DeepSeek R1 is said to be just as powerful as OpenAI's o1, but DeepSeek's model only cost $5.6 million to train - a modest figure compared to the more than $100 million OpenAI invested in GPT-4.
DeepSeek development cost was not $5.6 million as claimed
Not everyone is convinced by DeepSeek's groundbreaking claims, though, including 2024 Nobel Prize in Chemistry winner Demis Hassabis — who, along with John Jumper, was awarded the prize for successfully using AI to predict the structure of most proteins.
At the AI Action Summit in Paris, France, Demis Hassabis, CEO and co-founder of Google DeepMind, made some interesting remarks about DeepSeek. In his statement, Mr. Hassabis admitted that the model was impressive and said that the DeepSeek development team was one of the best in China.
Impressed with DeepSeek but not the optimal solution
However, he also cautioned that many of DeepSeek's claims may be exaggerated and misleading. Hassabis explained that the cost DeepSeek published was only for the "final training run," which does not reflect the full cost of developing an AI model from start to finish. He also argued that DeepSeek "relies on Western models for refinement," a point OpenAI emphasized shortly after DeepSeek's launch.
Ultimately, Hassabis asserts that while DeepSeek is impressive, Google doesn’t consider it a “perfect solution” technologically. He says Google’s Gemini model is more efficient than DeepSeek in both cost and performance. “We just don’t talk about it much,” the 2024 Nobel Prize in Chemistry laureate stresses.
The launch of DeepSeek may be a step forward in the AI industry, but is it truly revolutionary or just a marketing gimmick as Hassabis claims? Only time will tell.
Source: https://thanhnien.vn/giam-doc-ai-google-ha-thap-dot-pha-cua-deepseek-185250211103753014.htm
Comment (0)