Why Chinese AI Is Suddenly So Good (ft. DeepSeek, SeeDance 2.0) | AB Explained

0:00 Do you remember by any chance what you were doing on March 10th, 2016?

0:05 I know it sounds pretty specific,

0:06 like a complete random date roughly ten years ago,

0:09 but on that day something happened that would go

0:12 on to change or rather redefine humanity's relationship with AI forever.

0:18 You see, on March 10th, 2016, there were two opponents sitting across from each

0:23 other in a quiet room in Seoul, South Korea.

0:26 I say quiet, but it was actually a fairly large room filled with cameras,

0:31 technicians, and journalists from all over the world.

0:34 Yet, despite all that attention,

0:36 the atmosphere inside was dead silent and tense.

0:40 It had to be because the two opponents were about

0:42 to begin the second match in the game of Go.

0:45 Now, if you've never heard of Go before,

0:48 it's widely considered the oldest strategy board game

0:51 that originated in China more than 2500 years ago.

0:55 And unlike chess, which most people are

0:58 familiar with, Go is vastly more complex.

1:01 The number of possible board configurations in Go is so large that some

1:06 experts say it even exceeds the number of atoms in the observable universe.

1:11 That's like trillions upon trillions

1:13 upon trillions of possible board configurations.

1:16 Meaning the number of ways a single

1:18 game could unfold is pretty much unimaginable.

1:21 Which is exactly why for decades, up until that point,

1:25 the general consensus was that machines could never

1:28 beat a human master at the game of Go.

1:30 And even if they were capable one day,

1:33 that would have been years, if not decades away.

1:36 Anyway, on one side of the board that day sat Lee Sedol,

1:39 one of the greatest Go players in history and a national hero in Korea.

1:44 He wasn't ranked number one at the time,

1:47 but he had already won 18 international titles,

1:50 making him pretty much a global Go legend

1:52 known for his creative and aggressive playing style.

1:55 So who was sitting across from him as his opponent?

1:58 It was a computer program called AlphaGo,

2:01 developed by the British AI company DeepMind,

2:04 which had recently been acquired by Google.

2:06 Before the match began, which was the best of five series,

2:10 many Go experts believed that Lee would defeat AlphaGo fairly comfortably.

2:14 After all, humans had dominated the game of Go for over 2000 years.

2:19 But in game one, after a competitive match that lasted nearly four hours,

2:24 something unexpected happened.

2:26 AlphaGo defeated Lee Sedol.

2:29 Most spectators in the room looked on absolutely stunned.

2:33 Well, except for two people who looked

2:34 as if they had expected the result all along.

2:37 They were Demis Hassabis, the founder of DeepMind,

2:40 the company that built AlphaGo, and Sergey Brin, the co-founder of Google,

2:45 who had flown to Seoul to watch the historic match in person.

2:49 But what happened in the next match,

2:51 game two of the series on March 10th of 2016,

2:54 would shock the entire Go world, and particularly China, even more.

2:59 Midway through the game, AlphaGo made a move that at first looked completely

3:04 bizarre to the professional commentators watching the match live.

3:07 It placed a stone on the board in a way

3:10 that almost no top human player would normally do in that situation.

3:14 For several minutes, even the commentators paused their explanation,

3:18 struggling to make sense of what they were seeing

3:20 and wondering if the machine had made a mistake.

3:23 Lee himself was so stunned that he left the room

3:26 for about 15 minutes to take a break, which,

3:29 by the way, players are actually allowed to do

3:31 during a match as long as their clock keeps running.

3:35 That single move, later known simply as move 37, wasn't a mistake,

3:40 and it would go on to completely reshape

3:42 the way professional players thought about the game of Go.

3:46 AlphaGo would go on to win the second match as well,

3:49 and at that moment, something suddenly dawned on everybody watching.

3:53 Not only had a machine beaten one of the best

3:56 human players in the world two games in a row,

3:59 it had also revealed an entirely new strategy that no

4:03 human had even considered in more than 2000 years.

4:06 By the way, Lee Sedol did manage to beat AlphaGo once in game four,

4:11 using an extraordinarily creative move that many

4:13 later referred to as the divine move.

4:16 But in the end, he still lost the series 4 to 1,

4:19 and ultimately it was AlphaGo's sheer dominance,

4:22 especially that alien-like move 37 that really set China into motion.

4:28 In the West, many saw it simply as another impressive milestone

4:32 in a clever algorithm beating a human at a board game.

4:35 But to many Chinese scientists and policymakers,

4:38 the AlphaGo matches served as a powerful wake

4:41 up call that dramatically accelerated China's push into AI.

4:44 In the following year,

4:45 the Chinese government released a national strategy declaring AI a top priority,

4:51 with an explicit goal.

4:52 To become the world leader in AI by the year 2030.

4:56 Fast forward less than a decade and suddenly we are seeing

4:59 Chinese AI tools like DeepSeek and Seedance going viral across the internet.

5:04 So how did China pull it off?

