Game Changing AI Innovation in China

This will end the hype of the AI industry in the USA including all story telling...
Disruption, Dematerialization of Hardware and Demonetization of industries is the norm of our technological era. In this case the AI industry.
The whole AI industry is talking about DeepSeek.

So What is the disruption and innovation?

Deepseek was founded approx 2 years ago, has 200 employees, and has invested approx $5 million to develop their innovation. In contrast, OpenAI was founded 10 years ago, has around 4,500 employees, and has raised $6.6 billion in capital. Deepseek is a customer of OpenAI and could scrape their data.
Deepseek disclosed that they used a process called distillation using the Llama open source model to develop their model by asking hundreds of thousands of questions and analyzing the answers.

First what is the disruption?

While tech giants like OpenAI and Anthropic invest $100M+ just to train their AI models, Deepseek from China built an AI system matching GPT-4's performance with:

- Training costs approx $5M. There are a lot of discussions about this number that it is much higher but per training run it is definitely a factor lower than the western counterparts.

- Innovated the algorithms used (see later)

- Reduced dramatically the GPU requirements from 100,000 GPUs to only 2,000 gaming GPUs instead of specialized hardware more expensive ones. This will demonetize the GPU hardware industry with Nvidia the first victim.

- Opensource

- Multimodal capabilities


And second what is the Innovation?

1. Reduced computational needs with less decimal digits using simplicity. Instead of using 32 decimal places they found out that 8 are enough.

2. Instead of token reading they used multi-token system reading whole phrases at once which is faster & slightly less accurate initially but useful when processing billions of tokens. The result is 2X the inference hashtagspeed.

3. They built a system of specialists using less active parameters at once. From more than a trillion parameters to a few tens of billions active at once


It seems is not only Deepseek AI but more Chinese companies introducing AI applications almost at the same time, like ByteDance (the TikTok owner) Alibaba Group introducing Qwen and Moonshot AI introducing KIMI. This is clearly a China orchestrated response.

Concluding, the result is faster/cheaper/much less resource intensive AI applications.
The more the hashtag#USA blocks China to access innovative semiconductor computing resources the faster China will catch up and even leapfrog.

Any comments?