DeepSeek, a Chinese AI lab, recently updated its R1 reasoning model, which excels in math and coding benchmarks. However, the source of its training data remains undisclosed, leading to speculation that it may include outputs from Google’s Gemini AI. Developer Sam Paech claimed to have evidence suggesting that DeepSeek’s model, R1-0528, resembles Gemini outputs, while another anonymous developer noted its “thoughts” align with Gemini traces.
Previously, DeepSeek faced allegations of relying on data from competitors, notably OpenAI, leading to suspicions of unethical practices like distillation—using outputs from superior models. OpenAI indicated it had found evidence linking DeepSeek to this practice, prompting increased security measures in AI firms to deter data extraction.
Despite these controversies, experts acknowledge that AI outputs can overlap due to contamination from “AI slop” on the web, complicating the issue of replicating rivals’ models. Security enhancements from companies like OpenAI and Google aim to address these challenges.
Source link
