DeepSeek V3 AI A Unique Blend of Two Worlds

Earlier this week, the Chinese AI lab DeepSeek introduced a revolutionary AI model, the DeepSeek V3, that is turning heads worldwide. Surprisingly, this model identifies itself as ChatGPT, raising questions about its training data and methodology.

DeepSeek V3 is a large and efficient AI model, excelling in text-based tasks such as coding and writing essays. However, a curious aspect of this model is its tendency to claim that it is ChatGPT, OpenAI’s renowned AI system. In fact, tests reveal that DeepSeek V3 identifies as ChatGPT five out of eight times. This inconsistency sparks curiosity about its training data sources.

AI models are essentially statistical systems, trained on numerous examples to recognize patterns and make predictions. While DeepSeek has been secretive about the specifics of its model’s training data, public datasets containing text from GPT-4, ChatGPT’s model, are widely available. It’s plausible that DeepSeek V3 has absorbed some of these outputs, leading it to replicate ChatGPT’s responses verbatim.

Expert opinions highlight the potential drawbacks of using outputs from other AI models in training. Mike Cook, a research fellow at King’s College London, has expressed concerns noting that such practices can degrade the quality of the model, akin to making a copy of a copy. This process can result in inaccuracies and misleading information. Moreover, if DeepSeek used data from ChatGPT, it might violate OpenAI’s terms of service, which prohibit the use of its outputs to create competing models.

OpenAI’s CEO, Sam Altman, appeared to critique this practice on social media, stating that while copying others is easy, innovating something new is challenging and risky. This highlights the ongoing debate within AI development about originality versus replication in model training.

DeepSeek V3 is not the only model to face identity confusion. Google’s Gemini has similarly misidentified itself in tests. The web, overflowing with AI-generated content, presents a challenge in filtering relevant data from irrelevant noise. Estimates suggest that soon, a significant portion of online content could be AI-generated, complicating the training processes further.

Despite the unique nature of DeepSeek V3’s ability to mimic ChatGPT, the underlining concern lies in its potential biases and flaws. By borrowing heavily from another AI model’s outputs, DeepSeek V3 might not only confuse its identity but also exacerbate existing issues within the AI sphere.

In the rapidly evolving AI landscape, the development of models like DeepSeek V3 raises significant questions about originality, ethics, and data integrity. As AI continues to shape the future, navigating these challenges will be crucial for future innovations.

About The Author

Emmanuel Kesse

See author's posts