DeepSeek, a Chinese start-up, made waves in the finance industry last month by unveiling an artificial intelligence system that required significantly fewer computer chips than traditional systems. While most A.I. companies use supercomputers packed with 16,000 specialized chips or more, DeepSeek only needed about 2,000, cutting costs dramatically.
The company detailed its innovative methods in a research paper, revealing that it used a technique called “mixture of experts” to split the system into smaller neural networks focused on specific fields, paired with a generalist system to coordinate interactions. DeepSeek also simplified calculations by using fewer decimals during training, making the process more efficient.
Additionally, the company utilized a method that involved squeezing numbers into smaller bits of memory, sacrificing some accuracy in individual calculations but still producing a powerful neural network overall. By stretching the answers of multiplication problems across more memory bits, DeepSeek was able to maintain precision where it mattered most.
While these methods may seem simple, they required a high level of expertise and specialized coding to execute effectively. Other A.I. labs may already be using similar techniques behind closed doors, but the success of DeepSeek has demonstrated the potential benefits of taking risks and investing in experimentation to find breakthroughs in A.I. technology.
By sharing its innovative methods with the A.I. community, DeepSeek has paved the way for significant cost reductions in building A.I. systems, potentially setting a new standard for efficiency and affordability in the industry.
Note: The image is for illustrative purposes only and is not the original image of the presented article.
