FDE Might Not Be the Answer

It's been almost a month since I started running Claude Bloom. We held our first Anthropic Claude community event on April 14. I've been thinking hard about what should set our community apart. I'm not a developer, so I can't cover the technical side, and there are already plenty of great communities, influencers and content out there. So my approach is simple: I invite speakers on topics I'm personally interested in, and now and then I write a blog post or make a YouTube video.

What interests me most is news from abroad. Korean news is easy to come by, since group chats and social media already filter it into Korean (and Koreans pick up news really fast). Overseas news is different. Plenty of it spreads online, of course, but you almost never hear the inner-circle stuff that gets passed around offline. As you know, by the time something shows up on X or in the news, Silicon Valley has already picked it clean. It isn't under NDA anymore and it's basically a public good.

What is Silicon Valley's inner circle actually worrying about, what are they thinking, what kind of lives are they living? Not knowing won't hurt my ability to make a living one bit. But I'm curious all the same.

As it happens, I have a younger schoolmate from high school who built the predecessor to Gemini at Google DeepMind, then worked at xAI as a manager over 10 people, dealing with Elon Musk at least three times a week and sometimes seven days a week. Now they've joined Thinking Machines Lab, the AI frontier lab founded by former OpenAI CTO Mira Murati, as an early member. They're definitely younger than me, but their career is honestly top tier even by Silicon Valley standards. They happened to be in Korea for a short trip, partly for a health checkup and partly for personal reasons, so I grabbed them right away as a Claude Bloom speaker, haha. As an engineer they hadn't had many chances to speak at events like this (xAI is so security-conscious that they say they're still a little nervous about it), and since this was in Korea, people in the US probably wouldn't follow what got covered anyway, so they kindly agreed. Of course they can't really talk about anything internal. They'll walk through it based on public information plus their own thoughts and thinking.

We covered way too many topics, but the one I personally care about most isn't the frontier companies' grand narratives. It's how American companies actually make money with AI.

The concept getting the most buzz in Korea right now is the FDE (Forward Deployed Engineer). A lot of startups these days pair a PM with an engineer and send them into a large company to fix on-site problems with AI. Put simply, it's contract staffing.

But according to my friend, FDE isn't actually a very efficient model. Cost aside (and these firms bill project fees several times what the deployed engineer is paid), just telling a client to change everything wholesale is very hard to communicate and very hard to sell. No matter how hard leadership pushes from the top down, it ends up meaning existing staff have to change how they work, or a lot of them have to go. Between regulation and everything else, overhauling a company's fundamentals is extremely hard. So with FDE, the engineer has a hard time, the client's decision maker who ordered the work has a hard time, and the people inside the company feel like some outsider just showed up to torment them. It's hard for anyone to get along.

So after leaving xAI, my friend started toying with the idea of launching an AI roll-up with some friends (BZCF has a good explainer on this). Simply put, private equity firms raise money, scoop up companies in very outdated industries, run hard AX (AI transformation) on them, and then either sell them again or pull out cash at much higher margins. The difference from FDE is that you aren't serving a client. You buy the company outright, so you can run it however you want. Of course, that only works if you have someone with deep pockets.

In the US, though, Anthropic and OpenAI are already doing this through joint ventures with global private equity firms and top consultancies like McKinsey. The concept existed before, but now the frontier companies are stepping in themselves. It makes me wonder whether they plan to take the whole pie.

Then again, they'll go after enterprise companies first. So startups could either team up with them (as a subcontractor's subcontractor) or target small and midsize companies outside Seoul (under 50 billion won, roughly $35 million, in revenue) and try running AX on those. Korea is a manufacturing powerhouse, and there must be a lot of factories in the provinces. The problem is that finding a fund that shares this vision probably won't be easy. In any case it's a risky playbook, and there's no success story in Korea yet, so some people might decide it's more comfortable to just buy a building on Teheran-ro (Seoul's main business and startup strip).

