Welcome to Edition #90 of Gorick's newsletter, where Harvard career advisor and Wall Street Journal bestselling author Gorick Ng shares what they don't teach you in school about how to succeed in your career.
Sign up now to receive Gorick’s weekly career strategies!
→ Read time: 6 min
STORY
How DeepSeek AI got Silicon Valley’s attention
Last week, a little-known startup in China called DeepSeek AI came to market—and instantly became the #1 app on the Apple App Store.
In response, NVIDIA’s stock lost nearly $600 billion in value, the NASDAQ dropped by 3.6%, and the S&P 500 dropped by 2%.
Why?
Because DeepSeek AI claimed to have spent $6 million to develop what took OpenAI $100 million to develop.
(Missed my story on OpenAI from 2 weeks ago? Here it is.)
Here’s the backstory from the limited amount of public information available so far:
In 2022, 37-year-old engineer turned hedge fund investor Liang Wenfeng allegedly used his hedge fund, High-Flyer, to hoard 10,000 of Nvidia’s most powerful A100 computer chips.
Called GPUs (graphics processing units), these chips are used for tasks that require a great deal of processing power, such as mining Bitcoin and training AI models.
These chips are also the same ones used by OpenAI for ChatGPT.
Several months later, when new regulations in China “forced the closure of one of [High-Flyer’s] main investment products,” Liang Wenfeng saw his opportunity: To crack open his “substantial stockpile” of A100 chips (which were subsequently banned from export to China as a result of new U.S. sanctions) and to found a new project: DeepSeek.
DeepSeek’s mission? To achieve artificial general intelligence—or a “superintelligence” that not even OpenAI has developed yet. According to Liang, A.G.I. is “China’s only real chance to catch up with the United States” when it comes to AI technology.
Fast forward and on January 10, 2025, DeepSeek announced its new V3 chatbot. Similar to OpenAI’s ChatGPT, the DeepSeek-V3 is trained to answer questions, write poetry, solve logic problems, and more. It also “matched the abilities” of ChatGPT.
The difference? V3 was trained on fewer and less powerful chips than ChatGPT.
For ChatGPT, the training cost an estimated $100 million and used roughly 25,000 Nvidia A100 chips.
For DeepSeek, the training cost an estimated $6 million and used roughly 2,048 of Nvidia’s less capable H800 chips.
(And even if DeepSeek did use all 10,000 of its original A100 stockpile, DeepSeek still would have had less than 50% of ChatGPT’s stock of A100 chips).
How did DeepSeek do it? By finding a way to be more efficient with its computing resources.
Picture this: You have a skin problem and want to get it diagnosed. But, you don’t just get a dermatologist to take a look—you ask every medical specialty from cardiology to radiology to neurology to anesthesiology to oncology to psychiatry and more.
Sure, you’d get a thorough checkup from all of these doctors, but at the expense of time and money. This is ChatGPT’s method of problem solving. It’s called the “traditional transformer model” architecture.
Now, imagine if you just went straight to the dermatologist with your skin problem. This is what DeepSeek does. It’s called a “mixture-of-experts” (MoE) architecture.
The full story is still unfolding, though! In just the past week alone, OpenAI started investigating whether DeepSeek copied or “distilled” its model. Meanwhile, some AI experts call the costs reported by DeepSeek “not accurate” and “misleading.”
Allegations aside, the key takeaway remains: DeepSeek found “clever and impressive ways of building A.I. technology with fewer chips”—and, in doing so, created an “existential threat to America's AI industry.”
UNSPOKEN RULE
Find efficiencies
DeepSeek’s origin story is still unfolding, so some of the details I’ve shared in today’s story may prove to be wrong with time. Nevertheless, there’s a career lesson here:
A proven way to stand out in your career is by finding a more efficient way of doing things.
Every organization—and indeed life in general—is just a collection of processes.
If you can pick even a single process and find a way to do it more cheaply or quickly, you’ll be a star—especially if this process impacts a metric (corporate speak for “number”) that the higher-ups care about.
