Last week, a model

1
Last week, a model called DeepSeek appeared like a comet and turned the AI industry upside down. It has almost the same performance as existing models, but the cost is almost the same as that of a Big Tech executive. Of course, Google, openai, and Anslogic seem to be in a collective shock. (This is about 27 times the cost savings compared to OpenAI O1...)
It seems all the more meaningful because of the skepticism that the scaling law has reached its limit. DeepSeek proved that AI can be created in a "smart way" rather than "more computing power." Anyway, everyone is looking back on how inefficient the existing methods have been. To put it the other way around, it has definitely proved that cost-effective methods are possible in artificial intelligence development.

2
How was DeepSeek-R1 able to achieve high performance at low cost?
To paraphrase it very roughly, the existing method checks all the books in the library one by one to solve a problem. Naturally, it takes astronomical time, effort, and cost. However, DeepSeek can be said to be a method of quickly finding only books in the necessary fields by using the library's index system.
The learning process itself is much more efficient. First of all, as before, supervised learning is rarely used, and an initial model is created only with reinforcement learning. It is much faster and more efficient because it trains with only a relatively small amount of high-quality data without human supervised learning. After that, it seems that they are learning step by step through reasoning-oriented reinforcement learning, rejection sampling and map fine-tuning, and reinforcement learning for all scenarios. It reminds me of building basic learning skills step by step through middle school high school, rather than having an elementary school student solve all college-level problems and die.

3
Small and medium-sized enterprises have also shown the potential to create efficient AI, but this does not always mean that David can topple Goliath in the future. On the contrary, there is a possibility that the technology gap will widen if big tech companies adopt this efficient method and combine it with their enormous resources.
DeepSeek's innovation demonstrated the possibility of "creating AI more efficiently," not "creating AI at less cost." It's like introducing a systematic indexing system from the way all books were checked. But big techs with larger libraries may have a stronger competitive edge with this efficient indexing system.
Anyway, it seems to be a historical event that suggested the possibility of solving problems such as astronomical cost and power by optimizing datasets and training methods.