Claude Bloom × AB180: Google DeepMind, xAI, and Thinking Machines Lab

Today I held a Claude Bloom meetup at AB180 with Sungsik, a younger schoolmate from my high school. I've written about Sungsik on my blog a few times already. At a young age, Sungsik has racked up a lot of experience at Silicon Valley's top frontier labs: Google DeepMind, xAI, and now Thinking Machines Lab.

Even after running a lot of these events, I get nervous every time. There's a new audience each time, and I can never predict what will come up. I also tend to make eye contact with people in the audience one by one, and if anyone looks even a little bored, I get tense too.

During my opening today, people seemed a bit bored, so I said we'd get right to it and started the fireside chat. We went back and forth through the questions I'd prepared, and before I knew it more than an hour had gone by, which threw me for a second. But when I looked around, I saw about 120 people, all the way to the back row, completely locked in. That's when I relaxed: OK, the topic and the speaker's insights are keeping people interested.

Then we did small groups, which are Claude Bloom's signature. Because the fireside chat ran so long, we only had 40 minutes for them. Even so, everyone dove in and talked actively, and at the end each group came up and presented without a hitch, so I relaxed again.

After the event, I stayed in the space chatting with people until another hour had slipped by, and I kept talking with attendees during the five or six minute walk to the subway station. Oddly enough, everyone said today's questions and small groups were really good. Honestly, I'm not exactly sure what they liked so much. ;;

I think it's because they got to meet someone who actually builds AI models in Silicon Valley, at the very front line, at a place like Thinking Machines Lab, so every piece of that experience felt special and different. Not many people in Korea build AI models themselves. The stories about working shoulder to shoulder with Elon Musk, and about coming in at 9 a.m. and leaving at 2 a.m. (sometimes 6 a.m.) with no weekends for more than a year, were wild to us. (For Sungsik it's still a bit of PTSD.) The memes of people sleeping in tents were all real, and Sungsik slept there too.

Anyway, I plan to keep hosting Claude Bloom events often with people I'm interested in and curious about. If we get lucky, I'll bring in people from overseas too!

1. Seongpil Nam's opening

Seongpil Nam of AB180 opened with the topic of AI-native organizations: an organization designed around AI from the start, instead of AI bolted onto an existing one. A big example was A-bot, which AB180 runs internally. He said this one automation system has handled 800 PRs. That single number explained what "AI native" means.

2. How one career gets built

Sungsik Kim majored in CS at Northwestern University and started their career at Uber. Then they moved to Google DeepMind, worked on the AI Assistant team, and joined the Bard project, the team people called Code Red. Next came Cruise, and from there Sungsik joined xAI as its 100th member. Now they're at Thinking Machines Lab, founded by Mira Murati (former CTO of OpenAI) and John Schulman (an OpenAI co-founder). Strung together in a line it looks simple, but the stories that followed made it clear none of these steps were simply handed over.

3. Between tents and capsule hotels

Daily life at Google and at xAI was completely different. At Google, Sungsik worked 20 hours a week on light weeks and never more than 50 on heavy ones. New hires got about three weeks of onboarding. xAI was different from day one. After 30 minutes setting up the computer, Sungsik joined the team right away. The first assignment was a paper someone showed them, with orders to implement it and ship it by that night. That first day, they left between 11:30 p.m. and midnight.

There was no such thing as office hours. Some people came in at noon and left at 6 or 7 the next morning. Sungsik came in between 9:30 and 10 a.m. and left at 2 or 3 a.m. Someone asked whether the tent photos that went around Twitter during the Grok 3 launch were real. Sungsik was there at the time and slept in a tent. They slept at the office three or four times a week. Later, when the company had money, rooms like a capsule hotel were built, 16 in total. It was hard, but strangely it felt like pulling all-nighters on projects with friends in the college library. That intensity actually brought people closer.

4. Working with Elon Musk

Elon met every team in person at least once a week, and he didn't work by getting reports from executives. Instead of carefully prepared slides, he'd point at individual developers one by one and ask what they were working on right now. Then he'd dig at least three levels down. Why did you do it that way? So what is that? What happens next? He went pretty deep technically.

