An xAI Engineer on Working with Elon

Silicon Valley is too fast, and Korea is too slow.

When you think about AI, three companies dominate: OpenAI, Google, and Anthropic. Then there's Elon Musk's xAI, taking a third path of its own. So I sat down for a long, in-depth interview with a friend who works at xAI as an AI engineer.

First, given Elon Musk's personality, he meets with employees in person at least three times a week, and every time he gives very blunt, direct feedback and orders. That process is pretty stressful and hard. Still, the chance to work with Elon Musk alone has drawn a lot of talent. And unlike the other big tech companies, which are software driven, xAI is building its AI business with a somewhat different vision and direction, on top of Tesla's hardware and SpaceX's space business. That always gets a lot of attention.

Side note: one of our company's investors recently came back from a trip to Silicon Valley and told us that Silicon Valley is too fast and Korea is too slow. Personally, that was a real shock. Silicon Valley is where tech trends start, so of course we've always been a fast follower. I assumed there would be a lag. But hearing our investor point it out again stung, because I thought we were chasing pretty hard. What threw me even more was that I couldn't even tell what the structural problem was, or where we actually stand objectively.

Yesterday I also talked with a friend who works at an American financial firm in Hong Kong. This friend studied in China and the UK and now works in Hong Kong, so they have a balanced understanding of both East and West. Once a month they go to Shenzhen, only 15 minutes from Hong Kong by train. Just like you see in the media, there's drone delivery, and at hotels a robot arm comes out, knocks on your door, and hands you your food. Asia, China included, used to be an ecosystem that ran on subcontracting work from the US and the West. But now that nearly half of the engineers in the US are Chinese, there's a growing trend of people moving back to China. So a market that was built on manufacturing is now getting the idea people back too. With 1.4 billion people, China has a big enough market to be self-sufficient, plus aggressive pushing from the Communist Party. The US and China have completely opposite systems, but each is surging ahead in its own way.

Anyway, back to xAI. When I called my friend there, it was past midnight on a Sunday, local time. On the other side of the screen, the office was as bright as noon and buzzing with coworkers talking. If China has 996 (9 a.m. to 9 p.m., six days a week), Silicon Valley right now is even worse. Coming in on weekends is a given, and there were team meetings at 3 p.m. on both Saturday and Sunday. In his words, "Everything except breathing is work time." San Francisco is so expensive that if your job falls apart, just getting by becomes hard. That pressure pushes the intensity even higher. Silicon Valley is at war right now. Every AI lab shares the same fear: if you fall even a little behind in the AI market, there will be no way to catch up later.

My wife read this and said, "Wow... if work is just going to get harder, why does AI even exist? That's ironic." Tell me about it, haha.

Looking at all this, Korea is lucky K-pop took off recently. Our traditional strength, manufacturing, is being eaten away piece by piece by China, and in the US autonomous robots are arriving and making our know-how look second-rate. Everyone can feel that the national strategic crisis is getting bigger. Feeling anxious about this isn't strange. I've heard that in the UK it has recently become hard to walk around after 10 p.m., and so dangerous you can't walk down the street holding your phone.

The UK is also now the world's biggest exporter of millionaires, at three to four times second-place China, with India third and Korea fourth (adjusted for population, Korea is effectively second). The Philippines, once called the next Japan, has become a haven for corruption and drugs. The UK itself lost the standing it had as the empire on which the sun never set. There are plenty of examples showing that once a country starts to decline, it can keep falling with no bottom.

That doesn't mean Korea is out of hope yet. So, to keep listening to global trends, I'm sharing the conversation I had today with an xAI engineer.

1. From big tech to xAI: betting on the AI underdog

My friend joined an AI project at a US big tech company in its early days, starting in 2020. He worked there for about three years on agents and tool calling, training models and building agent loops.

Then xAI reached out. xAI had a lot of momentum at the time, and above all, the talent pool was impressive. His future manager was a well-known deep learning researcher who created one of the optimizers used in almost all LLM training today. The chance to learn under people like that was a big draw. The offer was also much better than at other labs, and most of all, xAI had the energy of an underdog rising fast with Elon Musk's backing.

Once he got to xAI, the talent density was clearly different. His previous company was also full of people from Stanford, MIT, and Harvard, but he said their sense of urgency and work intensity came nowhere close to the people in SF. One thing that surprised him was that age didn't matter. People much younger than him were in leadership roles, and they were really good. Seeing that, he realized he couldn't lean on his age and experience and had to keep learning. In fact, if you look at how talent is moving these days, there are senior people from C-tier companies like Atlassian joining Stripe as regular engineers. They've decided that working close to AI is worth more, even if it means a step down in title.

