DeepSeek: How China shocked the world
A small weekend project that took down giants
Two years ago, an app with a name that looked like someone had mashed their keyboard was released. It was called ChatGPT. The world went bonkers. In less than five days, a million people had downloaded it and were making it do everything—writing their homework, asking it deep philosophical questions like why do people eat, and getting it to draft excuses for skipping work. The company behind this digital wizardry—OpenAI—became the tech world's new favourite child.
Google (unhappy about someone else taking the first place in the AI race) sulked in a corner for about five minutes before launching its own AI app, Gemini. Now, if you Google anything, Gemini shoves its answer right in your face.
Meta (the proud owner of Facebook, Instagram, and WhatsApp) wasn't about to sit back and let Google and OpenAI hog the AI spotlight. So they rolled out their own AI, Llama 3—because, apparently, nothing says cutting-edge technology like a farm animal for its name.
Meanwhile, Microsoft’s team had a brilliant idea: why go through all the trouble of making an AI app when you can buy one? So they waved $10 billion at OpenAI and, just like that, ChatGPT was theirs.
At this point, the United States was leading the AI race and wasn’t about to let anyone else catch up. Their master plan? Ban other countries (including India and China) from buying the fancy semiconductor chips needed to power AI. Without these magical chips, building an AI was like trying to run a marathon with a broken leg and torn shoes. President Biden even passed a law that basically said, "Tough luck, world. No chips for you."
Fast forward to last Friday. The U.S. was feeling smug. Trump had just been sworn in as President (again) and was busy making grand announcements: I will make America great again. I will buy Greenland. I will conquer Canada. I will rename the Gulf of Mexico to the Gulf of the U.S. The tech industry was thrilled—finally, a President who would make sure no one else built better, cheaper tech.
Then, disaster struck.
Over the weekend, out of nowhere, a sneaky little AI app from China shot to the top of the Apple Store charts, racking up over 5 million downloads. Unlike its American counterparts, which had names that sounded like your poor aunt’s robots (GPT-4.0, anyone?), this one had a name that actually made sense—DeepSeek. Users loved it. This app cost just a fraction of what it had got the US companies to build their AI model. DeepSeek cost just $6 million to develop vs Billions of dollars that it had cost US companies (like ChatGPT) to build theirs.
Silicon Valley (the tech hub of the US) was in shock. Its brightest minds sat with their heads in their hands (while their coffees were growing cold), asking themselves three urgent questions:
How on earth had the Chinese built an AI without the fancy semiconductor chips?
How had they built it this fast?
And most importantly—how had they built it this cheap?
Every week on The Lighter Side, I take great delight in picking one news story that catches my eye. But this week’s story didn’t just catch my eye—it nearly made them pop out of my head. Yet another David vs. Goliath tale was unfolding right before me, and I simply had to document it. So I sat down, grabbed my trusty fountain pen, and started scribbling furiously in my rough notebook (I like my thoughts to marinate in ink before they go digital).
It’s a fascinating story about human ingenuity.
What is the difference between DeepSeek and ChatGPT?
Let me tell you how ChatGPT and DeepSeek work, using a library as an example.
ChatGPT is like a super nerdy robot that reads every single book in the world. Yep, all of them! It sits there for days (okay, weeks, months—who knows?), memorising every book. Once it has read everything, it becomes the ultimate expert. So, when you ask it a question, ChatGPT pulls out its encyclopaedic brain and gives you an answer.
Now, DeepSeek is a little different. Instead of trying to read every single book like an overachiever, it splits up the work! Imagine it has a group of book-smart robots as friends and they are all hanging out together. Each of them specialises in a subject. One robot reads only Math books, another reads History, and another gets all the Spanish books. So, when you ask a question, DeepSeek quickly checks which robot (or a combination of robots) is the best expert for your question and asks them to answer.
To understand why DeepSeek is more efficient in terms of cost, let us imagine that an AI model is like a school, and you (the user) are the student. You walk into the school, type out a question (just like asking your teacher), and the school (the AI model) gives you an answer.
In this school, there are hundreds of teachers, each one a specialist in a subject—Math, Science, History, you name it. The school is packed with super smart teachers who know everything.
What ChatGPT and other traditional AI models do
Every single teacher in the school would stop what they’re doing and try to answer the question, even if it’s just basic math. This wastes time, energy, and resources.
What DeepSeek does differently
Only a small group of the most relevant teachers (a set of experts) is called upon to answer the student’s question. If it’s a math question, maybe only one math teacher gives the answer if the question is simple. If it is more complex, three math teachers discuss and give an answer. The rest of the teachers continue their work, saving time and energy.
Now, let’s say one math teacher is getting too many questions while others are just sipping tea. The school principal notices this and adjusts who gets the next set of questions, so no teacher is overwhelmed while others sit idle. This ensures efficient resource use without adding extra work.
As a result, DeepSeek requires far less computing power than traditional AI models. Computing power comes from running high-end semiconductor chips. For example, ChatGPT uses around 16,000 of these powerful chips, whereas DeepSeek operates with just ~2,000.
Every time you ask ChatGPT a question, it's like turning on thousands of lights in a giant, super fancy building—a lot of power and lots of chips are needed to get you an answer. It’s like asking every teacher in the school for help. But if you ask DeepSeek the same question, it’s like turning on just a few lights to find the right teacher for your question. It uses way fewer chips, and boom, you get your answer quickly and without wasting all that energy.
