GenAI, Then Agents, Then Physical AI

I went to the Dojeon-gwa Nanum ("Challenge and Sharing") CEO breakfast again this month. It's run every month by Chairman Lee Geum-ryong, one of Korea's first-generation business founders. It's held at GS Tower near Yeoksam Station in Seoul, and doors open at 6:30 a.m. Every time I go, I'm struck by how many people show up that early and stay completely locked in. I believe tickets are around 150,000 won (about $100) per session, but as far as I know we were lucky enough to attend for free as Chairman Lee's guests. (I could be wrong about that.)

Last month's breakfast was a big help in understanding the larger flow of AI, and since two sessions in a row covered AI, I could really feel how fast the trend keeps moving. For the record, this breakfast isn't a tech event. The speakers are experts from all kinds of fields.

As I shared in my post about last month's breakfast, the age of humanoids is coming soon. Of all the shapes a robot could take, the reason it should look like a human is that the world's infrastructure and workplaces are already built for humans. Matching that shape is the fastest way to put robots to work. And the biggest difference between LLMs and physical AI is the impact they have on the real world. The obvious leaders in physical AI are the US and China, and I suspect this was the big picture behind Elon Musk building Tesla in the first place. (Broadly speaking, self-driving cars are physical AI too.) That's why he's been able to get results so quickly in the humanoid market.

China, of course, is putting out a lot of explosive, experimental technology. If you look at the tech stories that go viral these days, more of them come out of China than you'd expect.

But according to today's speaker, because humanoids operate so close to the real world, security could become a big issue. (Software can be a serious strategic weapon too, of course.) Tied to national security, Western countries may decide it's risky to buy Chinese humanoids in bulk and deploy them. His point was that even if they're cheaper and more capable than you'd think, they may have a hard time spreading quickly around the world, because they're not just robot vacuums. So political and security issues could get involved.

A few takeaways stuck with me.

First, Cliwant, the company I work at, is a startup founded in September 2023, only a year and a half old. Back then, businesses built on GenAI were very trendy and hot. We showed up in a lot of news stories and even got attention from OpenAI and Google.

We didn't start the business to chase a trend. We were just lucky that the problem we were trying to solve lined up with the trend of the moment, so we got off to a great start. But the trend moved fast: to AI agents in 2024, and to physical AI in 2025. I came away thinking trends will change even faster from here.

I'm not saying we should run the business by chasing trends. What matters is seeing clearly what our business really is. Still, it's true that we got a big boost when we started, and it reminded me again how much timing and speed matter.

Second, this is an age that calls for a sense of urgency. Early this year, DeepSeek and Manus AI went hugely viral. The point is that even if you can't create trends this fast, if you can't keep up with them either, you'll fall behind easily. It feels like a new mega trend shows up about once a month. Someone at our company suggested we hold an AI study session once a month. We don't need to cover every issue, but we agreed that anything that lasts more than a month is meaningful and can be covered after the fact.

My real point is that the US and China are driving innovation at this speed, and it's a shame Korea can't. The US is the US, but China is far more culturally closed than we are, and it still builds things with enough influence to shock not just its own country but the whole world. That amazes me. It's probably because China treats AI as a national strategic weapon and the government backs it all the way. Korea's government supports AI too, of course, but China seems to have a stronger energy and drive, and I honestly don't know what it is. (My guess is that patriotic global expansion, capitalism and a pure drive for research and development all fit together there, while in Korea only the capitalism is left.)

Also, unlike LLMs, physical AI needs much bigger upfront investment. Above all, it needs energy, meaning a steady supply of electricity. The speaker said a time will come when electricity costs more than labor, and that more and more AI companies will move out of Korea, where labor is cheap, to the central US, where power is cheap.

Even so, I think Korea is one of the countries that supports AI best in the world. We've received a lot of benefits and support ourselves, so blaming the country would be foolish. Instead, we should always be looking at the global market and thinking about how to beat American and Chinese companies and become number one. These days I try to keep a healthy sense of urgency and actually enjoy it.

To me, a startup means being tightly connected and quick to respond, and size has little to do with it. A giant ship like the Titanic can see the iceberg and still hit it, and most big organizations can't help reacting the same way. A small boat has the advantage of turning fast. Ironically, not every small boat does. I hope Cliwant stays close to the market, responds to trends quickly and keeps moving forward.

The Physical AI Revolution: A Game Changer for Korea's Advanced Industries

Speaker: Jeong Ji-hoon (adjunct professor at DGIST, a science and technology university in Daegu, and co-founder of Asia2G)

### Tech mega trends

Artificial intelligence: AI used to need huge amounts of energy, but recent AI, like DeepSeek, is moving toward efficiency.

Semiconductors: With Broadcom and other companies working together alongside Nvidia, the monopoly structure is breaking down.