5:07 Well, as I'll bring it up later in the video,

5:09 the ultimate weapon in the AI race is data.

5:12 Big tech companies are ruthlessly scraping the entire internet,

5:16 collecting every piece of text, video, and personal information they can find.

5:20 But these tech giants aren't the only ones harvesting your digital footprint.

5:24 And that brings me to today's sponsor, Incogni.

5:27 In this new age of automated AI web scraping,

5:30 your personal information is more exposed than ever.

5:33 Right now, an entire industry of data brokers, people-search websites,

5:38 and background check companies are quietly scraping the internet,

5:42 to build highly detailed profiles on you.

5:44 And that's exactly where scammers and identity

5:47 thieves go to buy your personal data.

5:49 Use it to impersonate you.

5:51 Open credit cards in your name and steal your money.

5:54 Imagine someone else taking out a loan under your stolen identity,

5:58 and you don't even realize it until it's too late.

6:00 That's why you need Incogni.

6:03 Incogni doesn't just remove you from data broker sites, they go way beyond that.

6:07 Through their Custom Removals feature,

6:10 Incogni targets the specific websites and platforms where

6:13 your personal data is being sold and traded.

6:16 They contact these companies on your behalf.

6:18 Demand your information be deleted and actually verify that it's been removed.

6:23 They handle everything so you don't have to spend months doing it yourself.

6:27 Think of it as taking yourself off the internet.

6:30 The same internet where your data is being

6:32 bought and sold every single day without your permission.

6:35 I actually use Incogni myself because of my relative

6:39 public profile and the results had been eye-opening.

6:42 So if you want to protect your data from being exploited,

6:45 try Incogni by scanning this QR code or using the link in the description.

6:50 You can use the code ASIANBOSS at incogni.com/asianboss

6:54 to get 60% off an annual plan.

6:56 That's 60% off to protect your identity and remove yourself from the internet.

7:01 I highly recommend taking advantage of this offer.

7:04 Now let's get back to the deep dive.

7:10 So before we can answer how China suddenly got this good at AI,

7:15 why don't we clear up a really basic question

7:18 that almost nobody in these conversations seems interested in clarifying?

7:23 What exactly is AI or artificial intelligence?

7:26 Because when most people hear that term,

7:29 maybe they think of ChatGPT or Gemini writing emails or essays.

7:33 Maybe they imagine all the ever-more

7:36 realistic-looking fake videos on social media.

7:38 Or you might even be thinking about those humanoid robots

7:42 dancing or doing kung fu with their crazy somersault kicks.

7:46 But here's the thing.

7:47 AI is not really one single technology.

7:50 It's really more like a stack,

7:53 multiple layers of technology sitting on top of one another

7:56 with each one depending on the layers below it to function.

8:00 And if you want to understand why the United States and China are

8:04 locked in what is arguably the most

8:07 consequential technological rivalry in modern history,

8:10 you have to understand that stack from the very bottom up,

8:13 because this race is happening simultaneously across the hardware layer,

8:18 the model layer and the application layer

8:20 that billions of people use every single day.

8:23 So let's start at the very foundation.

8:26 At the absolute bottom of the AI stack is the hardware layer.

8:30 This is essentially the physical infrastructure that turns

8:33 AI from super complex math operations into reality.

8:38 We are talking about massive data centers, cooling systems, GPUs,

8:42 and most importantly, the microchips, which,

8:45 as the name implies, are incredibly small.

8:49 If you've never actually seen how tiny these chips are,

8:53 take a look at this mind-blowing footage of a microchip under a microscope.

8:57 It's so insanely small that when I watched this for the first time,

9:02 it made me wonder how human beings

9:04 are even capable of creating something like this.

9:06 To put it simply, a microchip is a highly engineered

9:10 product carved out of a raw material called a semiconductor.

9:14 And the most common semiconductor material used today is silicon.

9:18 So physically, a microchip is literally just a tiny

9:22 piece of silicon that has billions of microscopic electrical transistors,

9:27 or switches, built directly into it.

9:31 But what are these switches actually doing?

9:33 They're simply flipping electricity on and off

9:36 to create the language that computers understand.

9:39 If the switch is off, it's a zero.

9:41 If the switch is on, then it's a one.

9:44 String enough of these together and you have computer code.

9:47 But here's the cool part.

9:49 And it's why we call it a "semiconductor" in the first place.

9:53 If you try to build a computer chip out of copper wire,

9:57 it wouldn't work because copper is a perfect conductor.

10:00 The electricity would just flow through it nonstop.

10:04 So the switch would always be stuck on one.

10:06 On the other hand, if you try to build a chip out of rubber,

10:10 it wouldn't work either, because rubber is a perfect insulator.

10:14 It completely blocks electricity.

10:17 So the switch would always be stuck on zero.

10:20 But silicon is special in that it naturally sits right in the middle.

10:24 Under normal conditions, it blocks electricity like rubber,

10:28 but if you hit it with a tiny electrical charge,

10:31 it suddenly transforms and lets electricity flow through like copper.

