DeepSeek's cost-saving AI tech might lead to inaccurate results: Study
What's the story
A recent study has raised concerns over the memory-saving technology used by DeepSeek, a Chinese artificial intelligence lab. The research, which is yet to be peer-reviewed, was conducted by ByteDance researchers along with co-authors from Princeton University, Stanford University and the University of California, Berkeley. The study found that DeepSeek's approach to reduce memory for processing long inputs could compromise reliability.
Model performance
Inconsistencies in model performance
The study found that DeepSeek's models are more likely to miss the same information when it appears in certain positions within a document.
This issue was observed in DeepSeek-V4, V4.1, and all compressed models pretrained by the lab.
The findings suggest that while DeepSeek's cost-saving measures may make its AI more affordable, they could also lead to inconsistencies in model performance.
Technology aim
Cost-saving measures may lead to reduced reliability
DeepSeek's technology is designed to save memory by compressing models, but it may compromise their performance.
This approach has helped the lab compete with US rivals such as OpenAI and Anthropic by offering a low-cost alternative.
However, the new research suggests that these cost-saving measures may come at a price, reducing reliability in information retrieval from documents.