide called it "building your own AI tutor." You're scrolling LinkedIn and get fired up. A famous engineer says building a wiki with a language model is huge. Oh, I need to do that. Usually you bookmark it, but you have so many bookmarks that nobody ever looks at them, so don't do that, they said. Instead, copy the link and run the skill, and study material gets generated in Notion.

There were four reasons it works. First, it goes beyond the article and gives you the background. If you don't know who that engineer is, it does nothing for you, and if you don't know why it came out, it means nothing, so it fills in that prior knowledge. Second, it explains how this differs from existing approaches, with comparisons and alternatives that aren't in the original. The third was the good one: based on an understanding of you and your organization, it lays out practical applications and problems. For example, it'll ask, "Our company has hundreds of documents piled up in Notion and Drive. If we adopt this new approach, what happens to those assets?" The explanations are folded inside toggles, so the team discusses first and only then opens them. Fourth, everyone's study material collects in one place in Notion, so you can trade comments with coworkers. Studying this way, they said, they could finish in five weeks what would take a semester at school.

They first tried AI at the end of February and started studying seriously in mid-March. They organized everything they'd learned, spent six hours reviewing it with AI, and published it as a Notion doc that runs 18 pages. It covers things like how to set up your folders and how to write documents, and it's aimed less at developers and more at humanities types: people in consulting, analysts, and those making business decisions who want to use AI. They agreed to post it in our Discord too.

From people who learn tools to people who talk to tools

The second session was Genspark: Seoyeon Park, the Korea marketing manager, and Seulgi Kang, an ambassador. The title was "How the Marketer's Language Is Changing," with the subtitle "Say Less, Ship More."

Seoyeon went first, laying out how marketing work is changing. They graduated not long ago, started as an intern and became full-time, and I liked how they introduced themselves as having used AI from the early days back in college. Since the days of asking a chatbot how to make braised baby octopus, they said, calling theirs the generation that grew up with AI. Their argument was clear. For the past few years marketers have been people who learn tools. You had to use design tools, look at analytics tools for performance marketing, and learn data analysis piece by piece. Now marketers stop learning tools and become people who talk to them.

"When the tools change, the job changes."

The marketer's job has changed three times so far, they said, and we're in the middle of the fourth. First was the outsourcing era: working with agencies or freelancers, handing over requirements in documents and getting deliverables back. Next was the self-service era: rather than handing it to an agency, I'll do it myself, so you looked up YouTube tutorials, learned the tools and made things yourself. Then came the prompt AI era, where you explain things to AI in detail to get output, and there's a lot to learn there too. Tell it not to do something and it may latch onto that keyword and head that way anyway, and you carry the burden of explaining everything in great detail. Now is the era of AI that understands context: you don't learn complicated tools, you just give a direction. One line on the slide summed it up exactly: "Context is the new prompt." The old prompt AI asked how you'd explain it. Now it starts out already knowing what you want to do.

That's where a feature called Second Brain comes in, they said. It understands your project history and connected apps, and produces output already knowing your weekly recurring tasks, the tone and style you usually use, and the assets you've already made. The email example landed. When several projects run at once, you can't remember what was said where, but it drafts emails already understanding your tone, so email that seemed like it would take all day gets done in minutes. This is where the marketers in the room nodded hardest. There's that gray zone that's too small to ask a designer for and awkward to do yourself, and the design features cover it, which is probably why so many marketers seek it out, they said.

Save the whole brand page

Next, Seulgi Kang demoed real examples. They've worked at ad agencies and media companies and now work as a marketing data AI consultant, and a track record of making a huge number of proposals and designs showed throughout the demo.

The most practical tip came here. Brands usually have a CI page on their website laying out the logo, images, color combinations, fonts and usage examples. Don't just leave it there. Right-click and "Save As," and the HTML page and all its images get saved. Upload that whole folder and register it as a design system. If you have AI generate a whole image at once, even if you told it to reference the logo or text, it's hard to verify that it followed the brand guide perfectly. If you register a design system and, instead of one flat image, have it work like front-end design, laying a background image and combining text and logo on top, you get quality output while sticking to the guide much more safely. They showed the two side by side, and the difference was clear. The left one felt somehow AI-made, and you couldn't be sure it followed the guide. On the right, the logo placement and tone lined up.

The demo featured two fictional brands. One was Velora, an athleisure brand set up as a brand for people who move at their own pace. They made a poster for a pop-up store in Seongsu (a trendy Seoul neighborhood) and an Instagram carousel. The campaign was called On Rhythm, with the copy "On the rhythm, as myself," and the logo placement, colors and fonts matched the brand guide exactly. The other was Mochino, a mochi brand. With the concept "Shape it at your fingertips, one a day, your own mochi," they built a kiosk interaction page for a Touch Room pop-up. You press, stretch and shake the mochi on screen, pick your favorite of three (Seullangi, Wakppu and Mallangi), finish the flavor and coating, and take it home with a QR code. The surprising part was that it generated and added the sound effects itself.

Finally they showed how to export the project records piled up in Notion as HTML and turn them into a portfolio website. A résumé that didn't look professional as a Notion page had turned into a one-page site under the headline "Connecting data and people."

What struck me most in the demo was the process, more than the output. You put in a request and it doesn't just build it right away. It asks a few questions back: how to handle the sub-copy, what programs run at this pop-up, what the visual direction is. Even if you toss in a lazy prompt, you only get a more complete result once you answer these follow-ups. Then they added, "You'll be typing in your own answers later, and you'll find that answering these questions makes you think a lot more than you expect." I heard that as: it's closer to a tool that organizes your thinking than a tool that makes things.

The five points they wrapped the demo with were good. You need a solid brand guide and philosophy up front, in the sense that you need something to register as a design system. You need a sense of direction about what to do, which you can work out by brainstorming with AI in chat. When the follow-up questions come, you need the conviction to decide. Even when things come out automatically, you need the eye to judge whether it's really well made and current. And you need follow-through. They admitted that they'd automated their own Instagram carousels and never posted a single one, which got a laugh. The easier making things gets, the more what's left is judgment and execution. That conclusion connected exactly with the slop-versus-assets talk from the first session. The last slide of the deck closed with: "Say Less, Ship More. Now focus on setting the direction."

Watching from the back

After the two sessions, 100 people opened their laptops. They built for an hour, spent 30 minutes passing their work around their tables, and spent 30 minutes sharing it on Discord. I stayed in the back the whole time.

Our community started on April 14. A piece I wrote called Claude Blue spread much further than I expected, so we held a first event, and within three days we had over 2,000 applicants. We couldn't fit everyone at once, so we started running events every week, and that's how we got to about 30. Now we have over 7,000 unique applicants and close to 3,000 people on Discord. And that night, the community ran without me holding the mic.

In the opening, Junsu said that when people hear "AI event," most assume it's for developers, but developers aren't the only ones afraid of AI. In fact our community is 30 percent developers and 70 percent non-developers. The three people up front that night, a CEO who used to be an analyst, a marketing manager and an ambassador from advertising, were all non-developers too. It was an evening of them teaching 100 non-developers how to build.

Originally written in Korean on Brunch, August 28, 2026.