Anyway, talking like this, I find the frontier labs' views on humanity's future shocking, but I'm reminded again that even at the business level, the gap in vision and execution is huge. Teheran-ro has its own glitz and a kind of megacity polish, but it made me wonder why it's so hard for us to dream bigger and act on it, and what's holding us back. It gave me the kind of energy where I want to just go for it myself. Some people might say this is a fantasy, a story that doesn't fit someone like me. But when a big dream breaks, the pieces are big too.

Oh, and here's the article from the pre-interview I did today with BZCF (Biz Cafe, a Korean business media outlet): A conversation with Sung-sik Kim, engineer at DeepMind, xAI and TML. Sung-sik is Korean American, and while recently in Korea, they spoke candidly about what they've experienced firsthand at frontier labs.

1. From Google to Cruise, then to xAI

Sung-sik's career hasn't been a straight line. They built the predecessor to Gemini at Google DeepMind, then moved to the self-driving company Cruise. The reason is a bit unusual. Before getting married, they wanted to move to Chicago with their wife. They moved quietly without filing an official request with the company, got found out, and ended up having to go back to San Francisco. This was right after the proposal.

At that point they moved to Cruise, which allowed remote work. Cruise was GM's self-driving subsidiary, and a month or two after they joined, GM cut off funding. The company basically shut down. That's how Sung-sik ended up at xAI. Looking back, they said it was lucky the company went under when it did. If it hadn't, they never would have gotten into AI post-training.

2. Inside xAI

Daily life at xAI is hard to sum up. Five hours of sleep a night, no days off. From 10 a.m. to 6 p.m. it was nonstop meetings. After dinner, from 8 or 9 p.m. until early morning, they worked on projects alone. Going back and forth between home and the office was such a hassle that on many nights they just slept at the office.

Sung-sik was a manager. It wasn't a role they asked for. They said the founder just put them in it without a word. At the peak they managed as many as 14 people. They'd get through interviews, one-on-ones and team meetings all day, then go back to hands-on development at night.

Elon Musk met with the team in person often. When things were busy there were face-to-face meetings almost every day. Sung-sik had heard Tesla and SpaceX were run in a similar way, and xAI was the same. After keeping up that pace for a while, they burned out badly. After that they moved to Thinking Machines Lab.

3. What Thinking Machines Lab is building

Thinking Machines Lab is a frontier lab co-founded by former OpenAI CTO Mira Murati and John Schulman. Sung-sik joined as an early member. (After it was founded, it raised a seed round of about 3 trillion won at a valuation of about 18 trillion won, roughly $2 billion at $12 billion. The founders' names make that possible.)

One of the things they're working on there is GenAI UI. Instead of AI spitting out its answers as text only, it produces them as interactive, rich web UI. If the result of a user's request becomes the interface itself, the meaning of today's app and OS structure changes. You'd no longer need to install and open separate apps at all.

If this direction actually gets built, we'd end up in a world quite different from the kind of software we picture today.

4. The day models train models

When news broke that Elon would provide compute to Anthropic, a lot of people around Sung-sik were surprised. So was Sung-sik. Elon hadn't exactly been a fan of Anthropic on Twitter, and suddenly he changed course.

Sung-sik's read was closer to a technical calculation. As coding models keep getting better, at some point a model reaches the level where it can train itself. Past that point, competing becomes practically impossible. Whoever gets there first takes the game. If Elon judged that Anthropic was getting close to that point, investing and joining their side is the right move. As Sung-sik sees it, it's the result of that calculation more than any feeling.

5. How do you define AGI

Everyone defines AGI differently. As Sung-sik put it, Sam Altman sees AGI as the level where AI can replace almost all the economic activity humans do. Demis Hassabis sets the bar higher. Given the same information Einstein had in his lifetime, it should be able to derive the theory of relativity on its own, and it has to meet the baseline of humanity's top experts in every field to count as AGI.