He was direct with his emotions too. If he liked something, it was about as far as "good job." If he didn't, he said things much harsher. The example Sungsik gave was the word "*ucktard," a twist on a word that's restricted in the US as a slur against people with disabilities. He actually used it directly on a specific employee. Sungsik said, "Every time I heard stuff like that, I thought, he really is smart, though."

What Elon hated most was going in a direction that didn't hold up when you thought from first principles. Self-driving was the classic example. At the time, most companies used a range of sensors like lidar and radar. Elon's logic was that people drive with only their eyes, so camera vision alone should be enough. Back then many people thought he was wrong, but looking at it now, Sungsik said, he got quite a bit right.

5. A project that started with two people

When Macrohard was first announced, Sungsik and teammates thought it was a joke. Starting with the name. When Elon unveiled the project at an all-hands meeting at xAI and asked who would take it, Sungsik's manager immediately pointed at Sungsik. They'd been on another coding task and looked up thinking, no way. After the meeting, they were told to drop everything and do this. One other person started with them. Two people in total.

Microsoft is a company that runs almost purely on software, with no hardware. If AI could replace a company like that with no human involvement, that would be AGI. They picked Microsoft as the company that would show that bar most clearly, and the name Macrohard meant going the opposite way. The goal was a digital human emulator: a model that could do everything a person can do at a computer, from planning to execution, the way a person would.

Inside, Elon's direction and the team's direction clashed often. When they couldn't agree, they'd sometimes split the team in half and try both things at once. That slowed them down, and going in directions people didn't believe in hurt focus too, Sungsik said. In the end there was a restructuring and the team was reorganized.

6. Coding until the morning of their own wedding

Several reasons added up to Sungsik leaving xAI. A big one was that the two people they looked up to most left first. One was Jimmy Ba, who co-created the Adam optimizer. The other was the person who built the first StarCraft AI at Google. Those two were Sungsik's mentors and managers, and both left xAI.

Then Sungsik hit a personal limit. They were newly married, but the life of sleeping at the office with no weekends kept going. On their wedding day, Sungsik was coding all morning until 10 minutes before walking into the venue. When they messaged their manager to say they really had to go now, they found out later that the manager had also just gotten married. So the manager let them go. After leaving, Sungsik seriously considered starting a company with a few people around them. Frontier lab experience changes how much funding you can raise, and there were people who wanted to join. But they'd come out of it burned out, so for now they wanted to recover on the research side first.

7. What Thinking Machines Lab is trying to build

Sungsik said TML is headed in a different direction from other frontier labs. Most labs are building technology that replaces people. What TML wants is a relationship where people and AI can work together.

One of the things they recently released was about multimodality. Right now, when you use Claude or GPT, you take a screenshot, paste it in, describe the situation in text, and then ask your question. That's already an email format. When people talk to each other, they see and hear each other, and can show something directly if needed. TML's direction is to connect much richer context so you can communicate with AI that way.

Another is full duplex. Most AI systems today are turn-based: I talk, then the other side takes it and responds. When people talk, you can interrupt while speaking, or think about something else while listening. Full duplex means building that into AI. Sungsik used the JARVIS analogy from Iron Man. JARVIS may be smarter than Tony Stark in many ways. That doesn't make Tony Stark disappear. Iron Man comes out of the two of them interacting. Sungsik also said full duplex will matter even more in physical AI. An email can wait for you. A car moving in the real world can't.

8. When models train models

There was also talk about the growing number of startups working on RSI (recursive self-improvement). One direction Sungsik described is models training models directly, a setup where the model keeps improving itself with no researchers. The other is RSI on the product side. An agent like Claude builds and ships a product, checks the logs, looks at bugs and feature requests, and iterates on its own. A company called Humans&, founded by one of Sungsik's former xAI colleagues, is going in the AI companion direction. It started from the belief that we should invest in AI you can treat like a friend, more than in productivity tools, and it has raised $500 million.

9. How do you define AGI

Sungsik asked the audience a question: who here feels they couldn't work as efficiently as they do now if AI were switched off? Almost every hand in the room went up. Sungsik said that means we're already pretty close to the dependence stage.