10:35 And that's the magic.

10:37 Because silicon can instantly switch back

10:39 and forth between blocking power and conducting power,

10:42 human engineers can actually control it.

10:45 We can treat those billions of microscopic switches like tiny gates,

10:49 commanding the silicon to flip between 0 and 1 billions of times per second.

10:54 A single advanced microchip today can

10:57 contain tens of billions of these switches,

10:59 all crammed into a piece of silicon roughly the size of your fingernail.

11:04 And that's just for a regular microchip.

11:07 If you start talking about the chips used to power AI,

11:10 everything is much bigger and much more expensive.

11:13 The single most important type of chip in the AI

11:16 world right now is something called a GPU,

11:18 which stands for Graphics Processing Unit.

11:21 Now, a GPU is still technically a microchip,

11:24 but physically it's built completely differently than

11:27 the normal chip you would find inside your laptop.

11:30 If you were to look at a GPU,

11:32 the surface of the silicon is divided up into a massive

11:36 grid of thousands of tiny brains all packed tightly together.

11:40 Originally, GPUs were designed for one very specific purpose,

11:44 rendering the complex graphics in video games.

11:48 Things like shadows, reflections, and three dimensional environments.

11:52 But researchers eventually discovered that the exact

11:55 same architecture that made GPUs so good at rendering images also made

12:00 them incredibly powerful for running the massive,

12:03 parallel mathematical calculations that AI systems require.

12:06 So when you're text prompting an AI tool to generate a video clip,

12:10 what's actually happening is that the system is performing hundreds of billions,

12:16 sometimes trillions of mathematical operations in a fraction of a second.

12:20 Pattern matching against everything it was trained

12:23 on to generate an output that seems coherent.

12:26 To do that, you need an engine called GPUs.

12:29 The absolute newest, most advanced GPU architecture currently

12:33 on the market is called the Blackwell B200.

12:36 It's so massive that it's not even a single piece of silicon.

12:40 It is actually two separate chips

12:43 seamlessly stitched together to function as one,

12:46 packing a mind-bending 208 billion microscopic switches.

12:50 And the technology is moving so fast that in January 2026,

12:55 it was unveiled that an even more

12:57 advanced GPU architecture is already on the way, called the Rubin platform.

13:01 Because these chips are essentially the engine of the entire AI revolution,

13:06 a single one of these GPUs costs between $30,000 and $40,000.

13:11 And to train a massive AI model like ChatGPT, you don't just need one,

13:16 you need tens of thousands of them,

13:18 all wired together in massive, multi-billion dollar data centers.

13:22 By the way, if you've seen the viral clip

13:25 of the GPU that Nvidia CEO Jensen Huang gifted to Elon Musk,

13:29 you might be thinking, "Wait, that doesn't look like a tiny chip,

13:33 it looks like some kind of a heavy hard drive." And you'd be right.

13:37 What he handed to him wasn't just a GPU,

13:40 but an entire mini supercomputer that has massive cooling fans built in.

13:45 The actual piece of silicon doing the math,

13:48 the GPU itself is only about the size of a playing card.

13:52 But here's where things start to get deeply geopolitical,

13:56 because who actually controls the production of these GPUs?

14:00 As of March 2026, Nvidia is the most valuable company on the entire planet,

14:06 worth roughly $4.5 trillion.

14:09 They design the architecture,

14:10 and they essentially control the global supply of these top tier GPUs.

14:15 But here's the thing.

14:16 Nvidia doesn't actually physically manufacture the chips themselves.

14:21 They design them and then they hand those blueprints over

14:24 to be manufactured by one specific company in Taiwan, TSMC.

14:29 TSMC, or the Taiwan Semiconductor Manufacturing Company,

14:33 is the only factory on Earth capable

14:36 of mass producing these hyper-advanced chips reliably.

14:39 Today, they control nearly 70% of the entire global chip manufacturing market

14:45 and over 90% of the market for the most advanced chips used in AI.

14:50 Now, you might be thinking,

14:52 if TSMC is a Taiwanese company and China claims Taiwan is part of its territory,

14:59 why can't China just force TSMC to give them the chips?

15:03 The answer comes down to leverage.

15:05 Even though TSMC is a Taiwanese company,

15:08 their factories rely heavily on American software,

15:11 American patents and American-made machinery.

15:14 And under U.S.

15:15 law, any foreign company that uses American technology to build

15:19 a product is strictly banned from selling advanced AI chips to China.

15:23 If TSMC were to break that rule, the U.S.

15:26 would instantly cut them off from the tools they need to survive,

15:30 effectively shutting the company down.

15:32 So then you might wonder if these chips are so incredibly valuable,

15:37 why can't China just spend billions of dollars

15:39 to build their own version of TSMC from scratch?

15:42 Because manufacturing these chips is arguably

15:45 the most complex physical process in human history.

15:48 It requires decades of accumulated engineering know-how,

15:52 and it relies on extremely specialized equipment,

15:55 like $200 million laser machines built

15:58 in the Netherlands and ultra pure chemicals from Japan.

