Snowflake × Bloom: Daytrip, Data and Context

Today we held a Bloom event with Snowflake.

The man who set a standard for taste

We opened with Seokho Yoon, CEO of daytrip. You've probably come across daytrip on Instagram at some point. It's a travel curation page that puts together lists of places worth visiting in Seoul, in that format everyone knows now.

I first met Seokho at a previous event, and what he told me was interesting. I'd always thought that however good AI gets at automation, making content that actually moves people, content that feels made by a person and not like AI slop, still takes a high level of skill and is still hard. So I wanted to hear how much AI is being used on Instagram, and in a role where you have to pick out the very trendiest things. His answer filled half of that night's event.

Even his introduction was unusual. He describes himself as having four jobs: Debate Korea, a nonprofit he's run for 13 years with TV announcer Oh Sang-jin; daytrip, the travel curation platform; and being a husband and a father of two. He used to be proud of that, he said, but not these days. "A human can live at 100 percent, and if you become a superhuman you can do 120. But working four jobs means each of the four expects 100 percent of me. In reality I'm giving each one 25. So I end up living with all four of them mad at me." He spends his days figuring out how to turn 25 into 30, then 35. As a marketer at an AI startup who also runs a community, I related to that a lot. It's fun, but resources are stretched thin, so naturally you get interested in using AI well.

Why he started Debate Korea connected to this community too. Korea has no debate culture, he said. People never learn it, it never caught on, and the word "debate" makes people think of a shouting match. Having done structured debate in high school and college in the US, he came to feel that America's creativity and ability to scale come from a debate culture built into the country. Korea, he said, brings it in without any structure. His example stung: using English names at work doesn't create a debate culture. The way people think has to change first, but companies just change the names and then say, let's have an open conversation.

As an interesting recent collaboration, he described a UN project. They built a way for young people to bring their stories to the UN, and 27,000 people from around the world applied through an online platform. In the past, reading 27,000 applications would have been impossible. Now they had AI sort them by region, area and topic, and ran a decision-making process on top of that. They traveled to six continents in person, bringing AI into the process of turning individual opinions into a collective conclusion. The response was good, and in the end they invited 150 people to Korea and built a global community.

Moving on to daytrip, he made one thing clear first: he's not the founder. His younger sibling founded it and now runs the US version in the US, and he took charge of building out the Korean side.

The service started from the fact that there was no standard for taste. As humans we clearly feel that something is good, pretty or something we want, but there's no concrete way to find that feeling. In the end everyone looks up places in a map app and decides for themselves whether a place is good, and everyone's criteria are different. Some want close, some want cheap, some want nice. So the idea was to find one sense of taste. A few creators and curators started collecting places with great taste as "our own standard," and at some point people started trusting the places that showed up on daytrip. That trust became the business.

He said daytrip was the first to do the format that's now everywhere: card-style posts with white text listing a handful of places worth going to. That was before COVID. Posts like "Five hot spots in Seongsu-dong" (a trendy Seoul neighborhood of converted factories and cafes) weren't an Instagram trend from the start. Daytrip basically created that culture.

Why they turn down 70 percent of ads

They have about 550,000 followers. There are plenty of bigger accounts, but what he always cared about was not being disposable. People should be able to come back to it anytime, and it should share trustworthy places they've vetted. So they chased sensational virality less. He admitted they gave up some profit and business upside for it.

One number stood out. They turn down about 70 percent of the ads that come in, because the moment they take an ad for a place that's not quite up to par, it shows up right away as unfollows. They know the trust lost is far bigger than what one ad brings in, so they approach each one carefully, and that has actually worked better for them.

But taste is something the people who create it can recognize and still struggle to explain. I asked how they tell what's daytrip's taste and what isn't. His answer was honest. "To this day, I don't think I can explain it in a word." You can't put into one word what's pretty or what makes a place good, but we know it by feel. So what daytrip did was turn the places users all feel are good into its own kind of algorithm. Internally they track a number of factors and have turned into data the fact that when those factors come together, a place gets favored as good. But pointing to a place and saying "that one, exactly" is still something only a person can do.