Each lab defines AGI differently. OpenAI sees it as the point where AI can do everything in every job at the level of an average expert in that field. Elon's standard showed in Macrohard: AGI is when a company can run on AI alone, with no human involvement. Demis Hassabis has yet another standard: if you gave AI the same context Einstein had when he discovered special relativity, and AI made the same discovery, that would be AGI.

Sungsik's own view was that on knowledge and logic, AI may already be ahead of most people. What's missing is the generalist part. Two things still fall short compared to people: the ability to learn completely unfamiliar information quickly from just a few examples, and continual learning, where it keeps learning as a conversation goes on.

10. Why Korea has an edge

What impresses Sungsik about Korea is how fast it reads trends and adopts them. It's something Sungsik often talks about with friends in Silicon Valley. Terms like "DX" and "AX" (digital and AI transformation) are already everywhere in Korea, but Sungsik had never heard them in the US. That alone shows how quickly Korea gets on board.

Korea's AI token usage is No. 1 or No. 2 in the world per capita. That's why Anthropic opening its first overseas office in Korea, and OpenAI running a Korea office too, are more than symbolic. On whether Korea really needs its own foundation model, Sungsik said there's no right answer. But if Korea's token usage is exponentially high, global labs will have no choice but to train their models to fit Korean data. Even without building sovereign AI yourself, using it heavily can itself give Korea bargaining power.

11. Why people will still matter after AGI

Sungsik said it will take longer than people think for every job to disappear. Look at Claude Code today: on certain tasks it's better than most developers, yet adding more developers still raises productivity. The era when one skill could be a job for life is over, but people haven't become useless. What's become important is the flexibility to move into other fields while working with AI. Not working doesn't make life meaningless either, Sungsik said, just as retired people's lives aren't meaningless.

In a post-AGI world, Sungsik expects entertainment to grow a lot. The areas AI replaces quickly are tasks you can grade in binary terms. With coding, you run the tests and know right away whether it's right. But it's hard to write a spec for whether a film is a masterpiece or a flop. So the role of people who can set those standards of judgment becomes more important.

12. What came up in Q&A and small groups

In the Q&A, Sangcheol Kim, who's building a note app called Tiro, said they were curious about Macrohard's internal objective function. They had expected something like organizing processes, along the lines of Palantir's ontology, and were surprised to hear it was about model training. Sungsik's answer was the digital human emulator. The goal was a model that handles the whole process like a person, from planning to execution, which goes well beyond automating interfaces. Someone doing agent research at KRAFTON AI (the AI arm of the game company behind PUBG) asked about the structure of TML's model and its development goals. Sungsik said it's built so that a System 1 layer calls a brain layer, that different companies will each reach AGI in their own way, and that TML's strength is interaction with people. Eungyu Lee, who makes AI films and immersive media, asked which area creators will hold most strongly after AGI. The answer: people who can define qualitative judgment.

The small groups only had 40 minutes, but they were dense. At Bongsu Choi's table (from Laf Labs), the question came up of whether AGI can really exist if, even in multi-agent systems, a person still has to be left to make the decisions. At the table of a growth marketer from Tiro, people talked about the LLM wiki idea Andrej Karpathy proposed and building a personal knowledge base. They agreed that lowering communication costs is the key to using AI well. At Byeonguk So's table (from Fujifilm), someone said that separating what humans do from what machines do might be a distinction in name only, so maybe we should see the two as one fused idea, like a married couple, instead of splitting them. Sungsik briefly joined Jeongmin Oh's table (from Morenter) to talk about sovereign AI. Sungsik said AGI-level AI could be weaponized like nuclear weapons and won't be made public, so pursuing sovereign AI for the sake of self-reliance makes sense. At Wuwon Shin's table (from LG Electronics), Lee Sedol and Shin Jinseo came up. In Go, a domain AI has already swept through, both players are still doing fine and the prize money is the same, so looking at domains AI has already passed through might show us our own future.

Thank you to AB180 for the delicious catering.

Originally written in Korean on Brunch, May 15, 2026.