Energy and analog transformation: Electricity in the US Midwest is far cheaper than in Korea. Energy costs are now a more important deciding factor than labor costs, especially for energy-hungry industries like semiconductors. On the other hand, industries that need huge infrastructure, like shipbuilding, won't find it easy to move to the US.

Mobility and data cities: Changes in mobility change how cities are built. Which country the advanced cities end up in will create gaps in national competitiveness.

ESG: Using AI to make environmental systems more efficient (UV, energy and so on) is an important angle.

### Gartner's 2025 AI trends

Security: How much an AI system can reduce risk matters.

New computing environments: Quantum computing, spatial computing and physical AI are arriving.

### CES 2025's main keyword: physical AI

The progression goes perception AI, then generative AI, then agentic AI, then physical AI.

We saw robots that project faces, humanoids that connect with people emotionally, and robots meant to replace pets.

Physical AI (AI that interacts with the real world): The difference from LLMs is that it perceives the real world and acts in it. Robots are taking over the roles of TVs and projectors and developing ways to communicate with furniture.

### Key companies at CES 2025

John Deere: Went from traditional farm equipment to AI tractors. This year it announced AI tractors specialized for five different terrains. For farm AI vehicles, the key is data from nature, not from roads. Tesla has a lot of road data, but John Deere is collecting off-road data from farms, orchards and so on.

Sony: Moving beyond its TV focus and developing self-driving cars with Honda. It wants to connect content IP like the PlayStation 5 to the world of AI, and it's working with Disney on a character strategy for the AI era.

Umi (Panasonic and Anthropic): Panasonic worked with Anthropic to build a home AI called "Manus" based on Claude 4.5. It's a wellness AI that connects with your home.

### AR/VR and the metaverse

AI needs to understand the devices around it, not just language.

Nvidia's World Foundation Model brings in the idea of spatial computing.

Nvidia tried to build its own AI ecosystem, but it open-sourced it to spread it quickly.

The metaverse isn't going away. Jensen Huang (Nvidia's CEO) was the first to talk about the metaverse.

Combined with AI, the metaverse may actually get bigger.

It's expanding under the name Omniverse, a new term in place of "metaverse."

### AI chips and the competitive picture

Nvidia's monopoly has become a risk factor, so Broadcom and other companies are looking for alternatives.

Small language models (SLMs) are on the rise. More and more countries have their own SLMs, and those can run without Nvidia. As a result, Nvidia's stock is falling.

Vertical integration (Tesla, Apple and others): closed systems, but they deliver high-quality products and keep customers interested. In the end, creating value matters more than the technology.

### AI factories and on-device AI

AI factories led by Nvidia (centered on AI data centers) versus on-device AI (Apple, Qualcomm and others).

On-device AI runs on personal devices, and it has become more important since DeepSeek came out.

The key is making models lighter: real-time speed (responses as fast as the internet), plus personalization and security (keeping data leaks to a minimum).

### Multimodal AI and interaction with humans

AI is getting better at reading complex meaning: emotion, irony, cultural nuance.

AI interaction is expanding into VR/AR, robots, healthcare, manufacturing and more.

The hardware best suited to AI will likely be smart glasses.

But smartphones won't disappear completely.

### Robots and regulation

The risk with Chinese robots: they could be turned toward national security threats or war.

Friendly countries may end up forming technology blocs.

Competition between countries will likely be decided by a mix of factors, not just price.

### Smart cities and drone AI

Autonomous drones are being set up to inspect aging power lines.

AI runs them in a fully automated way, with no human at the controls.

How AI Is Changing Medicine, with Korea in the Lead

Speaker: Dong Jae-jun (CEO of BNJ BioPharma and professor at Yonsei University College of Medicine)

The first computer (the Turing Bombe, 1939) was developed for war. Medicine is like war too.

### The biggest challenges for medical AI

Feedback from doctors is essential to improving AI.

What matters is whether AI can improve itself through its own internal reasoning.

### Trends in medical AI

AI can find brain tumors faster than humans.

Gemini can explain checkup results better than a person can.

Services are moving away from single features toward integrated ones.

### Hospital AI and security

Because AI use is restricted inside hospitals, on-premise AI could be the alternative.

Cost and performance are the key considerations.

A tenfold increase in parameters needs a hundredfold increase in computing power.

### Changes in medical data

Doctors now spend twice as much time with computers as with patients (in the US).

If Korea's medical expertise is used as input data for AI, Korea can compete globally.

### AI drug development

The goal is to develop better drugs faster than humans can.

AI can design new drugs, but the hallucination problem has to be solved first. Right now, AI's hallucination error rate in drug development is 45%, and it's expected to be solved within five years.

Originally written in Korean on Brunch, March 13, 2025.