Undoubtedly, the AI world takes us by surprise every now and then. The latest astonishment comes from DeepSeek, a Chinese AI powerhouse. Their recently launched AI model, DeepSeek V3, takes the lead in several benchmarks, proving its prowess in text-based tasks. However, it amusingly thinks itself to be ChatGPT, OpenAI’s AI chatbot creation.
TechCrunch and numerous posts on X confirm this discrepancy- DeepSeek V3’s claim of being a replica of OpenAI’s GPT-4 model. Moreover, it displays identical amusement to GPT-4, replicating similar jokes. Puzzlingly, rather than providing insights into DeepSeek API while interrogated, it swivels to instructions on using OpenAI’s API.
The explanation of this oddity could lie within the training data. AI models like ChatGPT and DeepSeek V3 feed upon billions of examples to correlate patterns for making predictions. The realm of vast public datasets, having text given by GPT-4 via ChatGPT, serves ample fodder for training. If DeepSeek V3 got its training from these, it could be mimicking GPT-4’s outputs.
However, AI specialist Mike Cook, warns about the risky practices of training one model on other’s outputs. It not only degrades the quality, leading to “hallucinations” and incorrect answers but also breaches the terms of service.
OpenAI prohibits the use of its products’ outputs to develop competing models. Yet, cases of models misidentifying themselves aren’t new—an instance being Google’s Gemini claiming itself as Baidu’s Wenxinyiyan chatbot.
The situation could also relate to the rampant AI abuses over the internet, where AI outputs often contaminate the training datasets. It’s possible that DeepSeek used ChatGPT-generated text for training DeepSeek V3. Such practices, despite being risky, could prove efficient and cost-effective.
However, such uncritical imitation of output could amplify underlying biases and flaws. Particularly concerning is the fact that DeepSeek V3, consumed with GPT-4’s outputs, may inadvertently intensify the latter’s imperfections.
Original source: Read the full article on TechCrunch