16:02 Even massive tech giants like South Korea's Samsung

16:05 have spent billions trying to catch up to TSMC,

16:08 but they still struggle to get their advanced AI chips to work as reliably.

16:13 The global bottleneck is so severe that Elon Musk recently announced

16:17 that Tesla is launching its own massive AI chip factory in the U.S.,

16:21 called Project Terafab.

16:23 He claimed that he has no other choice because third-party suppliers,

16:27 like TSMC, simply cannot manufacture enough chips to meet his future demands.

16:31 Remember when AlphaGo put fire under China's ass

16:34 to come up with a sweeping national strategy on AI?

16:38 That strategy was called the Next

16:40 Generation Artificial Intelligence Development Plan,

16:42 and part of China's plan at the time,

16:45 in order to become the global leader in AI by 2030,

16:49 was to dominate the hardware layer as well.

16:52 China's initial plan was to invest in their local tech

16:55 companies like Huawei and SMIC to start producing their own chips,

16:59 but China likely did not anticipate that in a few years later

17:03 they would be completely cut off from the global supply of top-tier GPUs.

17:08 So it seems like catching up on the hardware front

17:11 is virtually impossible for China right now, because the U.S.

17:14 and its allies have essentially blocked China from accessing TSMC,

17:19 blocked them from buying Nvidia's best GPUs,

17:22 and blocked them from buying the specialized European

17:25 and Japanese machines needed to manufacture them domestically.

17:29 You can see how this creates a massive bottleneck for China.

17:33 Sure, China is perfectly capable of pouring the concrete and building

17:38 the giant data centers and cooling systems needed to house an AI supercomputer.

17:43 That's the easy part.

17:45 But getting their hands on the actual engine to power that supercomputer,

17:49 that's a completely different story.

17:51 Because China had been locked out of the most

17:54 critical hardware supply chain in the world.

17:56 Most Western experts assumed their AI ambitions were dead in the water.

18:01 No one thought that China would be able to catch up to the U.S.

18:05 But what the West underestimated was how

18:08 much China could innovate on the software front.

18:11 Instead of brute-forcing the hardware,

18:13 Chinese engineers found a way to rewrite the rules of the game.

18:17 And that blind spot is exactly why the entire Western tech world

18:21 was caught completely off guard when

18:23 China announced an AI model called DeepSeek.

18:32 Okay, so hopefully by now it makes sense

18:34 that AI is really a stack of different technologies.

18:38 We talked about the hardware layer at the very bottom,

18:40 which is mainly those insanely powerful GPUs.

18:43 Now, sitting directly on top of that hardware is

18:45 what I referred to earlier as the software layer.

18:49 Now, to be more accurate, we should really call this the model

18:53 layer though there's a software element involved.

18:55 And this is where the actual brain of modern AI lives.

18:59 And arguably the layer where China proved that it could compete.

19:03 Now, to be fair, the hardware layer is still super important.

19:07 You cannot build your model on nothing.

19:10 Yes, China was cut off from the most

19:12 advanced AI chips coming out of the United States,

19:15 but Chinese researchers could still work with older,

19:18 less capable GPUs from Nvidia that they had stockpiled before the U.S.

19:23 export bans.

19:24 Because they were limited on the hardware front,

19:26 they had to be more innovative and squeeze far more performance

19:30 out of those chips through smarter

19:33 engineering and lower cost training strategies.

19:35 They became obsessed with efficiency, constantly asking questions like,

19:40 "How do we get better results from weaker hardware?" Or, "How do

19:44 we reduce wasted computation?" And that brings us to the model layer.

19:48 This is where terms like foundation model comes in.

19:51 Take ChatGPT for example.

19:53 The ChatGPT app is not actually the model itself.

19:57 It is the consumer-facing app wrapped around the foundation

20:01 model or the brain that ordinary people can interact with.

20:04 And the latest underlying brain powering ChatGPT

20:08 as of March 2026 is a model called GPT 5.4.

20:12 But you might be wondering, how do engineers even build the foundation model?

20:17 The answer traces back to a massive

20:19 engineering breakthrough in 2017 by American researchers.

20:23 They introduced a brand new architectural

20:25 blueprint for AI called the Transformer.

20:28 Think of the Transformer like a revolutionary new engine design.

20:32 Before 2017, AI processed text strictly one word at a time,

20:37 left to right, the same way a human reads a book.

20:41 It was slow, and if a sentence was too long,

20:45 the AI literally forgot the context of what it

20:47 was reading by the time it reached the end.

20:49 The Transformer completely changed the game because it

20:53 was mathematically wired to look at an entire,

20:56 massive block of text all at once.

20:58 More importantly, it drew invisible

21:00 mathematical connections between every single concept,

21:03 so it never lost the context.

21:06 And when you take this transformer engine and design

21:09 it specifically to process an absolute mountain of human text,

21:14 that's what we call a Large Language Model or LLM.

