News Summary

Chinese lab DeepSeek has recently updated its R1 reasoning AI model, R1-0528, which is performing well on various math and coding benchmarks. There is speculation among AI researchers that DeepSeek may have utilized Google’s Gemini AI outputs to train this latest model. This claim is supported by similarities in the language and expressions used by DeepSeek’s R1-0528 and Google’s Gemini 2.5 Pro. This isn’t the first instance of DeepSeek being accused of training on rival AI data; there are past allegations involving OpenAI’s ChatGPT as well. Despite the controversy, DeepSeek’s model outperforms some of Google’s and OpenAI’s models on certain benchmarks and can reportedly run efficiently on a single GPU. In response to concerns over data distillation, companies like OpenAI and Google are enhancing their security measures to prevent unauthorized use of their AI outputs.