AGI Is 99.9% Here (No Exaggeration)

Starting today, we're doing member demos

Bloom usually works with companies, but last week and this week we brought in people from our own community instead. For anyone new, here's one thing that sets our events apart: we almost never do keynotes. When people watch a keynote, they're bored and checking their phones within five minutes. Their eyes light up when questions start going back and forth. So whenever we can, we go with a fireside chat. I bring a list of questions, but I mostly ask whatever I'm curious about in the moment and let it flow. It's a bit rough and casual, but that's what keeps it alive.

This was also the first night of something new: member demos. The idea came from a suggestion on our Discord. In our small-group meetups, a lot of people turn out to be building side projects or their own products, but they have nowhere to show them or find their first users. So we made a spot for it right away. At the start of the event, two teams get exactly five minutes each. Think of it as an elevator pitch for what you're building. It's less about promotion and more about sharing what you made and where you're stuck.

How do you get it to think of what it can't think of

The first demo came from a developer who works on infrastructure by day and builds side projects at night.

https://liner-chat.vercel.app/

Liner Chat

A conversational curator that digs through your own Spotify library, following a song's musical lineage (its liner notes) to find its relatives.

https://liner-chat.vercel.app

The starting point was fun. SHINee's "Lucifer" and "Dokdo Is Our Land" (a patriotic Korean song every kid here knows) blend together perfectly. The beat and the chords line up, so they just fit. Your ears catch that in half a second. So the idea was to build a service that figures out why two songs fit. You put in one song, and it shows you four groups: songs that were its roots, songs from the same era, songs that inherited from it, and distant lookalikes. Then it explains in plain sentences why they sound alike. The developer's thinking was that as long as the connection can be explained, you can jump across eras and genres.

But the heart of the talk wasn't showing off the service. It was about where things got stuck. The first problem was the "goes-with-everything song." No matter what the seed was, the same song kept coming up. Portishead's "Glory Box" showed up 9 out of 9 times when they tried nine different songs. So they tried scoring similarity, but the score wasn't tracking similarity. It was tracking which song came to mind first. The number one result was supposed to be the most similar song, and instead it was just the first one the model thought of. The way they checked this was impressive. They called the same model separately, purely as an evaluator, and it gave its own picks 2 out of 10. So they split the scoring into a separate call, and had the judging call see only the song information and never the explanation written by the generating call. The song that had shown up 9 out of 9 times dropped to one sixth of its previous exposure.

The second problem, they said, is still unsolved. The project started because their boyfriend said that whenever he listens to a certain Tame Impala song, he ends up wanting to hear a particular John Lennon song. The developer had a similar experience. They were listening to a cover song, and someone in the comments said it sounded like Radiohead's "Creep." Different era, different genre, different language, but to human ears it sounded similar. Yet when they put that song in, "Creep" never comes out. It doesn't fail verification. It never even gets suggested as a candidate. If you give the model a song title, it describes the chord progression accurately. But if you give it the progression and ask what songs use it, it failed to come up with "Creep" three times out of three.

Their summary was one of the best lines of the night. Recalling things is obviously something the machine does better than they do, since it has so much more data. But they still haven't found a way to get it to think of what it can't think of. The judging setup they fixed earlier only works within the candidate list. If something isn't on the list, it doesn't matter whether you separate the scoring or plug in a stronger model. Right now the only option is to write directly into the prompt that this kind of song goes with that kind of song, but that only feeds in answers they already know, so it has limits. They wrapped up by saying that if anyone has solved this problem, they'd love to talk. That was exactly why we created the demo session.

Turning a five-minute appointment into thirty seconds

The second demo wasn't on the schedule. One of the two teams that signed up couldn't make it, so we handed the mic to someone who raised their hand on the spot. They're the CEO of a company that supports medical workflows. They'll get their medical license next year, and they've also been developing software for a long time. They figured they could translate between the two languages: where medical staff struggle, and how that needs to be explained to developers so it becomes a product. So they started a company.