21:17 Now, at its absolute core, the structure of an LLM is basically just

21:22 playing the ultimate game of guess the next word,

21:25 kind of like the simple autocomplete feature

21:27 on your smartphone when you text somebody, right?

21:31 But thanks to the Transformer architecture,

21:33 the foundation model never loses context.

21:36 And because it gets trained on grammar, logic, physics, math,

21:41 and basically everything else on the internet,

21:43 its predictions go way beyond simple autocomplete.

21:46 When it guesses the next word,

21:48 it starts looking and acting exactly like complex human reasoning.

21:52 Almost every major AI company in the world, including OpenAI,

21:56 built their model on top of this exact Transformer blueprint.

22:00 But building a massive Transformer brain requires thousands of top-tier,

22:05 state-of-the-art GPUs, which again, the U.S.

22:07 had banned China from buying.

22:10 So Chinese companies like DeepSeek had to take that Transformer blueprint

22:14 and fundamentally alter it to make it cheaper and more efficient.

22:18 And they pulled off two massive breakthroughs to do it.

22:21 First, they pushed an architecture called Mixture of Experts,

22:25 or MoE, to an absolute extreme.

22:27 An expert here is literally just a cluster of artificial

22:31 neurons inside the model that specializes in recognizing specific patterns.

22:36 So if you ask a math question,

22:38 the model routes it to the cluster of neurons that recognizes math equations.

22:42 Now, in older, dense AI models, every time you ask a question,

22:48 the entire massive brain would have to light up to calculate the answer,

22:53 which consumes enormous GPU computing power.

22:56 To address this, American companies like OpenAI pioneered the use of MoE,

23:01 which essentially divides the model into multiple specialized expert clusters.

23:06 However, DeepSeek engineers took this concept to an engineering extreme.

23:11 Instead of dividing the model into dozens of expert clusters,

23:15 they slice it into 256 tiny, hyper-specialized experts.

23:20 So when you ask DeepSeek to solve a coding problem,

23:23 an ultra efficient router instantly kicks

23:25 in and activates just eight of those tiny experts,

23:29 leaving the vast majority of the brain completely asleep.

23:32 But the Chinese didn't stop there.

23:34 They paired this extreme fragmentation with a brand new

23:37 technique they invented called Multi-head Latent Attention, or MLA.

23:42 To understand what this does, think about having a conversation with an AI.

23:46 When you ask it a long, complicated question,

23:49 the AI has to constantly hold the previous parts

23:52 of the conversation in its head so it doesn't lose context.

23:56 In engineering, this is called the key value cache,

23:59 but you can just think of it as the AI short-term memory.

24:03 Normally, storing all the short-term memory takes

24:05 up a massive amount of GPU power,

24:07 but DeepSeek's new MLA technology acts like an extreme memory compression tool.

24:13 It essentially shrinks the AI short-term memory footprint by over 90%.

24:18 This allows the model to perfectly keep track of what

24:21 it is thinking about while using dramatically less memory,

24:24 which makes the entire system incredibly cheap and highly efficient to run.

24:29 Second, they also made profound optimizations to the way

24:33 hardware communicates because DeepSeek had to use older,

24:36 less powerful Nvidia GPUs.

24:38 They couldn't just rely solely on Nvidia's default software.

24:41 Usually, every AI company in the world

24:44 relies on Nvidia's industry standard software,

24:47 called CUDA, or Compute Unified Device Architecture, to run their chips.

24:52 CUDA is like the automatic transmission of AI development.

24:55 Fast, reliable, and easy to use.

24:58 But what DeepSeek did was dig deeper,

25:00 building an intermediate assembly layer within the CUDA ecosystem called PTX,

25:06 which stands for Parallel Thread Execution.

25:09 By writing highly customized, low-level code using PTX,

25:13 they shifted from automatic into manual.

25:16 They forced those older chips to communicate and process calculations

25:20 much more efficiently than Nvidia's default general purpose software allowed.

25:25 Now, could American tech companies do this, too?

25:28 Absolutely.

25:29 OpenAI, Anthropic, and even Nvidia itself regularly develops highly optimized,

25:34 low-level operators.

25:36 The difference here is all about necessity.

25:39 DeepSeek had no choice but to pursue extreme software

25:43 optimization because they couldn't access the latest high-end GPUs.

25:46 For companies like OpenAI,

25:48 who have virtually unlimited budgets and unlimited access to top-tier chips,

25:52 it is often faster and more cost-efficient to just buy more

25:56 hardware than it is to spend months meticulously optimizing low-level code.

26:01 So these incredibly complex software tricks,

26:03 along with their version of MoE and memory compression technology,

26:07 are exactly how DeepSeek managed to build a world-class

26:11 AI model for reportedly just under $6 million,

26:15 while OpenAI was spending hundreds of millions to train GPT models.

26:20 But the ultimate killer feature of DeepSeek's

26:23 model layer wasn't just the architecture itself, it was distribution.

