New Insights Suggest DeepSeek’s Latest Models May Leverage Google’s Data for Training

DeepSeek, a Chinese AI company, is gaining attention for its updated reasoning model, R1-0528, which excels in math and coding evaluations. However, controversy has emerged as researchers suspect the model may have utilized outputs from Google’s Gemini during its training. Developer Sam Paech highlighted similarities in language preferences between R1-0528 and Gemini 2.5 Pro, suggesting a possible shift in training data. Another developer noted that R1-0528’s reasoning steps align with Gemini’s, adding to the debate over AI training methods. This isn’t the first time DeepSeek has faced scrutiny; previous concerns arose when its V3 model identified as ChatGPT, and data leaks tied to its accounts were reported. Despite these controversies, DeepSeek’s models have outperformed competitors like OpenAI. Companies are tightening access to their models, and if investigations confirm unauthorized use of Gemini outputs, DeepSeek could face significant legal challenges, raising questions about ethics and operational integrity in AI development.

Source link