2. One-on-one with Elon is much scarier than the media shows

At xAI, almost every team meets Elon in person once a week. During the stretch when he was really focused on xAI, they briefed him about three times a week. I asked what Elon Musk is like in person, and the answer was clear: much scarier than he looks in the media.

Elon's style is closer to a "king" than a CEO. He runs his companies like a dictator. I don't think that's necessarily bad. In a small, fast startup, strong leadership and a clear push in one direction can work better than the democratic consensus you get at a big company. And that's how Elon actually built Tesla and SpaceX. He gathers people in a room and tells them what to do, and since you get fired if you don't listen, everyone follows and works hard. It has produced huge results.

The message Elon stresses most to everyone is "usefulness." He says AI means nothing if it isn't useful. He wants products that actually help with people's work and lives, beyond fun or surprise. Don't work on anything that isn't tied to usefulness. My friend said this has become the yardstick for every team's decisions.

He also said Elon understands the technology very deeply. He enjoys technical conversations with engineers, and when something seems off to him, he digs in relentlessly. For example, if you report that training a model takes 100 hours, he'll say, "Why does it have to be 100 hours? What makes it 100 hours? There's definitely room to improve somewhere. Get it down to 20." He also tests the products himself every day and posts in the group chat, "This is weird, fix it." Even now, on a Sunday, Elon is at the office.

Of course, there are plenty of downsides. AI is a bit different from regular engineering in that R&D needs enough time. Building a model, cleaning data, and training it all require patience, and Elon didn't have that much patience.

On that point, my friend once talked with an xAI cofounder who was close to Elon. What that person said stuck with him. Elon cares much more about "eventually" than about whether something is right at this moment. Even if a decision looks bad right now and doesn't seem to make sense, if you look at what Elon has done so far, he eventually pulled it all off. The ironic part is that this cofounder left the company not long after saying this, haha.

Even so, lots of talented people still go to Elon's companies. Part of it is that you don't really feel it when you only hear about it. And the experience of working face to face with Elon is very rare and an honor in itself.

3. The rise and fall of Macrohard: drive a computer like FSD

Early this year, a project called Macrohard was finally made public. But it was actually xAI's first project, started last September. The ambitious vision was to beat Microsoft, hence the name Macrohard, and the goal was to have AI operate a computer. My friend was deeply involved in it.

Elon's instructions were clear: use exactly the same model architecture as Tesla's FSD (Full Self-Driving). In Elon's mind, operating a computer was the same idea as a car driving itself. He even thought operating a computer was much easier than driving, since a driving mistake can kill someone and a computer mistake can't. The development team saw it differently. Roads are fairly standardized. You follow the lines, stop at red, go at green. It's a setup that doesn't need much planning or reasoning. With a computer, pressing a button isn't the hard part. What matters is the thinking: why you're pressing this button and what logic goes into it. The team saw it as a fundamentally different problem.

After months of wrestling with it, the Macrohard project ended up being handed over to Tesla.

4. Elon's vision for xAI is different

Honestly, xAI is in a tough spot. All the cofounders have left, many of the lead researchers have gone, and people are still leaving. AI in particular is a field where it's hard to compete without talent, and he said that part worries him.

Right now, Elon's interests come down to three things.

First, catching up with Anthropic is the top priority. So xAI recently moved more than half the company over to coding and is now all in on building a coding agent.

Second, Optimus. Elon's view is that even if xAI can't be the very best at digital AI, once things move to physical AI, it will win no matter what, thanks to the manufacturing capability and data at Tesla and SpaceX. It's hard to send humans to space to do labor, so sending Tesla's Optimus robots to space would be efficient. That logic sort of makes sense.

Third, he is very seriously pursuing data centers in space. Nobody knows yet how feasible that is, but since it's Elon, it might actually happen.

xAI merged with SpaceX early this year, and they've been working more closely since. Elon pushes xAI hard, so even if Tesla or SpaceX want to use a different model, Elon tells them to use xAI's model first. For xAI that's good, since data comes in fast. For Tesla or SpaceX, it might be a different story.

5. Why Anthropic leads right now: coding is a gold mine

I asked him when he first felt that Anthropic was in the lead. He said early last year, when Claude 3.7 came out. More precisely, when Claude Code first launched. He tried it right away, and it was absurdly good compared to everything else.

Coding might look like just one field, but it comes with huge benefits. First, models are priced by tokens, and coding agents use far more tokens than ordinary conversation. The revenue from that is much higher. Second, almost all the work at an AI lab is done through code, so a good coding model makes the company itself more efficient. Anthropic used the tools it built to move faster. Last, every software service is built with code, so there's also the argument that if you're good at coding, you can be good at other verticals too.

There's also a structural reason coding advanced before other fields. Coding is the domain AI researchers know best, because it's what they do every day. So research there moved fast naturally. If you ask people in finance, they'll tell you the models still aren't good at things like Excel. People in hardware say there are a lot of hallucinations. Many other domains don't even have data pipelines yet, so they may take longer.