The talk about metrics was especially practical. On Instagram, 80 percent of a post's fate is decided by the reaction in the first 10 minutes, so what matters is the share rate and engagement in the 10 minutes after posting. What daytrip always focuses on is saves and shares, more than likes. By his account, shares run at twice the rate of likes and saves at twice the rate of shares, so saves come to four times likes. Content that beats that baseline is what matters.

What kind of content gets saved is a pattern we all already follow, he said. Their working hypothesis is that people won't go to the places daytrip posts today or this weekend. Good content gets people to think, I want to go there, I should go, I'll share this and go with my spouse or a friend, how about this place, want to check it out sometime? The type of content changes daily with the season, location and trends, but the goal is just one: get people to save it for later, for when they feel like going.

When an editor's instincts become data

I asked the question I was most curious about, the size of the team. At its peak it was up to 20 people, he said, and now it's much smaller.

The reason lines up exactly with the theme of the event. Before AI, each editor's personal instincts were the asset. Competitiveness came from how many channels an editor followed and how fast they caught new information. Since nobody can visit every place in person, editors pieced together a sense of which places were rising fast and which people caught things early. Now AI does that. There's far more data, it's far faster, and decisions are far easier.

When I asked what the raw material was, he answered in one sentence. "Everything we've posted on Instagram for five years is data." If you also train on the huge number of reactions collected in the app, you can see which places people like and when. Once you can see which places drew a lot of attention, you go out and find more like them.

Another side of the business came out here. Place advertising is a business model, but trend research itself is actually bigger. The biggest data on where people go belongs to the map apps. Where daytrip stands apart is in how it categorizes the high-end demand for places everyone flocks to and places with great taste. Everyday needs, and the step where those turn into spending, are something else again, and that's the next step they're working toward.

When asked where you could see proof of the automation, he gave a fun answer. If you follow the channel, you've probably noticed a lot more overseas content lately. They couldn't source that from Korea before, and now they can much more easily. That's the proof of their AI transition.

He was also honest about what doesn't work yet. They don't have data points or a read on overseas users. They can catch what Korean users look for and where they want to go, so they can predict where a Korean would want to go in New York. But they can't predict yet where Americans will go in New York. Japan is the same. They know what Koreans will do in Japan, but not where Japanese people will go in Japan. The plan is to keep training on data that way, broaden the user cases and expand country by country.

They already run a separate US channel. They gave it a different name on purpose, and in just over a year it reached 280,000 followers. The US actually moves much faster and feels like it has more room to scale, so these days he works on the US business more than Korea.

Here's what I thought at that point. It used to take several editors researching, designing and uploading to curate a single account. If that gets automated, the same resources can localize for Korea, Japan, the US and Southeast Asia. If the area of human taste, which I'd thought was the hardest to automate, has come this far, everything else will be much easier.

Your likes move the world

Finally, I asked what individuals and organizations should each work on.

His answer for individuals was a good one. What trends on Instagram is made by us in the end, he said. "The world moves on the likes you tap and the shares you send." So all you need is to clearly catch your own judgment and the judgment of the people around you. It's not hard, but you have to make an effort. Keep track of how much you use Instagram each day, what you look at there, what you like and what you share, and follow how a few people around you consume content. That gets multiplied by hundreds of thousands. The 200 people in this room each make different judgments, and Instagram just does that in the millions. The conclusion that people will act the way you judge they will can be more accurate than you think, so trust your own judgment.

He added another point I liked. Instagram has actually made the world smaller. With a few searches you can catch the flow and trends of how things move overseas, so there's no need to limit yourself to Korea at all.

Riding along with the customer

The second session was with Jooyeon Hwang, Principal Solution Engineer at Snowflake.

Their career was interesting from the start. They began in IT at eBay, and working with data made them want to handle bigger data, so they moved into Hadoop. That led to working with the Hadoop ecosystem in the cloud, and deciding that public cloud was more efficient than on-premise for big data, they chose Snowflake, where they're now in their fourth year. Coming from eBay, they support companies in e-commerce and companies like Amorepacific (Korea's biggest cosmetics group).