I liked how they explained it with numbers. In a three-hour morning outpatient clinic at a university hospital, a doctor has to see 30 patients. That's five minutes each. You have to get up to speed on the patient and finish the consultation in five minutes to get 30 patients through in three hours. The famous Korean "five-minute appointment" falls right out of basic arithmetic. The higher-ups say see more patients, and you also have to see them properly. So they built a tool where you feed in previous medical records and it summarizes 50 patients, organizing each one so a doctor can review them in 30 seconds. It even summarizes what needs to be checked that day, and when writing the chart, the objective findings section can be copied straight over.

They were honest about the hard parts too. Every doctor writes records differently. Some write "diabetes," some write "DM," some write "T2DM." Standardizing those terms is where they keep getting stuck. It was exactly the same point we heard at last week's session: the name people use is different from the value stored in the data. Only the domain had changed, to medicine.

One day before funding, 18 teams in the same cohort stopped

The first fireside chat was with Brian Jin, a Korean Canadian founder visiting Seoul for the first time in four or five years. Brian moved to Canada after the 2002 World Cup, finished high school and college there, and studied computer engineering at the University of Waterloo. After graduating, Brian worked in data the whole time, but had been starting companies almost since graduation. I asked what the startup scene in Toronto is like, and the answer was clear. Think of Canada as a whole as a testbed for the US. Even at Brian's own university, people who want to start companies basically go to the US, and there are a lot of cases of succeeding in Canada and then moving to the US. Given that everyone in Korea is aiming for the US these days too, being right next door to it definitely sounded different. But Brian's take on Canadians was funny. They're lazy. They don't do "hurry, hurry."

Brian applied to Antler Canada. It's a program where you join as a solo founder, they match you with a cofounder inside the program, and if things go well they invest. Even getting in was tough. Brian did four interviews over three months, and when Brian later asked why they'd extended an invitation, the answer was that Brian came across as steady and relentless. The program runs ten weeks, and in week six, 70 percent are cut. I asked whether people leave on their own partway through, and Brian said no, they cut you. They tell you it's time to go. I asked if it was like "Show Me the Money" (a Korean rap survival show), and Brian laughed and said officially they say it's not a survival competition, since you're not competing against each other. Brian also pointed out what makes Antler different. At first they don't look much at things like distribution. They look at people. They care about founder market fit more than product market fit, about the synergy of the founding team, and about whether you're someone who can be taught. Along the way, Brian met an Indian cofounder who had been selling ERP software.

Brian was candid here. Brian is a technical founder, not a business founder. Brian said the line from last week's meetup about being an introvert who has to act like an extrovert hit hard. Brian really hates the business side and spends ten times the energy on going to market itself, because it doesn't come naturally.

The product they were building was interesting. When a company buys a new ERP or switches to a different one, for every 1 dollar spent on licenses, it costs 6 dollars to actually roll it out. That covers scoping, configuration, data migration, support and training. There are a lot of AI-native ERPs these days. With something like Oracle, rollout takes three to six months, while an AI-native ERP takes about a week, even if it doesn't have as many features. The implementation partners have their own reasons too. They don't work at a fixed price. They quote 100 hours, but it often takes 150, and the partner has to eat those 50 hours. So everyone wants to do fixed quotes, but technically it wasn't possible. Brian's team wanted to automate everything from scoping to rollout with agents. I asked whether Oracle would hate that, and the answer was spot on. Oracle headquarters doesn't make much of its money on rollouts. Its partner network does that work, so the customers aren't Oracle, they're the partners. Even if a partner goes from earning 6 dollars to 4, they can cut six months down to a week, so their margins actually improve.

So I said Brian must have made it all the way to the final investment committee. The answer was short: "We didn't." In May, Antler's general partners from each country held a global summit in San Francisco, around week nine. When the partners came back, the mood had suddenly changed. Even if you get funding from Antler, you absolutely have to raise a follow-on round within three or six months, and the VCs in San Francisco had said they're no longer investing in vertical AI. "Even if I invest in you, San Francisco won't. What are you going to do?" That's what Brian heard. Eighteen teams in that cohort that were doing vertical AI stopped all at once.