Work at a restaurant? There are thousands of processes. Let’s take the 1 process of taking and fulfilling an order:
1. Server is assigned to a table
2. Server approaches the table
3. Server listens to the order
4. Server writes down the order along with the table number
5. Server walks to the kitchen
6. Server tears out a copy of the order sheet for the kitchen staff
7. Kitchen staff read the sheet
8. Kitchen staff work on each part of the order
9. Kitchen staff finish each part of the order and then slip a piece of paper under the plate indicating the table number
10. Server picks up order and reads the table number written on the piece of paper
11. Server walks to the table
12. Server asks who ordered the dish
13. Server places the dish in front of that person
Are you thinking, “Don’t many servers walk around with tablets so they don’t have to walk back and forth between the kitchen?” If so, you’re proving my exact point!
Someone found a way to cut down on the number of steps—and digitizing the order process is just one of the many ways. In doing so, they not only transformed how restaurants operate but also likely changed the trajectory of their own career.
This could be you!
The next time you find yourself doing anything more than twice at work (or in your life), ask yourself 3 questions:
1. “How can I cut down on the number of steps I need to take to get _____ done?”
2. “Of the remaining steps, which ones can I automate?”
3. “Of the (still) remaining steps, which ones can I do faster, even if it’s manual?”
For example, while writing this very newsletter, I found a new way of work that I’m super excited to try more: Dictating my to-do items into my Apple Watch as they come to mind so I don’t need to pull out a pen or my phone and so I don’t end up forgetting.
Now, back to the workplace: Once you get good at this 3-step process, it’s time to find the highest impact processes that are still in your “swim lane” to solve.*
As I write about on p.230 of my book, The Unspoken Rules, you’ll want to figure out what matters to those who matter:
“The better you understand what matters to those who matter, the better your odds of making an impact—and the better your chances of getting promoted.”
Start by paying attention to what the higher-ups are complaining about at work. What’s wasting their time? What’s causing them stress?
Find the process, make the process more efficient, and sell your outcomes—and it’s only a matter of time before people start noticing.
*A swim lane is corporate speak for “who’s in charge of what.” But be careful! People’s swim lanes don’t always match their formal job titles. It’s important to understand your organization’s hierarchy as you want to find an impactful yet unoccupied swim lane.
What’s an “unspoken rule”? They’re the things that separate those who get ahead from those who stumble—and don’t know why. You can learn more about these rules in the workplace in my Wall Street Journal bestselling book called—you guessed it—The Unspoken Rules.
Every newsletter is free and a fraction of my work.
Here are 4 of my paid offerings that may interest you:
1. Keynote speaking: My 2025 calendar is filling up! If your organization is looking for speakers for graduation season, manager training, new hire / early career orientation, AAPI Heritage Month, ERG summits, or something else, let’s chat!
2. HOW TO SAY IT: Flashcards that teach you to know what to say in every high-stakes professional setting. Get 260 fill-in-the-blank scripts and self-guided video tutorials. Free shipping on all orders over $40.
3. Fast Lane to Leadership: My online course that takes you from day 1 in a new role through to a promotion with 28 modules and 28 cheatsheets (3.5 hours of content).
4. The Unspoken Rules: My Wall Street Journal Bestseller that Arianna Huffington calls “a blueprint for anyone starting their career, entering a new role, or wanting to get unstuck.” Used by top companies and MBA programs.
Sources:
- became the #1 app on the Apple App Store
- NVIDIA’s stock lost nearly $600 billion in value
- the NASDAQ dropped by 3.6%
- the S&P 500 dropped by 2%
- what took OpenAI $100 million
- DeepSeek’s model was “impressive.”
- Liang Wenfeng
- Called GPUs (graphics processing units)
- forced the closure of one of [High-Flyer’s] main investment products,”
- “substantial stockpile” of A100 chips
- “superintelligence” that not even OpenAI has developed yet.
- “China’s only real chance to catch up with the United States”
- “Magic Cube AI supercomputing cluster.”
- “matched the abilities”
- For ChatGPT, the training cost
- For DeepSeek, the training cost
- “traditional transformer model” architecture
- “mixture-of-experts” (MoE) architecture
- “distilled” its model
- “not accurate” and “misleading.”
- clever and impressive ways of building A.I. technology with fewer chips”
- “existential threat to America's AI industry.”
