Inside the Frontier Labs: Google, xAI, and Thinking Machines

Joseph Kim, who passed through Google DeepMind, xAI, and Thinking Machines Lab, unpacked the real working intensity of the frontier labs, how they operate, and the competing definitions of AGI.

The speaker carries a rare resume: computer science at Northwestern, a career start at Uber, the AI assistant team and the Bard project at Google DeepMind, the 100th member of xAI, and now Thinking Machines Lab, founded by Mira Murati and John Schulman. In Korea you rarely meet someone who builds the models themselves, so every story landed differently. AB180 CEO Sungpil Nam opened the night on AI-native organizations, designed around AI from the start rather than layering AI on top, and one number explained the concept instantly: a single internal automation system that has handled 800 pull requests.

The intensity of xAI, and the First Principles of Elon

Daily life at Google and at xAI could not have been more different. At Google, light weeks ran 20 hours and never past 50, with three weeks of onboarding. At xAI, day one meant 30 minutes of computer setup, straight onto the team, and a first assignment of implementing a paper and shipping it by that night. Commuting as a concept did not exist; Kim arrived in the morning and left at 2 or 3 AM. The tent photos circulating on Twitter during the Grok 3 launch were real, and he slept in one, spending three or four nights a week at the office. It was hard, he said, but it felt like all-nighters in the university library, and the intensity actually pulled people closer.

Elon Musk met every team at least once a week, in person. Instead of taking executive reports, he pointed at individual engineers and asked what they were doing, drilling at least three levels down: why did you do it that way, so what is it, and what happens next. He was technically deep, blunt in expression, and unsparing when displeased. What he hated most was any direction that failed a First Principles test. Autonomous driving is the canonical case: while most companies used lidar and radar, he insisted on camera vision alone, on the logic that humans drive with eyes only. Many said he was wrong then; looking now, quite a bit of it holds up.

Macro Hard, and the competing definitions of AGI

The project Kim owned at xAI was Macro Hard, a name that starts as a joke: Microsoft runs almost entirely on software without hardware, and if an AI could replace a company like that without human involvement, that would be AGI. The name marks the opposite direction. The goal was a digital human emulator handling everything a person does at a computer, from planning to execution. It started with just two people.

As the project suggests, AGI is defined differently at every lab. OpenAI sets it at AI performing at the average level of a domain expert in every occupation. The Musk standard is Macro Hard itself: a company operating on AI alone without humans. Demis Hassabis sets it differently again: give AI the same conditions Einstein had when discovering special relativity, and if it makes the same discovery, that is AGI. Kim gave his own diagnosis: in knowledge and logic AI already beats most people, but in the generalist territory of learning brand-new information in a few tries and continual learning across a conversation, it still falls short of humans.

What TML wants: collaboration, not replacement

Kim said Thinking Machines Lab points in a different direction from the other labs. Where most build technology that replaces people, TML aims at a working relationship between people and AI. Two directions were made public. One is multimodality: today we paste screenshots, describe our situation in text, and then ask, which is basically email. People see and hear each other and show things directly; TML wants to connect that much richer context to AI. The other is Full Duplex. Most AI today is turn-based: I speak, it responds. People interrupt each other mid-sentence and think while listening, and TML wants that in the model. Kim reached for Jarvis from Iron Man: Jarvis may be smarter than Tony Stark in many ways, and Tony does not disappear; Iron Man emerges from their interaction. Full Duplex matters even more for physical AI moving through the world, because email waits and a car on the road does not.

The era of models training models

Startups built around recursive self-improvement are multiplying, in two branches. One is models training models directly, improving without researchers. The other is product-side RSI: an agent ships a product, reads the logs, watches bugs and requests, and iterates on itself. A company founded by a former xAI colleague of Kim raised 500 million dollars, aimed not at productivity tools but at an AI companion you treat like a friend.

Why Korea has an edge

What impressed Kim in Korea was the speed of reading and adopting trends. He said he often discusses it with Silicon Valley friends: terms like DX and AX are everyday vocabulary in Korea and unheard of in the US, which itself shows how fast Korea boards this wave.

The numbers agree: Korean AI token usage per capita ranks first or second in the world. That Anthropic opened its first overseas office in Korea and OpenAI runs a Korean office is not mere symbolism. On whether a sovereign foundation model is necessary, he said there is no single answer, but if Korean token usage keeps compounding, global labs have no choice but to train their models for Korean data. Even without building sovereign AI, heavy usage is itself negotiating power. For companies, that is a clear signal: the current speed of adoption and volume of usage are assets in their own right.

Why people remain after AGI

He expects the disappearance of every job to take far longer than people think. Even Claude Code today beats most developers on specific tasks, and yet adding developers still raises productivity. What ended is the era of one skill as a lifelong job, not the usefulness of people. What matters now is the flexibility to collaborate with AI and move across fields.

In a post-AGI world he expects entertainment to grow enormously. AI replaces fastest where right and wrong can be judged immediately: run the tests and coding has an answer. Whether a film is a masterpiece or a disaster cannot be written as a spec. The people who can define those qualitative judgments become more important, not less.

Questions from the small groups

The forty-minute small groups were short but dense. A note-app founder was surprised that the internal objective of Macro Hard is model training rather than process cleanup; the answer circled back to the digital human emulator. A game company AI researcher asked about the TML model structure and got an answer about a System 1 layer calling a brain layer, with the point that many companies will reach AGI by different roads and the TML edge is human interaction. An AI film creator asked where creators remain strongest after AGI: the people who can define qualitative judgment. Another table debated sovereign AI: AGI-level systems may be weaponizable like nuclear arms and never fully released, which gives sovereignty-level AI real meaning. One table reached for Go: in a domain AI has already swept, the top players still thrive and the prize money remains. Look at the domains AI has passed through, and you glimpse our future.

What companies can take away

Four things, in working language. One, ask again from First Principles: not what everyone does, but whether it holds at the root, three levels deep. Two, design for collaborative AI: structures where people and AI work together are what produce real productivity. Three, treat adoption speed and usage as assets: Korean token volume is negotiating power with the global labs. Four, keep the capacity for qualitative judgment: where right and wrong cannot be scored, the person who sets the standard matters more.

FAQ

How does Thinking Machines Lab differ from other labs? Where most frontier labs build replacement technology, TML aims at human-AI collaboration, with multimodality and Full Duplex at the core of AI that communicates like a person.

Why does the definition of AGI differ by lab? OpenAI: expert-average performance in every occupation. Musk: a company run by AI alone. Hassabis: reproducing an Einstein-level discovery. Different standards, different judgments of arrival.

What is the evidence that Korea has an edge in AI? Fast trend adoption and per-capita token usage at first or second in the world, which is why Anthropic chose Korea for its first overseas office.

Scenes from the night

Join Bloom

Bloom builds offline rooms where people and technology meet. We run them in Seoul, and now beyond it.

• 💬 Discord community, where events are announced first
• ▶️ YouTube, full talks from past events
• 📸 Instagram, photos from our events
• 🎟️ Upcoming events