Why did DeepSeek think of this approach? Why did none of the AI companies in the US think of this?
To be fair, the idea of using select experts isn’t entirely new. However, a major reason DeepSeek’s team adopted this approach can be summed up by the age-old saying: "Necessity is the mother of invention."
Remember, a couple of years ago, the U.S. declared that the rest of the world would no longer have access to high-end semiconductor chips. The assumption was simple: without thousands of these powerful chips, no one else could possibly build a competitive AI model. This, U.S. tech giants believed, would shield them from outside competition.
To understand why this spurred DeepSeek to think differently, let me tell you their story from the start: The company behind DeepSeek isn’t a traditional tech firm in China. Its founder, Liang Wenfeng, actually ran a hedge fund—which is a place that invests in things like companies, big projects etc. and earn lots of money from this.
Initially, Wenfeng simply wanted to set up a small research unit in his hedge fund to develop their own AI model. He never intended to create something that would rival the giant U.S. tech industry. His goal was much simpler: to build an AI tool that his fund could use to make better investment decisions. Otherwise, they would have had to pay U.S. tech companies a fortune to access their advanced AI models.
In 2021-22, Liang Wenfeng purchased a few thousand high-end semiconductor chips from a U.S. company called Nvidia. He had a gut feeling that relations between the U.S. and China were going to get worse.
By 2023, he set up a research unit to develop the AI model his company needed. This was going to be his weekend project.
Like he had predicted, relations between the US and China got ugly and the US banned any further sales of high-end semiconductor chips to China. This meant Wenfeng’s team would have to build their AI using only the chips they had already acquired.
Instead of hiring experienced engineers—who might rely on the same old ideas and frameworks used by other AI models—Wenfeng deliberately recruited a small team of young engineers, fresh out of college with little prior experience. He wanted people who could think differently.
They also understood a key limitation: they had very few high-end chips but could still buy an unlimited number of lower-end chips (since the US had not banned these).
Given these constraints, they developed an innovative solution: Traditional AI models (like ChatGPT) process vast amounts of data using only high-end chips—like forcing an AI to read every book in a library to become an all-knowing expert.
DeepSeek’s approach divided tasks: Instead of making a single AI process everything, they allocated specific topics (like math or geography) to smaller AI programs, allowing them to specialize. This meant that only critical tasks used high-end chips, while most processing was handled by lower-end chips.
The model worked—and spectacularly so.
They built it in less than two years (compared to OpenAI’s ChatGPT, which took over seven years).
By reducing reliance on high-end chips, they slashed costs—spending just $6 million, while other AI companies had poured tens of billions into their models.
There are plenty of memes about ChatGPT vs. DeepSeek. My favourite is this one: (For context—Iron Man, a Marvel superhero, built his powerful suit while held hostage in a cave in Afghanistan with limited resources. Meanwhile, the CEO of Stark Industries is frustrated with his team, who can’t develop the same tech despite sitting in cushy chairs and having billions of dollars at their disposal.)
This is all cool. But tell me why should the stock market around the world fall.
Nvidia, the company that makes super fancy, ridiculously expensive computer chips, was the golden child of the stock market. Everyone thought it would make piles of money, because, obviously, AI was the future, and everyone would need Nvidia’s high-end chips to run it. The stock price soared.
And then—BOOM! Reality check.
It turned out, you didn’t actually need Nvidia’s high-end chips to make AI work. Gasp! DeepSeek had cracked the code with cheaper chips, and suddenly, Nvidia’s future profits looked about as promising as a soggy sandwich.
And when people realized a company isn’t going to make as much money as they hoped - they run away from it and sell the stock. Nvidia’s stock price tumbled like a kid on roller skates for the first time.
By the end of Monday, Nvidia had lost $600 billion in value—the biggest one-day loss in U.S. history. Somewhere in a fancy office, some very rich people probably sat staring at their screens, whispering, “What just happened?”
Trivia Corner - How jigsaw puzzles got their name
The name jigsaw puzzle originated from the use of a jigsaw—a specialized saw with a narrow blade— to cut intricate shapes that form part of the puzzle.
The world’s first jigsaw puzzle was created around 1760 by John Spilsbury, a Londoner. He glued a world map onto a flat piece of wood and then cut it into pieces along the borders of the countries. His idea was that children could learn geography by assembling the puzzle. Ingenious!
You can now buy a printed copy of the ‘Lighter Side magazine’ for kids
I’ve handpicked 12 amazing real-world stories from 2024 and turned them into a delightful book. From SpaceX’s groundbreaking missions to a heroic dog saving lives in an earthquake, a Nepalese teen conquering the world’s tallest peaks, and Rohit Sharma’s emotional retirement, it’s got lots of stories and comic strips. Perfect for kids age 8-15. Stay curious, entertained, and ahead of the curve!
Get this one for INR 399/-.
Recommended Reads
If being environmentally friendly is something you want to do - start with small things you can do while traveling. It starts with how to pack your bag.
Birds need your help. What you can do to help them.
The true story of how a dog saved the lives of three kids.
To pick more stories of your own choice, here’s the ‘Lighter Side’ website - with ALL our past stories.
Loved reading it Sangeetha!