Deepseek has swept the world this week with a series of AI models trained with hardware that is much less powerful than those used by Western AI developers like Openai. Reliable surprised investors in the fear that companies like NVIDIA could have been very overestimated.
Lecun explained the sale as a “major misunderstanding” and explained that it is a Deepseek’s innovative approach to improve the training efficiency that promotes success.
Important connection news and insights -directly to the reception tray
“Deepseek is profitable from open research and open source, for example, Meta Pytorch and LLAMA,” Linkedin posted. “Looking at DeepSeek’s performance, people who think that” China is beyond the United States with AI. ” You are reading this by mistake. The correct reading is “Open source model exceeds its own model.”
Lecun, a winner of the Turing Award, and the vocalist of the open source AI, not relying on stronger hardware, but improves his architecture such as his JEPA concept for self -learning. I claim that there is a progress.
“(DeepSeek) came up with a new idea and built them on other people’s work,” Lecun said. “Because their jobs are published and open source, everyone can make a profit from it. That is the power of open research and open source.”
After appearing in the Davos’s World Economic Forum, Lecun claimed that open source AI is the key to building a variety of AI systems and tools, and added: From all languages and culture of the world. “
The market crash, following DeepSeek, has occurred after the team behind that the open source Deepseek-V3 model was developed for only $ 6 million.
Investors were surprised after Microsoft, Openai, Oracle, XAI, Google, etc. built AI infrastructure and spent billions of dollars to snap high -end hardware. Pocket change
In a later thread post, Lecun stated that stock sales were due to “major misunderstandings” on AI infrastructure investment.
“Many of these billions are in infrastructure for inferences, not training,” Lecun says. “To execute AI assistant services for hundreds of billions, a lot of calculations are required. If you place an AI system understanding, inference, large memory, and other functions, the reasoning cost will increase. Masu.
“The only real problem is whether the user is happy and happy to justify Capex and OPEX.”
Back behind DeepSeek’s hype: Explanation of cost, low risk, censorship
China’s Deepseek AI will erupt global technical shifts and challenge us great technology.
Is DeepSeek the future of AI? Industry reaction to the launch of R1
