At the beginning of the 21st century, everyone thought of a future (like a Wall-e) that would result in the supply and control of superstate mega-global corporations...
Ironically, I wonder if the cohesive technological prowess of the totalitarian state, which is becoming megacorpized, has given me a clue to real decentralization.
*Number 11 LOL Meta is cheap even if you get hit twice. It's cheap even if you get hit twice. It's cheap even if you get hit 200 times.
<Firm>
Morgan Brown, VP of AI at Dropbox
1.First, let me explain the background. Currently, the cost of training state-of-the-art AI models is prohibitively expensive.
Companies like OpenAI and Anthropic spend more than $100 million on calculations alone and run large data centers that require thousands of $40,000 GPUs. It's like you need an entire power plant to run a factory.
2.But DeepSeek appeared and said.
"LOL, we could do this for $5 million."
And I didn't just say it, I actually did it.
Their models outperform or match GPT-4 and Claude on many tasks. The AI industry has been hit with a "shock."
3.How was it possible?
They thought everything over again.
Traditional AI is like recording all numbers to 32 decimal places.
DeepSeek approached, "What if we only write to eight digits? That's accurate enough!" resulting in a 75% reduction in memory usage.
4.And their "multi-token" system is also noteworthy.
As an elementary school student reads, "The… cat… sat…I read it like "
DeepSeek, on the other hand, reads the entire sentence in one go. As a result, it is twice as fast and boasts an accuracy of around 90%.
When processing billions of words, this efficiency is critical.
5.But the real ingenuity is that we've built an "expert system."
Instead of making one giant AI know everything (e.g., one person plays all the roles of a doctor, lawyer, and engineer), DeepSeek designed it to call experts only when necessary.
6. Existing models must have 1.8 trillion parameters enabled at all times.
Of the 671 billion parameters, only 37 billion are enabled for DeepSeek.
It's like running a big team but calling only the necessary professionals.
7.The results are astonishing:
• Training cost: $100 million → $5 million
• Number of GPUs required: 100,000 → 2,000
• API cost: 95% reduction
• Runs on gaming GPUs instead of data center hardware
8. "By the way," one might say, "there must be some downsides!"
What's surprising is that everything is open-source.
Anyone can validate their work. The code is open, and a technical paper explains the whole process.
It's not magic, it's simply a very clever engineering.
9.Why is it important?
This broke the existing model of "only large tech companies can handle AI."
Now you don't need a multibillion-dollar data center.
I only need a few good GPUs.
10. For Nvidia, it's a scary story.
Their business model is based on selling ultra-high-priced GPUs at a 90% margin.
But if everyone can turn AI to regular gaming GPUs... The problem is clear.
11.And the point is that DeepSeek did this with a team of 200 or less.
Meanwhile, Meta's team is working on a higher salary than the entire DeepSeek training budget, but their model is not as good as DeepSeek.
12.This is a typical story of disruptive innovation.
Existing companies focus on optimizing existing processes, while disruptive innovators rethink the fundamental approach.
DeepSeek asked, "What if we approach it smarter than we put in more hardware?"
13.The impact is significant:
• AI development becomes more accessible
• a sharp increase in competition
• The "barriers to entry" of big tech companies look like small puddles
• Hardware requirements (and costs) plummet
14.Of course, big companies like OpenAI and Anthropic will not stay put.
They're probably already implementing these innovations.
But the lamp of efficiency is now out of the bottle, and we can't go back to the "let's put more GPUs" approach.
15.Last thoughts:
This moment is likely to be remembered as an inflection point later on.
It's like PC made mainframe less important, or cloud computing changed everything.
AI will be more accessible, and much cheaper.
How this change will affect current players is just a matter of speed.
'U.S stocks [2025] ISSUE arrangemet' 카테고리의 다른 글
Early morning, March (5) | 2025.01.29 |
---|---|
Let's participate in the Dipsych game (4) | 2025.01.29 |
In short, the development (4) | 2025.01.29 |
The artificial intelligence (4) | 2025.01.29 |
<First Founder vs Serial Founder> (5) | 2025.01.29 |