DeepSeek may have used Google’s Gemini to train its latest model

Last week, Chinese Laborcy released The DeepSee Updated version of R1 Reasoning AI This is a good performance of a number of math and coding criteria. The company has not revealed the source of the information used to raise the model, but some AI researchers assume at least one part of the EU’s Google’s Gemini family.

Sam Paech, which creates an Melbourne-based developer, which creates “emotional intelligence” for the AI, is the proof of Deepseek’s latest model prepares performances from the twins. DeepSeek model called R1-0528, Google’s Gemini 2.5 Pro’ın Similar words and phrases said at once X writing.

This is not a smoking weapon. However, another developer, the player of the “Free Speech Assessment” for the AI SphmaThe DeepSeek model celebrates the traces of the model – the “Thoughts” creates a conclusion – “Read as twins prints”.

Deepseleek is previously charged with the preparation of information from the opponent’s AI models. Developers in December observed The V3 model of the DeepSeek has often defined itself as a chairgpt as an open AI-Waterproof ChatBot platform, which offers Chatgpt chat logs.

At the beginning of this year, Openai explained financial time Deepseek, to prepare AI models, to develop AI models by extracting information from larger, more skillful, found evidence to develop AI models. According to BloombergMicrosoft, a close Openai joint and investor, found that in late 2024, the Openai developer accounts revealed that the accounts are related to DeepSeek.

Distylation is not an unusual experience, but Openai’s service conditions prohibit customers using model performances to build a company competitive AI.

Many models to be clear dishonorable themselves And the same words and expressions are combined. Because of the open Internet where the AI ​​companies are welding the main part of the training information littered with ai clap. Content uses AI to create farms Clitbaitand bots are reddit floods and X.

This “contamination”, if you want it did it quite difficult Thoroughly filtering AI copies from training information.

Again, AI specialists, a researcher like a non-profit research institution AI2, a researcher like Nathan Lambert, I do not think it is outside the question that Google is taught by the twins.

“If I were DeepSeek, I would definitely create a ton of synthetic information from the best API model there” said Lambert write In an article in x. “[DeepSeek is] Flow with short and cash in GPUs. For them, literally calculated more efficiently. “

AI companies take security measures to prevent distillation.

Openai started in April demanding Organizations to complete an identity verification process to access certain advanced models. The process requires an ID card provided by one of the countries supported by Openai’s API; China is not in the list.

In another place, Google has recently a step that has recently been created by the models created by the models available through the AI ​​Studio Developer Platform, more complicating performance rival models for marks in Gemini. Said it could be anthropic Start generalization The traces of their own model are a need to protect the “competitiveness”.

We appealed to Google for comments and if we hear back, we will update this piece.