Their description of the solution engineer role was the best analogy of the night. It's software, but to make it easier to understand, think of a car. A typical salesperson explains the car's specs and how good it is. A solution engineer just gets in the car with the customer. They drive together, explaining how to use the turn signal and how to park, and if the customer wants more, they customize it. If the car keeps veering right, they might give a tuning guide that lets some air out of a right tire on purpose so it turns right better. You need to know development, but it's a role with that much contact with the customer.

Three reasons company data is hard

The main part started from a question I'd asked earlier: how much did revenue grow compared to last year? It sounds simple, but even the definition splits several ways. It's July now, so does "last year" mean January to July, or last July to this July? People differ. Change the definition and the report's answer changes. Even for revenue alone, it's unclear whether tax is included or returns are included. And since every team builds metrics that make itself look good, the answer differs between teams too.

Here's how Hwang summed up what makes company data hard.

First, nobody knows where the data is. The data you need might be in Excel or in a database, and even if it's in a database, nobody knows which table. Everyone looks at the BI dashboard, but almost nobody in the company knows where the data behind it comes from. Meanwhile, the big data team that builds the dashboards knows where the data comes from but not what it means.

Second, nobody knows what the data means. Even with a column and table called "sales," there's no clear definition of whether it's pre-tax or post-tax or includes returns, so users looking only at the numbers have a hard time understanding them.

Third, even with those two filled in, what's missing is how to give instructions to an AI agent and how to connect all this data.

Innisfree is stored as 0 in the database

The first step in filling these gaps is giving the data meaning.

The example was nice and concrete. Innisfree is one of Amorepacific's brands, but in the database it's stored as the code 0, and users might search in Korean or misspell the English name. To unify all that into the exact database term, you first need techniques like a dictionary or vector search, and the revenue column needs a definition attached saying that from now on, revenue here means pre-tax revenue.

On top of that come each company's own rules. When someone asks for this year's revenue versus last year's, you usually think January to December, but Amorepacific's fiscal year starts in July. So you have to add up July through the next July to get the annual comparison they want. Only when every company and every dataset has the definitions it needs can you find the answer you want accurately.

Defining each piece of data isn't the end either. To ask more complex questions, the data has to be connected, and that's the area called ontology. A few years ago it was seen as Palantir's unique technology, but now many products support it.

One demo question made it clear why ontology is needed: which product's sales rose the most during a promotion, and what do its reviews say? It's one sentence, but to answer it, promotions have to be linked to products, products to sales, and sales to reviews. You need at least four levels of linked data to get an answer. That's why you don't get an accurate answer if you just throw the tables at it and let it search.

Hwang also introduced features the product adds automatically. One assumes that a table many users rely on is likely the right one and moves popular data up the priority list. Another automatically traces where data came from, helping the system understand the meaning of a question more naturally.

Here's the order Hwang laid out for the whole structure. When a question comes in, first you need the data. Then you need a catalog that defines it, and context made up of each dataset's definition and the relationships between datasets. Only when that context is applied effectively to an agent do you get the answer you want. You can solve it with development, but it's not easy.

The last sentence was the conclusion of the session. The most important thing in enterprise AI is reliability. You first have to build AI you can trust through this complicated-looking work, and only on top of that can you ask it to analyze why revenue dropped or forecast future demand. If you have it repeat that automatically every hour or every day, anomaly detection happens automatically too. Build trustworthy AI first, then put business logic on top. The order is clear.

The hardest thing to find is people

In the conversation that followed, the answer that struck me most came when I asked which of the three problems was actually the biggest, and whether it was a technology problem or a people and culture problem.

Hwang said you could see it as a people problem. The hardest thing to find is the context itself, more than the data. Finding the meaning of the data is hard, and the reason that's a people problem is that it comes down to who knows it. Someone clearly built the system, but most people leave within four or five years, so 80 percent of systems become legacy. Nobody knows why it was built that way. The room burst out laughing, probably because a lot of people were picturing their own companies.

