Why is Dipsyk a shock (to the US)?

Why is Dipsyk a shock (to the US)?

✅ Existing U.S. AI companies are investing a lot of money to develop LLM
✅ Suddenly, Dipsyck said, 'We can do it much cheaper.' Show & probe
✅ Fundamentally Rethink and Develop LLM Development Methods
✔️ 75% reduction in memory usage
✔️ If an existing LLM is read word by word, DeepSeek reads the entire sentence at once, twice faster and 90% more accurate
✔️ Instead of traditional LLM calling all agents at once, DeepSeek calls agents only when needed
✔️ Existing models require 1.8 trillion parameters to be enabled, but DeepSeek only enables 37 billion of 671 billion parameters (call agents only when needed)
✅ Results
💰️ Training cost: $100 million → $5 million
💰️Number of GPUs required: 100,000 → 2,000
💰️ API cost: 95% less
✅ Even open-source code is released, so anyone can verify it
✅ Traditional LLM was created by big tech companies with huge amounts of money, while DeepSeek was developed with a small number of people
✅ A big part of Nvidia's current revenue is its ultra-high-end GPUs, such as the H100 for AI research. But the DeepSeek team shows it's also possible with regular gaming GPUs
✅ Meta has a much higher number of employees and a much higher salary than DeepSeek, but has a much lower performance

❗ Dipsyck belongs to Chinese hedge fund company Huanfang Quant. Quant is great!