If there's no technical moat, what is the moat

So which teams survived? Two out of 20. One was building lab tools with sensors, and they had a patent. Two people who couldn't do anything like sales, in Brian's words just nerds. They still got funded. The other was a marketing AI agent team led by a star founder. But Brian heard this team is now pivoting too, even though they raised in June. They fell into the same vertical AI trap and survived on the strength of the founder personally. Either you had a patent or you had the person. No team survived on the product itself.

Then I asked what San Francisco VCs are interested in now. The answer: anything that makes money and can grow fast. Vertical AI didn't die. In the end, distribution is the problem. Those people want a 10x return. And there's one thing they kept hearing from the very start of the Antler program.

There's no technical moat anymore.

When they asked what the moat was, they were told to go find it. I couldn't help laughing. That was exactly the topic of last week's session. What will you do when OpenAI releases its next version? It'll be copy-paste. Brian was frank about the relationship with VCs, too. You expect that since this person picked you, they must like you and will teach you a lot. None of that. Brian said it's brutal. They don't give much advice, and every week they ask one thing: how many design partners or letters of intent did you get? Even when you give them a number, they ask how you found them. They don't like it if you found them through your personal network or connections. You have to make it happen using every means and method available.

Here's one of the things they told the founders. Startup founders, especially business founders, have to become influencers. You need a celebrity mindset. Put your face out there a lot, take part in the marketing yourself, go to lots of events. And Brian added that this is exactly what a tech founder like Brian hates most. So I asked whether the standing of tech founders has changed, and the answer stung. It used to be 50/50. If a founding team didn't have a CTO, you couldn't even consider it a tech company. Now it feels more like 80/20 or 70/30. The funded teams Brian knows don't have CTOs. In marketing tech, there's a founder who used to be a CMO, there's a salesperson opening up the market, and the person in charge of product builds everything using Claude and OpenAI. From an engineer's point of view, that's unfair. Building a production-level product isn't easy, even with AI. But from a VC's point of view, that's not what matters. You can hire for it later.

So I asked what Brian plans to do next. Right now it's open source, for two reasons. Tech is what Brian does best, and open source is a distribution channel in its own right. That reminded me of last month's meetup. An engineer who'd spent nine years at Google said exactly the same thing. The reason they wanted to go abroad was to work with those kinds of people, and with open source you can start doing that tonight. Completely different setting, completely different person, same conclusion.

The open source project is called the Judgment Pack Specification. It started from two places. The first is the business side. In the end, every company runs through the judgment of its decision makers, and the most important basis for that judgment is know-how. If you pour all of that know-how straight into an AI, the company's intellectual assets eventually disappear. On top of that, in North America companies these days want to use open-weight models internally a lot, and you can't keep rewriting your prompts and skills every time the model changes. So the idea is to pull the basis for judgment out of the AI model entirely. The second was a technical reason, and I liked this part. Brian doesn't think LLMs have intelligence. They just have general knowledge, and agents are what create the effect of seeming intelligent. So why is coding where AI agents are used the most right now? Because coding is a closed environment. There are linters, compilers and unit tests, so a machine can judge whether the result is correct. You can't hand a company's important decisions to an agent that works probabilistically. So can you build a testable environment for business agents too? This project came out of those two considerations. Brian also mentioned that 95 percent of internal agent projects at companies end at the MVP stage, because you can't make decisions probabilistically.

Why a Toss developer is getting into YouTube

The second session was with Jooeon Kim. Jooeon works in DevOps and backend at Toss (a major Korean fintech company), and originally built systems for training, managing and retraining machine learning models. Right now Jooeon is getting ready to start their own business. When I asked whether it was okay to say that publicly, Jooeon laughed and said they hadn't even thought about it. I said the people here might end up being future customers or partners.

Jooeon has been with Bloom since the early days, starting with our second event. I asked why they keep coming, and the answer was simple. Jooeon was about to go on leave and knew that once the leave started, they'd stop seeing people, get discouraged and feel down. So Jooeon was thinking about how to meet people when they saw a post about Bloom on LinkedIn. When they came, the level of the attendees was high, the agenda was good, and the fireside chat questions didn't feel scripted. They felt human. Honestly, those questions are just me asking whatever I'm curious about at the time. But I figure if I'm curious about something, you probably are too.