I also asked how client companies react when Hwang first brings up revenue. Ask if they're pulling revenue numbers well, and of course they say yes. And they really are, because they pull them from BI tools or Excel, not from the database. But go one step further and ask if they know the process the revenue number goes through, and most can't answer. Because they don't know how the data flows, even a small change can give a completely different answer.

How they start was a candid answer. Sorry to say it, but they don't rely on the client. They ask for the BI tool or Excel files the client looks at instead, because the dashboard actually holds the most information. Open a dashboard and there are queries and sources, so they trace things back from there. If they're lucky and a legendary person who's been there 15 years is still around, the work goes much faster. If those people have all left, they figure it out piece by piece through reverse engineering.

I asked how they get past the wall of long-tenured employees who might react with, I've been here for years, how well could you possibly know this? The answer had real practical wisdom. Reverse engineering can make clients uncomfortable, so they keep sharing interim results. This used to take a lot of time, but now AI makes it fairly fast, so they build a prototype and show it. There's an order to it, though: you can't show it straight to the executives. You show it to the people doing the work first, get their OK, then take it to the executives, building trust slowly.

AI Ready Data, and a project that bit off too much

Hwang shared the Amorepacific case in as much detail as could be made public.

When Hwang first took on the account two years ago, the company was already ahead of the curve. The project they were running then was called AI Ready Data, not "next-generation data warehouse." They were gathering company-wide data with the goal of preparing it for AI, because that's where things were headed. It was a company that prepared that early, and since last year it's been seeing results, with agents built and launched one by one.

Behind that was conviction from the C-suite that decisions would be data-driven. But the downside Hwang added was the most practical lesson of the night. They were too ambitious and wanted to run AI over 1,000 tables from the start, so the project struggled early on. In the end, they interviewed the client, asked how many tables people actually look at, and cut it down to 100. They split those 100 into five subgroups like marketing, revenue and sales, built a block of 20 to 30 tables for each group, and attached an agent to each block, which got them to the service they have today.

Asked what they were missing most, Hwang answered like this. When revenue drops, people want to know why, and root cause analysis has to use source tables, not aggregate tables. But when you have AI handle aggregate and source tables together, it doesn't come together well. It's less a limitation of AI than the fact that you need a very finely designed ontology to control it, and that takes a lot of time. So right now they analyze mainly with aggregate tables, and the source data side is still in development. Since ontology only became widespread this year, Hwang expects the detailed technology in this area to become more important.

Saying you want AI to raise revenue is like asking for lottery numbers

I asked where a company that wants a company-wide AX (AI transformation) should start, and the answer flipped my expectations.

It's easy to think you should prepare the data first, but it's the opposite. What matters most is what you want to do with AI. Clients often say they want to see AI do something good, or they want to raise revenue with AI, and when Hwang hears that, the thought is: that's the same as asking me to pick your lottery numbers. Using AI without a goal almost never succeeds.

The good examples were ones with a sharp purpose: AI that judges whether an incoming purchase order is appropriate, or AI that checks whether a new product's ingredient list complies with each country's laws. You can only prepare once the purpose is clear, and the goal determines what data you need. So set the goal first and prepare the data to fit it, and the project succeeds quickly.

So how ready are Korean companies right now? Unfortunately, Hwang said, when the C-suite hands down an order like "we should do AI too," the people receiving it come back with vague proposals like "we want to raise revenue" or "we want to find our problems." That takes a lot of time and rarely produces efficient results, and even if the project succeeds, it becomes a system nobody uses. Fortunately, lately about 3 out of every 10 companies come into meetings with some guidelines and goals.

I also asked about the temperature of AI adoption in Korea. When I held a Bloom meetup in Singapore recently, I'd heard that Korea's high AI adoption rate is leading some global companies to put their Asia-Pacific hubs in Korea. Hwang's answer pointed the same way. Every company here has an order or a goal to adopt AI in some form, and the will to actually use it. In Korea, more people are taking it on with the goal of getting results from AI than are afraid of it or rejecting it. Considering that many places in other countries are still conservative, asking whether it can be trusted or whether it's unethical, that's a clear difference.

