Chinese AI lab DeepSeek has released an updated reasoning model, R1-0528, which is reported to perform well in math and coding benchmarks.

However, concerns have been raised regarding the potential use of data from Google’s Gemini AI family in training this model.

Developer Sam Paech, based in Melbourne, shared evidence on social media indicating that R1-0528 shows similarities to Google’s Gemini 2.5 Pro.

Another developer, known for creating SpeechMap, also noted that the reasoning patterns of R1-0528 resemble those of Gemini AI.

DeepSeek has not disclosed the sources of data used for training the model.

🔗 Source: TechCrunch

🧠 Food for thought

1️⃣ Model distillation creates an ethical gray area amid fierce AI competition

Distillation, the process of training smaller models using outputs from larger ones, has become a contentious but widespread practice in AI development, especially for companies with limited computing resources.

While distillation itself is a legitimate technique, DeepSeek’s alleged use of competitors’ models highlights the intellectual property challenges in AI development, with previous accusations suggesting they used OpenAI’s outputs without authorization1.

This case illustrates a technical reality: companies like DeepSeek, which are “short on GPUs and flush with cash,” may find it economically rational to create synthetic data from competitors’ models rather than building everything from scratch2.

The increasing adoption of protective measures by major AI labs, such as OpenAI requiring ID verification from countries that exclude China or Google summarizing model traces, demonstrates how seriously these companies view the threat of unauthorized knowledge transfer3.

These protective measures reflect a broader industry recognition that model weights represent the culmination of substantial investments, making them valuable intellectual property worth safeguarding4.

2️⃣ AI contamination creates attribution challenges for researchers and companies

The difficulty in definitively proving model copying stems partly from the growing “contamination” of the open web with AI-generated content, making it increasingly challenging to determine a model’s true training sources.

As content farms flood the internet with AI-generated text and bots populate platforms like Reddit and X, the lines between human-created content and AI outputs are blurring, complicating efforts to create “clean” training datasets5.

Recent DeepSeek developments