Function, power, ownership

I set up this session because of something Jooeon brought up at a small-group meetup. As a developer, Jooeon is thinking hard about how to survive in the AGI era, and isn't looking for the answer inside development. First, we went over the definition of AGI. There are two main camps, and OpenAI's bar is on the lower side. They define it roughly as automated intelligence that outperforms most humans at most economic activity, and by that standard the dominant view is that we've already reached the early stages. Jooeon follows the more conservative standard, Demis Hassabis's. Can AI surpass the best of human intellect? If an AI had the knowledge Einstein had, could it come up with the theory of relativity on its own? There was recent news that some model solved ten hard math problems, but Jooeon said if you look inside, it's closer to filling gaps in the logical structures humans have built up so far. Tasks that are hard for people but easy for machines. The consensus is that there's been no emergence at the level of a paradigm shift.

The key test Jooeon laid out was clear. Today's AI, once a person gives it a task and a direction, pushes on it for a very long time and reaches conclusions faster than a person. But can it set its own direction when no one gives it one? Jooeon sees that as the line that will define AGI. When I asked when that will happen, the answer was 2030. Three to four years from now. When asked whether we're headed for utopia or dystopia, Jooeon has always said dystopia. Jooeon is a pessimist by nature, and as a devout Christian, that belief is part of the foundation too.

Curious about what will be left after AGI, Jooeon went through human history. From hunting and gathering to agriculture to the Industrial Revolution, they looked at where value is produced and how that value spreads. There turned out to be exactly three ways an individual can produce value: function, power and ownership. Function is going to a company, selling your abilities and getting paid. Power is the ability, based on control over resources, to decide who gets assigned which ownership rights. Ownership creates value by lending out or letting others use what you own, backed by the ownership rights that power guarantees. Then came the high point of the night.

The essence of the AI era is replacing function. The reason we're afraid is that we don't own much.

Someone who owns a big building in Gangnam has nothing to fear from AI. AI can come, and my building is still my building. If you already have power and ownership in this era, AI is just something that replaces functions you weren't doing anyway and cuts your costs. I'd never heard anyone explain where the fear comes from that cleanly.

So Jooeon broke ownership down further. The first kind is physical ownership: land, materials, resources. The second is intangible ownership, which splits in two again. One is ownership defined by law: patents, trademarks, legal tender, securities, data. The other isn't defined by law but clearly exists. It's owning a place in people's minds. What is a brand? It's the image of that brand etched into consumers' minds. So owning a brand means owning a space in the minds of a large, anonymous crowd. Owning a platform means owning people's perception of that platform, and creating meaning through art means buying a space in someone's mind rather than selling a function or an object. Listening, I thought of Hermès. Will Hermès disappear in the AI era? If anything, the more knockoffs there are, the more its value goes up.

Of all the kinds of ownership, Jooeon picked network. With nothing to work with for chasing power, they chose the kind of ownership that looked most within reach. So the first goal is to own their own personal media. Grow it around YouTube, build deep emotional connections with subscribers, and form relationships where both sides trust each other and find it meaningful. Jooeon thinks that will be the biggest asset of the future. Who owns a media audience, and how close they are to that audience, is a fortress nobody can easily take. I agree with this.

I asked whether Jooeon had ever done YouTube before. They had, around college graduation, and it ended up becoming a Christian channel. The channel's catchphrase back then was "Your thoughts make you." When Jooeon asked the question "Who am I?", the conclusion was that their deepest essence is what they think. The first gate through which people perceive the world is the senses, but senses are just electrical signals. Interpret them once, and you get emotions. Thoughts and emotions create desires, desires create actions, actions create habits, and habits create a life. So if you want to change your life, you have to start by changing your thoughts. Then I asked how much of their development knowledge and skills they'll use in the new media work, and the answer was refreshingly blunt: "I don't think I'll need much of that." These days, having built up knowledge isn't much of a strength. Undocumented, hands-on experience would matter, but the development knowledge Jooeon learned is all documented. I asked if they had the readiness and courage to live a completely new life, and Jooeon said they were never all that passionate about development. Development was what they did to put food on the table, and what they were really interested in was something else.