There's one more important point. A company that builds an agent harness and a company that builds the model itself are fundamentally different. The closer the harness used in training is to the harness used in production, the better the model performs. So a company that owns the harness and trains with it has an overwhelming built-in advantage. No matter how well a wrapper like Cursor is built, Anthropic, which owns both the model and the harness, has a structural edge.

When I asked when AGI is coming, he said he takes what Sam Altman says with a grain of salt, since he's a promoter and a businessman. He also said AGI itself has no clear definition, so it's a fuzzy question. But by his own standard, AGI has already been reached to some degree. His reasoning: ask tech companies today whether they could run as well as they do now without Claude Code and Codex, and most would say no. To him, that's evidence we've gotten pretty close to AGI.

6. The DNA of the AI labs: Google, OpenAI, Anthropic, xAI

I asked about the differences in direction and company culture between the labs. He said they're quite different.

Google is organized like a classic big company, so there's bureaucracy and it tends to move slowly. On the other hand, its resources are overwhelming, so it puts out a lot of good work. OpenAI has been at it a long time and has focused heavily on consumers. He said it feels like they put more effort into building models that can interact well with people than into pushing hard on reasoning or coding. Anthropic is clearly focused on agents, coding, and enterprise. xAI owns the world's largest GPU cluster (Colossus), so its strategy has been to make the most of that compute advantage.

We also talked about scaling laws. If you increase data linearly, performance goes up on a log scale. Flip that around and it means that once you start feeding in all the data in the world, each improvement keeps getting smaller. In other words, today's transformer-style models may all end up roughly the same someday. So he expects Google and OpenAI to catch up with Anthropic soon. He doesn't think Anthropic has some special secret technique. It paid a lot of attention to data and focused well, and that's what worked. He thinks other labs can catch up.

So this race looks a lot like the way Google took over the search market. As investment in and dependence on AI keep growing, once you start falling behind, there's no way to catch up later. That's why every AI lab is fighting tooth and nail to get even a little ahead. He's also watching Chinese companies. Quite a few of them don't get the credit their performance deserves. Just look at the recent GLM, which he says performs almost at the level of a top model like Claude Opus. The Chinese style is to pour your soul into something once it looks like it'll work, and he said most of the engineers inside xAI are Chinese too. What they care about is getting ahead and having their ability recognized. They don't care much about political ideology. They just want to be recognized wherever things are working.

7. He feels an AI bubble

He said there's definitely a bubble in AI. Here's how it works. Revenue at AI startups is growing exponentially, but the way they make that revenue is unusual. Almost all of these companies build wrappers on top of foundation models. The better the models from Anthropic, OpenAI, or Google get, the more people use these products on their own. Even if the wrapper company does nothing, the product gets better because the model gets better. It also means many users use the product in order to get to the model.

Then almost all the money that company makes flows back to Anthropic or OpenAI. Startups look at revenue, not profit, so the numbers go up exponentially. They raise another round on those numbers and use even more tokens. A loop forms.

Even so, he said AI has already passed the point where we can't do without it. The foundation itself isn't going to shake. And wrappers aren't automatically doomed. Cursor is an example. It lets you use multiple models, it has a Tab model that autocompletes your code as you write, and its UI/UX makes coding comfortable, so a lot of people still use it.

But if you ask whether AI will replace everyone right away, reality is a bit different. The area AI replaces best is digital work. Jobs where you only work on a computer, typing, organizing data, and sending emails, are the most at risk. On the other hand, AI still falls far short at work where you deal with people face to face, build relationships, and read emotions. He said understanding people's deeper feelings and the context of relationships still seems a long way off.

There's also a reason AI isn't getting adopted on the ground. Even if AI does 99% well, 1% is still errors. When a person makes a mistake, there's room to fix it or adapt to the situation, but AI's 1% error is always there. He said quite a few businesses can't adopt it because of that 1%. So he sees a middle stage coming first, one that maximizes the productivity of existing people, rather than aiming to replace people outright.

What's interesting is Anthropic's ultimate goal. Right now Claude Code needs a person next to it the whole time. He said what Anthropic is after is a recursive harness (a loop that keeps improving itself on its own), where instances of Claude Code consult each other to decide on the next feature and analyze logs to raise product quality. Inside that harness, there's no role for people. He said nearly all the research is heading in that direction. That said, the field where we're having this conversation is still about people at the very edge of the trend, not the general public.

8. The grammar of Silicon Valley: America grows on capital, not labor

At the start, I mentioned our investor's comment about the speed gap between Silicon Valley and Korea. My friend's view on the reason was clear. In the end, there's far more funding in the US and China than in Korea. It's not just that there's more money. Labs that build foundation models also invest directly in companies a lot. Sometimes they invest money, and sometimes they invest in tokens, as in "use our model to build it." That creates a setup where builders can build and use things much more easily.

