Earlier this year, DeepSeek emerged as a noteworthy player in the AI landscape with a model that competed well against leading technologies. Recently, the company announced an update to its AI model, which has sparked discussions regarding its potentially controversial training methods. Reports suggest that the latest version of DeepSeek might have used Google’s Gemini as a foundational training tool.

This insight was shared by Sam Paech on social media, where he noted that the updated model exhibits differences due to its training methodology. The developer of a competing AI platform, SpeechMap, also observed that the processes of DeepSeek closely align with those of Gemini. For context, “traces” refer to the thought processes that AI models engage in while drawing conclusions.

These allegations are not new, as DeepSeek has previously faced scrutiny for potentially leveraging other AI systems for its own training. When DeepSeek first launched, OpenAI expressed concerns that its model may have utilized ChatGPT’s outputs. This is part of the reason DeepSeek claimed its training costs were significantly lower than those of its rivals.

Distinct from traditional AI strategies that depend on raw data, DeepSeek employs a process known as distillation. This technique involves using outputs from existing AI models, allowing it to learn in a manner akin to a student absorbing knowledge from a teacher. While this method is more efficient, it raises ethical concerns, particularly given that OpenAI explicitly prohibits the use of its outputs for developing competing AIs.

Despite the ethical implications, some experts believe DeepSeek’s approach could be justified. Researcher Nathan Lambert from AI2 remarked that it may be logical for DeepSeek to create synthetic data using resources from established API models. Additionally, the ongoing US-China trade tensions have made it challenging for companies like DeepSeek to access advanced technologies, prompting them to seek alternative training methods.