Is DeepSeek Using Gemini to Train Its AI? Here’s What We Know So Far

The AI community is buzzing with allegations that Chinese lab DeepSeek may have leveraged outputs from Google’s AI model, Gemini, to train its new reasoning-focused version, R1-0528. This model achieved impressive results in math and coding without revealing its training data, prompting scrutiny. Developer Sam Paeach observed similarities in phrasing and logical processes between R1-0528 and Gemini 2.5 Pro, further raising suspicions. DeepSeek has faced similar accusations in the past regarding its V3 model, which referred to itself as “ChatGPT,” and allegations of using a distillation method to train its models with data from larger AIs, like those of OpenAI. The AI landscape is increasingly competitive, with firms tightening security measures, including identity verification and summarizing internal model processes, to safeguard their technologies. This controversy highlights the ongoing tensions in AI development, with significant financial stakes involved in maintaining proprietary knowledge.

Source link