What comes after functional identity

An audience question delivered the last twist of the night. It came from someone who works as a marketer at a global beauty brand. They majored in philosophy and said marketers are feeling a sense of crisis too. The question went like this. Everyone runs to get into college, and after graduation we live as if there's a set route. Won't that kind of education and that way of life change a lot from here on?

Jooeon's answer was long, and it was good for that reason. Society became centered on college as the Industrial Revolution matured and work shifted from physical labor to mental labor. As the quality and number of physical labor jobs dropped, the belief spread that you had to do mental labor to survive, and a society that worships college was born. Jooeon brought up something they'd overheard on the subway. These days, among moms in Gangnam (Seoul's wealthiest district), it's apparently considered important for kids to start building up a GitHub account in middle school. Never mind undergrad grades. What matters now is "here's what I've done, here's what I've started." That's where this idea came in. Right now we're in the era of functional identity. People introduce themselves by their job. So when their function gets replaced, their identity gets shaken too.

So what comes next? The most hopeful version people talk about a lot is mission identity. Since direction is something AI still can't do, the history of "this is the problem I'm going to solve, and here's what I've done so far" becomes your storyfolio, the story that explains you. But Jooeon thinks this has five years left too. Because AGI aims for general intelligence, it has no particular character or direction, but once general intelligence is achieved, Jooeon expects a lot of AIs with very distinct character. It's like clothing: when clothes were expensive to make, everyone wore similar things, and once they got cheap, everyone started wearing different things. At that point, setting a direction can't be your unique identity either.

What becomes important in that era is belonging identity. Historically, people who lost their function all came back to belonging identity. They fell into nationalism, joined groups or turned to religion. People need to feel their own worth in some form, and belonging is the most fundamental identity, one that goes back to prehistoric times. That's why voluntary communities will matter most in the future. Places where people don't just say hi and leave, but take a genuine interest in each other and help when someone's going through something hard. Jooeon thinks a community should raise that floor, so that just belonging to it fills part of the meaning in your life.

Honestly, this is where it hit me hard. I work at an AI startup, and community isn't my actual job. It just sort of came together, people kept showing up, and it was fun, so I kept going. But I've always had the question: why are we doing this, and why do people keep coming? I'd planned to stop if people stopped coming, but they keep coming. You could explain it as FOMO, or as people chasing a trend. But what if, as the functional parts of being human fade, people need to find self-worth and a sense of belonging somewhere, and that's why they naturally gather in communities? I'd never thought about it that way.

What each of the six tables came up with

The roundtable had two questions. Do you think AGI is already here? And in that era, how do software companies survive, and what should individuals build up?

The common answer was that low-level AGI is almost here and high-level AGI is still far off. There were other views too. One pushback was that, as Yann LeCun has said, you need physical AI before you can call it real intelligence, and that without emotions there's no motivation to want to do anything, so how would an AI ever set its own north star? The most interesting example was a case where someone analyzed the movements of an F1 pit stop crew and applied them to the operating room. The view was that if AI can make cross-domain leaps like that, that's real general intelligence. The navigation analogy was good too. People used to look at maps and trust their own judgment, but now they trust their navigation app more than themselves. We don't seem to do that with AI yet, and the moment that flips might be AGI. An attendee who's a doctor said that within medicine, people already see AI as clearly better than an average doctor, and everyone sees the possibility of physical AI getting better than people at surgical procedures. But even that comes down to trust in the end, the question of whether we'll hand responsibility over to AI.

On software companies, most agreed that the era of showing off features is over. One idea was that most companies will end up going the way of entertainment companies. One person said the essence of a software product hasn't changed: companies that make good products well still sell, the cycle has just gotten faster. One table converged on the view that AI alone isn't sustainable, and businesses tied to something physical, like hardware or offline, have the better long-term outlook. The example was fun. Entertainment agencies are trying to stack something more on top, since an individual's IP can now be quickly cloned with AI somewhere else, and so one agency is expanding into a robotics business. They also concluded that plenty of clients still confuse AX (AI transformation) with DX (digital transformation), so software companies won't disappear any time soon.

