- One more Thing in AI
- Posts
- Wall Street was Deeply stung by DeepSeek this week!
Wall Street was Deeply stung by DeepSeek this week!
Latest edition of One More Thing in AI Newsletter.
Renjit Philip
February 01, 2025 • Estimated Reading Time: 8 minutes
Date: 1-Feb-2025
Hey AI enthusiast,
Welcome to the latest edition of the One More Thing in AI newsletter.
We dive deep into DeepSeek (sorry, could not avoid doing that!)
DeepSeek App shot to #1 position on the IOS store in a few days after its release.
Best regards,
Renjit
In this edition:
DeepSeek: The AI Underdog Shaking Up Big Tech
A Chinese startup just rocked AI. DeepSeek built a model that matches OpenAI’s best—at 1/200th the cost.
Wall Street freaked out. Nvidia crashed 17%. The Nasdaq dipped. Investors realized: AI might not be as expensive to build as they thought.
Image by Bravo research shows the extent of the pull back in the S&P 500
The $5.6 Million Shock
DeepSeek’s new R1 model is no toy. An economist tested it on Trump’s 25% tariffs (to test the impact on the US economy). R1 crunched the numbers in 12 seconds, matching a major bank’s estimate. Get this:
• OpenAI and Google spend $1B+ training AI.
• DeepSeek did it for $5.6M—with 10,000 chips instead of hundreds of thousands.
Why Big Tech Is Nervous
Tech giants are burning billions to dominate AI:
• OpenAI, Oracle, and SoftBank: $500B AI bet
• Microsoft: $80B on data centers
• Meta: 1.3 Million AI chips by 2025
DeepSeek’s success suggests spending big isn’t the only way to win. When you have no resources as an underdog, you tend to get resourceful!
The Open-Source Wild Card
Unlike OpenAI and Anthropic, DeepSeek’s model is open-source. Anyone can use, tweak, and improve it—lowering barriers for startups. Generally, in the world of tech, open source and free usually win. Look at Linux and Android and how they beat Windows and IOS and others.
But open AI isn’t all sunshine. Big Tech argues closed models are safer, reducing bias and misinformation. DeepSeek, for instance, dodges politically sensitive topics in China (ask it about Tiananmen Square, for example).
The AI Power Struggle
Some U.S. lawmakers are pushing for upwards of $32B/year in AI funding to compete with China. DeepSeek’s rise questions that logic. Do we need more government spending, or just fewer regulations?
The Takeaway- DeepSeek’s rise changes the AI game:
• AI is getting cheaper to build.
• Big Tech’s lead isn’t guaranteed.
• Startups can still disrupt the giants.
The AI race is heating up.
How DeepSeek Outsmarted Big Tech and Washington
Using a trick called distillation, it learned from bigger models instead of training from scratch. Now, Silicon Valley is panicking.
The Einstein Hack- Distillation
Distillation is like grilling Einstein for a few hours and walking away with 80% of his knowledge.
• OpenAI and Google train from scratch, burning mountains of cash.
• DeepSeek asked existing AIs millions of questions—and learned fast. Makes sense, right?
This Hack is called the Mixture of Experts method.
DeepSeek didn’t invent Mixture of Experts (MoE), but it made key improvements.
How It Works
• Instead of one giant AI model, MoE splits the model into smaller specialized networks (experts).
• When processing text, the AI picks the right experts for each task, using only a fraction of its total parameters.
• This reduces computation and speeds up training.
DeepSeek’s Key Tweaks
1. Memory Efficiency:
• AI models store massive amounts of data while processing text.
• DeepSeek developed Multi-head Latent Attention, which compresses this data, saving memory and making communication between chips faster.
2. Quantization:
• Normally, GPUs store AI model numbers using 16 or 32 bits.
• DeepSeek reportedly used 8-bit precision, cutting memory use in half while maintaining performance.
3. Better GPU Connections:
• AI training isn’t just about raw power—it’s about how fast data moves between GPUs.
• DeepSeek optimized how its 2,048 GPUs communicate, improving efficiency without needing more hardware.
The result? A cheaper, faster AI model compared to the hundreds of thousands used by OpenAI and Google.
The U.S. Export Loophole
Washington had tried to block China from getting top AI chips. But Nvidia, looking to keep business flowing, built a “weaker” chip for China that was still almost as powerful.
Observers believe DeepSeek bought those chips before the rules tightened. Some suspect they also stockpiled high-end Nvidia chips before restrictions hit. U.S. officials admitted Nvidia followed the law—but found a way around the spirit of the ban.
Why This Matters
DeepSeek didn’t just prove China could compete in AI—it exposed flaws in U.S. strategy. The U.S. spent billions restricting AI chips, but China still found a way. Silicon Valley assumed only the richest companies could build leading AI—DeepSeek proved them wrong.
DeepSeek R1 vs. OpenAI o1: A Head-to-Head Comparison
Feature | DeepSeek R1 | OpenAI o1 |
|---|---|---|
Core Strengths | Strong in reasoning, problem-solving, and cost efficiency. | Strong in reasoning, problem-solving, and industry leadership by virtue of installed base. |
Transparency | Shows step-by-step reasoning using "chain-of-thought" logic. | Uses "chain-of-thought" reasoning but does not display its steps. |
Performance | Competitive with OpenAI’s o1, outperforms older models like o1-mini. | Still leads in some problem-solving areas but faces stronger competition. |
Cost Efficiency | Trained on an estimated $5M worth of chips, making it significantly cheaper to develop. | GPT-4 training cost is estimated to be $100M+, and GPT-5 is projected to cost an additional $500M+. |
Processing Approach | Uses "mixture of experts", breaking tasks into specialized sub-models to reduce chip requirements. | Trains AI from scratch using vast amounts of data, requiring extensive computing power. |
Privacy & Security | Open-source but censors political topics related to China. Users can download and modify the model. | Proprietary model with strict safety policies to prevent misuse. No political censorship. |
Open-Source vs. Proprietary | Released model weights for public use but did not disclose training data. | Fully proprietary; users pay to access and cannot modify the model. |
Business Model | Free to download and use; businesses can customize and host it independently. | Subscription-based model with built-in security and ongoing updates. |
Adoption & Market Impact | R1 app is a top download on the U.S. App Store, showing rapid adoption. | Said to have 300 Million monthly active users. Still dominates enterprise AI applications but faces pricing pressure from open-source competition. |
Graph by Bravo research shows the low operating costs of using DeepSeek on Azure
Reference:
Do you want to reach a high-quality audience?
To sponsor this Newsletter, subscribed to by startup founders and business leaders:
Make your voice heard!
I would love to know more about what you like and don’t like, specifically in this Newsletter. Please take this 2-minute survey to help me produce more useful content for you.
Reply