Scaling laws came up too. Adding more data and compute raises performance, but the returns come on a log curve. Past a certain point, the return on investment drops sharply. Nobody knows where on that curve we are right now. The three of us sketched curves by hand, arguing whether we're here or there, but there was no answer.

Sung-sik's personal definition leans toward autonomy. Even with coding agents, there are still points where a human has to be in the loop. In that sense, they think we're still in the early days.

6. AI doesn't close the gap

You often hear that AI is a tool for democratization. Sung-sik sees it differently. AI is an expensive technology right now, and running multiple agents or turning them into a business takes capital. Token usage follows a power law too. A handful of users burn billions of tokens, while most people use the free version a few times a month. It may be a technology that lifts the people at the top even higher, rather than one that narrows the gap with everyone below.

One-person unicorns came up as well. Sung-sik pointed to funding structure as the difference between Korea and the US. A unicorn is about market cap, and market cap ultimately comes down to how much funding you raise. In the US, a company started by two or three Stanford freshmen really can get a $500 million valuation. Standard Intelligence is one example. In the middle of the AI investment frenzy, that market makes it possible.

7. Korean intensity vs. Bay Area intensity

We compared Korea and the Bay Area. The Bay is intense too. Maybe even more so. But the direction is different. In the Bay, the intensity is about getting better and climbing higher. In Korea, a lot more social status is mixed in.

Someone in Oklahoma doesn't try to copy a New York lifestyle. It's far away and they don't see it directly. Korea is different. Gangnam is a subway ride away. You can see that a Han River view apartment costs 7 billion won (about $5 million). When it's close enough to compare, you compare.

8. Humanity's brain, and humans as the front end

When Elon bought Twitter, he said he was buying humanity's agora. An LLM, then, is humanity's brain. Each of us has our own brain, and an LLM is another brain that runs on compute. Past a certain point, it runs self-recursively across every domain, and hardware becomes the muscle for that brain. Whoever builds that brain first wins it all.

The BZCF team shot back: "Humans are the front end." Our experiences, even this conversation we're having right here, are ultimately a dataset. Hand that context to a model and the model processes it. Seen through that structure, humans are just the front end.

Along those lines, Sung-sik brought up an idea they'd sketched out on their own back at xAI. xAI had built an AI companion service, and they said the voice model and animation quality were good. They thought building K-pop avatars on top of it and taking them into the entertainment market could be a huge market. They never got to execute it, but they said they still think the direction is right.

9. Just buy the company

I was struck by the way the BZCF team laid out market strategy. There are two ways to make money. Be number one in a hot market, or be the only hot player in a lukewarm market where nobody else is. In tech, if you're not number one, it doesn't mean much. Otherwise, you're better off taking the technology somewhere no one has gone yet.

That led to the limits of FDE. The assumption that the client will change is itself unrealistic. Regulation, internal resistance and competing interests are all tangled together. The approach Sung-sik and the BZCF team both landed on was the PE model. Instead of consulting, you buy the company. Once you acquire it, you can change it however you want.

I've actually met people who showed interest in this kind of investing. But it's hard to knock down, in one go, the argument that you can buy a building on Teheran-ro and collect rent, so why go through the headache of AX. Still, it's also clear that this opportunity isn't a one- or two-year game.

10. Wrapping up

It was a long conversation. Sung-sik was severely burned out at xAI from the first day to the last, and now works in a more open research environment at Thinking Machines Lab. They stepped back from xAI's extreme pace and found a place where they can actually think. Someone like that coming to Korea briefly and having this conversation is a scene that's only possible right now.

The world as seen by someone who has worked inside the frontier and the world as seen by someone in Korea trying to do something with that technology run at different temperatures. Creating spaces where people can feel that temperature difference firsthand is one of the things Claude Bloom wants to do.

Originally written in Korean on Brunch, May 12, 2026.