Українська правда

Chinese DeepSeek again used someone else's model to train its own, this time Google Gemini

Chinese DeepSeek again used someone else's model to train its own, this time Google Gemini
deepseek
0

Chinese company DeepSeek, which recently unveiled an updated version of its R1 AI model, may have partially trained it on Google's Gemini, according to developer Sam Page, who researched the updated R1-0528.

According to him, the DeepSeek model shows linguistic similarities to Google Gemini 2.5 Pro. According to TechCrunch, another researcher under the nickname SpeechMap drew attention to the similarity of the "thought trails" - the intermediate reasoning of the model during the solution of problems.

This is not the first suggestion that third-party models were used in DeepSeek training. Previously, it was found that the V3 version sometimes presented itself as ChatGPT, which could indicate that it was trained on OpenAI responses. In January, the Financial Times reported that OpenAI had found evidence of using distillation techniques to extract data from more powerful models, likely through accounts associated with DeepSeek.

Although distillation is a common approach, OpenAI prohibits the use of its models to create competitive solutions. As a result, major AI developers, including OpenAI, Google, and Anthropic, have tightened security measures, including introducing customer verification and hiding intermediate model data.

Google has not yet provided any official comments on the situation or indicated whether it has launched an investigation into the allegations against DeepSeek.

Share:
Посилання скопійовано
Advert:
Advert: