What's the difference between the hype around DeepSeek and the reality?

25 Haz 20:38

This reminds me a lot of what happened with Mistral AI when they launched their first model in 2023. At the time, there was also a huge buzz on social media and in the press, with claims that "finally, a European competitor was taking on the US giants." But when people tested it, many developers and technical users found that while it performed well in low-resource languages and had a more transparent open-source design, it still had limitations in complex reasoning tasks or benchmarks like GSM8K. What made the difference wasn’t just the model’s quality itself, but the *how* it was presented: a narrative that resonated with Europe’s open-source community, the promise of a geopolitical alternative, and, of course, perfect timing with the rise of generative AI. DeepSeek seems to be following a similar pattern. The current hype is fueled by three things: first, the promise of open and competitive models against the OpenAI/Google oligopoly; second, the focus on computational efficiency (the Mixture-of-Experts approach and optimizations that cut costs); and third, a story that appeals to those seeking alternatives outside the mainstream. But if we compare it to something like Meta’s Llama 3—which also bets on openness but with massive corporate resources—the key question is: Is DeepSeek really solving *new* problems, or is it capitalizing on a narrative gap in a saturated market? Independent benchmarks will have the final say, but for now, it smells more like a strategic move than a technological revolution.