26:28 What I mean by that is DeepSeek made the model open source.

26:32 In practice, that means researchers, startups,

26:35 and developers around the world could inspect it,

26:38 run it, fine-tune it, and build on top of it themselves.

26:42 This is also how the general public was able to understand

26:45 the key design choices inside the model in the first place,

26:49 including details about its mixture of expert system

26:52 and exactly how many experts they divided the brain into.

26:55 That is the exact opposite of the closed

26:58 model approach used by companies like OpenAI.

27:01 The bottom line is that DeepSeek effectively turned its model

27:04 into a platform and started handing out its secret recipe behind the model.

27:09 And once that happens, progress no longer depends on one company alone.

27:14 Thousands of outside engineers can start experimenting,

27:17 refining, and extending the system in parallel.

27:20 Anyway, this incredibly structured model, whether open source or not,

27:24 is still completely useless by itself without one thing to get it going.

27:28 It's pretty much just an empty brain at this point, and it needs fuel.

27:33 And that fuel is data.

27:35 And this is exactly where China's AI story gets interesting,

27:38 because once you understand that you need not

27:40 just the hardware and the model or the brain,

27:43 but the massive amounts of data to fuel it,

27:46 you realize that China holds a structural advantage that the U.S.

27:50 fundamentally cannot match.

27:52 And that brings us to the third and most visible layer of the entire AI stack,

27:57 the consumer app layer.

27:59 And the unique way China actually generates and collects its data.

28:08 To say that the arrival of DeepSeek, the Chinese open source foundation model,

28:13 sent shockwaves through Silicon Valley and the broader

28:16 Western tech community would be an understatement.

28:19 For many Western engineers and investors, it felt like a Sputnik moment.

28:23 If you aren't familiar with that term, it is a reference to 1957,

28:27 when the Soviet Union unexpectedly launched the first

28:30 satellite into space and completely stunned the United States.

28:34 In tech, a Sputnik moment is a sudden,

28:37 shocking reminder that a strategic rival is catching up much

28:41 faster than expected and doing so through a completely different path.

28:46 But if you thought DeepSeek's advantages

28:48 stopped at software tricks and model architecture,

28:51 you're missing the final piece of the AI stack.

28:53 We covered the hardware layer, right?

28:56 And then I just talked about the model layer.

28:58 Now comes arguably the most important layer of all

29:01 and the true bottleneck of the entire AI industry.

29:04 The data layer.

29:06 An AI model, no matter how elegant its design, begins as an empty brain.

29:10 It has no inherent understanding of the world.

29:13 To become intelligent,

29:15 it has to be trained on an enormous amount of human knowledge and behavior.

29:19 You've probably heard the term big data, right?

29:22 We are talking about text from the entire internet.

29:25 Books, news articles, academic papers and so on and so forth.

29:29 During the training process,

29:31 the model is repeatedly exposed to billions of patterns inside the scraped data.

29:36 Every time it reads a sentence,

29:38 it adjusts the invisible connections inside its brain over and over again,

29:42 until it becomes extraordinarily good

29:44 at predicting exactly what the most coherent,

29:47 logical answer to your question should be.

29:50 And for years, the assumption in the West was that the smartest

29:53 model would come from the companies that could buy the most human expertise.

29:58 They hired huge teams of annotators,

30:01 researchers and domain specialists to create premium training data.

30:06 Step by step math solutions,

30:08 carefully written coding examples and highly structured explanations

30:11 designed to teach the model how to reason.

30:14 That approach still worked,

30:16 but it was also brutally expensive because expert knowledge is expensive.

30:21 Imagine paying not just one lawyer, consultant or software engineer by the hour,

30:26 but thousands of them at scale to handcraft the AI study material.

30:30 But DeepSeek pushed much harder into a different approach,

30:34 reinforcement learning.

30:36 Instead of constantly showing the model

30:38 the correct reasoning process written by a human,

30:40 you let the model generate many possible answers on its own.

30:44 Then a scoring system checks the result.

30:48 If the answer is correct, clear, or useful, the model gets rewarded.

30:53 If it's wrong, sloppy, or inconsistent, it gets penalized.

30:56 So the model is not learning the way a student memorizes an answer.

31:01 It is learning more like a person

31:03 solving practical problems over and over again,

31:05 slowly discovering which strategies lead to success.

31:09 Over time, the system starts reinforcing

31:11 the patterns that produce better reasoning.

31:14 Now, at this point, you might be asking, "If this is so efficient,

31:18 why didn't OpenAI, Google, or Meta just take this approach themselves?

31:22 The truth is, American AI companies do use reinforcement learning,

31:27 but they rely on a hybrid approach.

31:29 They spend hundreds of millions of dollars on human

31:32 expertise to guarantee that the AI is highly controllable,

31:36 safe, polite, and commercial ready.

31:39 And then they layer reinforcement learning on top of that.

31:42 DeepSeek, on the other hand,

31:44 didn't have billions of dollars to spend on human tutors.