There was a time when bootstrapping was praised. But the way VC works is to inflate a bubble, grow the valuation, and then cash out. That's how financial markets around the world work, in Silicon Valley but also New York, London, and Hong Kong. If one falls, they all fall together, but it's still a real financial ecosystem, and huge numbers of people, institutions, and national funds have already put money into it. So understanding the American capitalist system well, riding the bubble strategically on top of it, and building the next bubble is also part of doing business. The point is to know how to surf the bubble without drowning in it.

Honestly, it's hard for a Korean to fully accept and understand this. It's hard for us to grasp how American companies with no profits can get valuations at least 10 times, and sometimes hundreds of times, higher than ours. Or how a company less than two years old can be worth tens of billions of dollars almost overnight. We always see bubbles as risky and avoid them. American founders have the nerve to make the bubble even bigger and plan to make money while it lasts.

My take is that America is the country of the reserve currency. Its infrastructure lets it print as much money as it wants, so to Americans a bubble isn't a mirage. It's a scenario they can fully control. (On the flip side, look at Japan: after the US pushed it into the Plaza Accord, its bubble burst and it went through its lost 30 years. Not just anyone can afford to face a bubble.)

So instead of spending the money in its own vault, the US keeps selling Treasury bonds and strategically takes on debt. With its overwhelming military power, the petrodollar, and the global credit and financial system behind it, it can keep its national credibility steady. That means it can keep growing bigger bubbles on top of bubbles without much trouble. (Every so often there's a financial crisis, but then again, the countries below it take most of the damage.)

This might be a bit off topic, but now that AI has drastically cut the resources needed to build a product, VCs say the bigger value they offer is no longer just money. It's the "multiplier" of their network, branding, distribution channels, and credibility. So the VC world also has a clear gap between the rich and the poor. That's why a top US VC like a16z created a new media division, puts real effort into podcast and social media teams and events like invitation-only dinners, and keeps building close connections with the US government.

Of course, the place that has always used online media best is Y Combinator. Getting into YC, or getting a mention from Garry Tan or Paul Graham on Twitter, is more powerful than being covered in TechCrunch. It's gotten to the point where there are tons of consulting services and YouTube channels that help you get into YC. At the same time, YC's biggest enemy is damage to its alumni's credibility. If even one alum causes a trust scandal, I've heard YC tells them internally to leave the community. If the credibility badge YC has worked so hard to build gets damaged, its branding, YC's biggest weapon, could collapse.

Marc Andreessen of a16z recently laid out five skills everyone should learn: communication, management, sales, finance, and an international perspective. The last one really hit home for me. His point was that spending time in other countries and other cultures pays off in many ways over the course of a career.

If you only stay inside the fence called Korea, you think we're the smartest, most diligent, hardest-working people. That might be true, but if you don't know the bigger picture, it may just be us shouting into the void in the global market. For a country like ours, where global exports are the only answer, I think we need to keep paying attention to other countries, including the US but also Europe, Southeast Asia, China, Japan, Latin America, and Africa. And I mean more than their tech ecosystems. We should learn their culture and history too.

These days Sweden seems to be becoming the startup mecca outside of the US and China. Finland is trying to build that through Slush, Europe's biggest startup event, but it's Swedish companies like Lovable and Legora (a legal startup) that have been taking off globally and in the US. I've traveled in the Nordics and I'm pretty interested in the culture there. Nordic countries like Sweden have harsh land and small populations, so historically they sent out the Vikings, expanded into neighboring countries, and earned an image as tough warriors. That's probably in their historical DNA. We, on the other hand, have no history or culture of invading other countries.

Finally, Yeojun Cho, CIO at The Ventures (a Korean VC firm), wrote a post quoting Andrej Karpathy, saying people who use AI fall into two groups. There are the people chuckling at the mistakes of a $20 older model, and the people running SOTA models (state of the art, meaning the newest models with the best current performance on a given AI task) whose faces are slowly going pale.

When I first wrote the Claude Blue post and it went hugely viral, I thought Koreans were just too vulnerable to FOMO. But thinking about it more, I started to feel the opposite. Not letting AI pass you by, finding your own weak spots, chewing on them, trying to get even a little better, and connecting with other people is actually a very healthy attitude. (On the other hand, I also wonder how many of the people who think they're already doing fine, with no problems and no fears, have really used AI deeply and thought hard about it. Though it's hard to generalize about everyone.)

Anyway, a tsunami looks scary from far away, but once you get on it, it might turn into a great ride. That's the mindset I want to close out the Claude Bloom event with.

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