One table even came back with six things individuals should build up. The ability to use AI tools. And since using AI will wear down our social skills, we should actually work on our humanity and social skills. Companies with lots of raw data will have the advantage. We need to develop active thinking. There used to be praise for people who were quiet but good at their jobs, but now the people who survive are the ones who keep asking questions. And aesthetics and design. AI can't taste and can't see with its own eyes, so people have to keep developing that sense.

The sharpest point came from a table made up entirely of people in fashion and beauty. While we argue about whether AGI is here or not, the world is already changing. AI writes the reports, the proposals and the code. But the truly scary part isn't AI doing everything. It's people no longer thinking. We used to spend a lot more time asking "Is this right? Is this the best way to get the best result?" Now it's "AI says so," and that's the end of it. So they suggested changing the question. Instead of asking whether AGI is already here, ask whether we're already acting like we're in the AGI era. For software companies, "we use AI too" isn't something to brag about anymore. Everyone uses it. So "what problem are you really good at solving?" becomes the most important thing, and someone who really knows beauty becomes even stronger once they add AI. Their answer about individuals was the line of the night: the people who ask questions will survive. When you hand a task to a new hire, something comes back in thirty minutes, and a lot of the time it's a report that looks like AI wrote it no matter how you read it. Then there's no reason to hand it to a new hire, and eventually companies stop hiring new people. So the most important skill becomes continually asking "Why did it come out this way? What's the basis for this?" They closed with this:

From now on, the ability to struggle through something will be rare.

The last table's take was the most fundamental. They said even if AGI arrives functionally, we probably won't recognize it. The reason is that people don't have the ability to explain what they want. A lot gets lost when people put their thoughts into words, so unless something comes along that solves that tacit knowledge problem, functional AGI won't really count as having arrived. Someone who works at a game company said something good too: we don't know what we want, but the companies that can suggest what we want will survive. The table leader's closing analogy has stayed with me. Back when TV was the only medium, we all influenced each other in faint colors inside one huge circle. Now there are tens of thousands of small, deeply colored circles. Owning one of those circles might be what individuals should aim for. It was exactly the same conclusion as Jooeon Kim's talk about ownership, even though it came from a different table on its own.

Two people were saying the same thing

One person learned firsthand, through a funding deal that fell apart, that there's no technical moat. The other went looking through history for what's left when function gets replaced, and laid it out. And the two ended up in the same place. Brian heard at Antler that founders need to become influencers, hates that more than anything, and so decided to use open source as a distribution channel. Jooeon sees that after function, what's left is power and ownership, believes network is the one they can grab, and is preparing personal media. Both were talking about the ability to reach people, rather than the ability to make things. About the space in people's minds, rather than the product. And the conversations that started out differently at the six roundtables ended up gathering in a similar place: people who know a domain deeply, people who can ask questions, people who can struggle through things, and people who have their own small circle.

The venue that day was sponsored by Neosapience. The truth is this event came together in a big rush. The most important thing when you're putting on an event is the space, and two days before, I posted on Discord asking if anyone could offer a place in Gangnam. They reached out right away. That's how the night happened. Neosapience makes Typecast, an AI voice service, and also runs AI character chat and conversational voice agents. They said anyone interested in content automation or voice AI agents should feel free to get in touch.

Bloom happens every week. Next week there are three events, and the week after, three more. Lately, oddly enough, overseas companies have been reaching out first with cold emails. Things like, "We're coming to Seoul in September, want to do a meetup?" I don't know how they find us, but I try to say yes to as many as I can. And if you have an idea for something you'd like us to try, I'll do my best to make it happen. One thing I've been thinking about lately is renting out an entire cafe and just doing coffee chats there all day long. I actually pitched it to one place, but they close early. Send me ideas and I'll try to make them real.

Originally written in Korean on Brunch, August 20, 2026.