31:47 So out of pure financial necessity,

31:49 they leaned into pure reinforcement learning much

31:52 harder and much earlier in the training process.

31:55 But there is a downside to this.

31:58 When DeepSeek first tested this pure reinforcement learning approach,

32:01 they ran into a massive problem.

32:04 The AI became brilliant at logic, math and coding,

32:07 but it became terrible at communicating with humans.

32:11 Because the AI was only being

32:13 rewarded for getting the right mathematical answer,

32:15 it stopped caring about how it sounded.

32:18 DeepSeek's engineers admitted that the model started suffering from what

32:22 they call "language mixing." It would start thinking in a bizarre,

32:26 unreadable hybrid of English and Chinese.

32:28 It basically turned into a genius mathematician who mumbles

32:32 to himself and doesn't know how to speak to normal people.

32:35 So to make the model useful for the general public,

32:38 DeepSeek still had to go back and use a small amount of that expensive

32:42 human-labeled data just to teach the AI how to format its answers clearly.

32:47 But still, DeepSeek proved that you could build a core reasoning

32:51 engine of a world-class AI for a fraction of the cost.

32:54 Alright, are you still with me?

32:56 I know this is all very technical, and I'm trying my best to explain

33:00 this in the simplest way possible based on our research.

33:04 If you made it this far,

33:05 you now understand the two crucial parts of the AI race.

33:09 The chips, the reasoning.

33:11 But solving the reasoning problem is really only the beginning.

33:15 I remember when ChatGPT first came out.

33:18 I was so blown away that I kept experimenting with it nonstop.

33:22 But fast forward to today and you hear people complaining about

33:25 how models like Claude or Gemini just aren't good enough anymore,

33:30 despite how amazing they are at reasoning.

33:33 And that is because, number one, people's expectations can never be satisfied,

33:38 and number two, the real world doesn't just operate based on text.

33:42 There is a hard limit to these large language models like ChatGPT and Gemini.

33:48 Because no matter how advanced they are or how much

33:51 data they were trained on, they're fundamentally based on language.

33:55 But how do you teach an AI to understand the physical world?

33:59 You cannot just describe a sunset

34:02 or the physics of water splashing in text, right?

34:05 You need images.

34:07 You need audio.

34:08 You need video.

34:09 Basically, you need a massive amount of high definition video,

34:13 audio, and images to train the next generation of AI.

34:17 In the industry, this is called multimodal data.

34:20 If a language model is like a brain trapped

34:23 in a dark room that only knows how to read text,

34:26 a multimodal model is a brain that has been given eyes and ears.

34:31 It can process multiple modes of information at the same time,

34:35 meaning it can look at a photo, listen to an audio clip and understand

34:39 exactly what is happening in the physical world.

34:42 And that is precisely where the U.S.

34:44 AI industry is currently hitting a massive brick wall.

34:47 American AI companies trained their early model by scraping the open internet,

34:52 often without permission and probably illegally.

34:55 I'm talking about platforms like YouTube, Reddit, X, and public websites.

35:01 But from our research, as of 2026,

35:03 they've essentially run out of high quality multimodal data.

35:07 Whatever they could scrape off the internet, they already have.

35:10 Even Elon Musk has publicly warned about this, stating

35:14 that because the industry is running out of real human data,

35:18 AI companies will have no choice but to rely on what is called synthetic data.

35:22 That is essentially AI generating its own data to train itself.

35:26 Very similar to the self-teaching

35:28 reinforcement learning loops we covered earlier.

35:30 Worse yet, the real human data they

35:33 have managed to scrape is often heavily compressed,

35:36 highly fragmented, and increasingly locked behind

35:39 strict copyright lawsuits and privacy laws.

35:42 But in China, the ecosystem is entirely different.

35:45 And that brings us right back to those crazy

35:48 viral AI videos you're seeing all over the internet.

35:51 These hyper realistic videos are being generated

35:54 by Chinese AI tools like Seedance 2.0.

35:57 To understand why a Chinese company is suddenly dominating this space,

36:02 you have to understand the consumer app layer,

36:04 which is the final piece of the AI stack.

36:07 In China, the digital economy is dominated by the so-called

36:11 super apps like WeChat and video platforms like Douyin,

36:15 the Chinese equivalent of TikTok.

36:17 And companies like ByteDance,

36:19 the parent company of both domain and TikTok, doesn't just host videos.

36:24 These are all consumer-facing apps, and they operate the most efficient,

36:29 high-volume video data pipelines ever engineered.

36:32 Every single day, hundreds of millions of Chinese citizens use these apps

36:37 to upload ultra-high definition videos covering

36:39 every conceivable aspect of human life.

36:42 From cooking and dancing to complex mechanical repairs,

36:46 drone footage, and daily vlogs.

36:49 So guess who owns Seedance 2.0?

36:53 That's right, ByteDance.

36:55 Because ByteDance literally owns the platform, they possess the native,

37:00 uncompressed video files directly on their own servers.

37:03 More importantly, that video is perfectly categorized

37:06 and paired with exact user engagement metrics.