The business side gets the spotlight, IT does the work

The topic of organizations came up from the question of what it takes to avoid building a system nobody uses.

AI needs help from both sides, Hwang said. Ideas come from the business side, but they don't know how to make them concrete. Even if the business side tries to build it themselves, enterprise AI uses company data, so it's not something you can solve by installing it on your own PC. They need help from the infrastructure team. But the infrastructure team doesn't speak up much even when they have ideas, because it all becomes their work.

The structural problem Hwang pointed out was the core of this part. When the business side gets the spotlight and IT does the work, it's hard to reach a good outcome. So you need an organizational culture where results come from collaboration, and when they do, the rewards get shared again. If that's hard, you can set up a task force or bring in outside help.

An audience question hit exactly this point. They were using AI on the business side, and the company kept pushing everyone to use AI aggressively, while the security team kept reminding them they needed a separate review and approval. The more they tried to analyze data, the more those two demands collided head on. Judging by how many people in the room nodded, everyone is going through this. It's the same problem as the organizational one. The business side, infrastructure and security each have their own demands, and without a structure to coordinate between them, nothing moves forward, no matter how sharp the goal is.

When I asked about the most common misconception about AI, the answer was firm. There's no all-purpose agent. Maybe in ten years it'll really be possible, but not in the next two or three. You need to build agents with a clear purpose, like verifying purchase orders or checking legal compliance, and you can chain several of them to make them more capable. But Hwang doesn't expect an agent within ten years that can answer a broad question like "our company's results dropped, analyze why."

What will definitely work within a year or two, on the other hand, is team-level agents. Even revenue analysis, at the team level, has fairly fixed rules, so it can be put to good use.

Asked what to do about data right now, Hwang put it this way. Even if you don't know where your data is, these days MCP or workspace connections can find it to some extent. So now the purpose is what matters. First figure out what you, your team or your company wants to do, then search backward for the data you need. If you build data first without a goal, you waste far too much money and time.

When know-how in a notebook becomes an agent

Finally, I explained that our community started from a post called Claude Blue and was created to move past Blue toward Bloom, and asked whether, as someone who works with data and AI, there are moments in the AI era when work gets really fun.

The story Hwang told closed the night well. They recently met the CFO of a food company, someone a bit older, who wrote down and organized ideas by hand in a notebook. One day Hwang saw it and suggested turning it into files. When they built an agent from that content, the CFO's knowledge became the agent, just as it was.

"If you organize the ideas you have well and can bring them to life with the power of AI, you can make a big contribution to your company, and you might even get the chance to start your own business." A session that started with data ended with an individual's accumulated know-how. Whether it's company data or a personal notebook, I heard it as: AI can't use context that isn't organized, and organized context turns straight into an agent.

Groups of six, the heart of Bloom

As always, the last hour was the roundtable, and this is the real heart of every Bloom event.

That night we set up groups of six. After trying it many times, six turns out to be the right number. Not too few, not too many. We give a question, each group discusses it, and the group leader comes up to share the insights from their group.

This overlaps with what Seokho said in the opening: Koreans think of debate as a shouting match and never had a chance to learn it. But the people who come to our events aren't only curious about the speakers. They want to tell their own stories too, and they come hungry for things you can't find on the internet. So we made space at the tables to share raw insights that aren't online.

That was especially true on a day like this, when the talk was about data and context: whether anyone who knows that context is still at your company, whether your team's goal is sharp, whether you can explain the flow behind your dashboard. Among coworkers, you might be careful about revealing what you don't know, but here everyone is a stranger, so you can talk honestly and take the ideas home.

In short, the night's answer comes down to one thing. When AI doesn't work, the usual reason is that our own data has no meaning attached, more than anything AI can't do. Five years of Instagram posts became data that let daytrip shrink its team and expand overseas. A CFO's know-how in a notebook became an agent. In the end, AI is a data problem.

Originally written in Korean on Brunch, July 29, 2026.