37:10 When their AI is training on this data,

37:12 it isn't just looking at a video of a person walking.

37:16 It has access to the metadata.

37:18 It knows the exact camera angle, the lighting conditions,

37:22 and the exact millisecond a human viewer lost interest and swiped away.

37:27 This is a perfectly labeled, infinitely growing database that exists

37:32 entirely behind the Chinese internet wall.

37:34 And that is precisely why Seedance can blow

37:37 away other AI video generators like OpenAI's Sora.

37:41 Because Sora still struggles with physical consistency,

37:45 audio synchronization and visual hallucinations,

37:48 largely because OpenAI hit the limits of the multimodal

37:52 data they could legally and cleanly scrape.

37:54 In fact, Bytedance's AI chatbot Doubao has already surpassed deep seek

37:59 in users and is currently the number one AI chatbot in China.

38:03 And the main reason for that is simple.

38:05 While DeepSeek is great at text and reasoning, it cannot handle images or video.

38:11 Doubao, on the other hand,

38:13 taps into Bytedance's massive data engine to seamlessly generate AI images,

38:19 cinematic videos, and realistic voices, all in one place.

38:22 So now you know why Seedance is

38:24 fundamentally better than any of its American equivalents,

38:27 due to the sheer quality and structure of its training data.

38:31 Engineers call this natural motion synthesis.

38:35 When Seedance generates a video of a person walking through a puddle,

38:39 the water splashes correctly.

38:41 The reflection matches the environment,

38:43 and the sound of the splash syncs perfectly with the visual.

38:47 And when you look at it from that perspective,

38:49 China's advantage starts to look much bigger.

38:52 Because if DeepSeek showed that China could compete in reasoning,

38:56 companies like ByteDance showed that China also sits on one

39:00 of the deepest reservoirs of multimodal consumer data in the world.

39:04 And of course, this is all possible because China has a population of 1.4

39:08 billion people and Douyin has over 750

39:12 million daily active users constantly feeding the machine.

39:16 That being said, the flip side of the coin is that these Chinese

39:20 AI tools might eventually struggle

39:22 with accessing multimodal data outside of China.

39:26 If they want their models

39:28 to perfectly understand Western culture, Western physics,

39:31 or Western cityscapes, they just cannot rely on Douyin,

39:35 the Chinese video platform, right?

39:37 They will eventually bump into the exact same

39:40 data wall problem as the American AI companies.

39:42 To me, this means that the AI race is

39:45 far from over and we are barely scratching the surface.

39:48 It's just shifting to a new battlefield.

39:51 When it comes to the future,

39:52 specifically the rise of AI agents and physical robots,

39:56 which is a whole other topic, my intuition, and I could be wrong here,

40:00 is that there's still a massive amount of real

40:03 life data waiting to be discovered and trained on.

40:06 It just might not be on the internet.

40:08 For example, at Asian Boss,

40:10 we've been conducting real-life street interviews for over a decade,

40:14 collecting and curating people's honest opinions on social and cultural trends.

40:18 What if we had the capacity to do a lot more of these straight interviews?

40:23 Or better yet, what if we created our own app

40:26 layer to help represent ordinary people's voices in a video format,

40:29 and show the world why people think the way they do?

40:33 If the future of AI requires understanding real human beings,

40:37 then maybe the most valuable data won't come from scraping websites.

40:43 Maybe it'll come from actually talking to people.

40:46 If you found this video insightful and felt like you learned something new,

40:50 I also want to let you in on something that would

40:53 be far easier to understand than what I've just explained.

40:56 Our analytics show that only about 23% of you

40:59 watching right now are actually subscribed to our channel.

41:02 And I'll tell you why that matters.

41:04 Because we put so much time and effort into our research,

41:08 we cannot really upload on a fixed schedule, like the same day every week.

41:14 So unless you actually subscribe to our channel

41:16 and turn on the notification bell,

41:19 the algorithm will often just bury our latest video whenever we upload it.

41:23 What that means is that when a new video comes out,

41:26 sometimes you won't even see it.

41:28 Even if your regular viewers of our channel.

41:30 So if you regularly watch our videos,

41:33 not just because of the interesting topics,

41:35 but because of how we deep dive and break things down,

41:39 please do me a huge favor.

41:41 Subscribe to Asian Boss and turn on the notifications.

41:44 It'll help us beat the algorithm, reach more viewers,

41:48 and build a brand big enough to truly rival legacy media outlets.

41:52 You can also participate by leaving comments, emailing us,

41:56 or filling out the form in the description

41:58 if you'd like to join our future live streams.

42:00 The goal here is to build a real community of culturally curious

42:04 people and future leaders who help power the content that we create,

42:09 so that we can be the go-to source for authentic,

42:12 nonpolitical insights on all things Asia.

42:15 I really hope you'll be a part of it.

42:18 Of course, I'm Stephen Park.

42:20 Thank you for watching all the way to the end.

42:22 And as always, stay curious.