DeepSeek R1 is coming, Meta executives are panicking

Recently, Meta engineers expressed their feelings on the anonymous social platform TeamBlind, revealing the tremendous pressure brought by the AI ​​model R1 developed by the Chinese company DeepSeek. DeepSeek is considered to be the world's first open source AI model on par with OpenAI's o1 model. Compared with OpenAI, R1 is not only more open source, but also has an astonishingly low training cost of only $5.5 million.

In contrast, the annual salary of Meta's executives often exceeds the training cost of the entire DeepSeek V3, which makes Meta's management feel embarrassed.

According to internal sources, the emergence of DeepSeek V3 has put Meta under considerable pressure last year. Meta's engineers are working hard to analyze DeepSeek's technology, hoping to replicate the key technologies as soon as possible. In the past, AI researchers around the world were chasing big American models. But now, the situation has changed, and American engineers have begun to reverse engineer Chinese AI technology.

DeepSeek not only surpassed Meta's Llama4 in benchmark tests, but also attracted the attention of the industry in its performance in multiple fields. DeepSeek achieves extremely high inference performance through technologies such as large-scale reinforcement learning (RL) and unsupervised fine-tuning (SFT). This kind of technological confidence has caused some netizens in the United States to begin to reflect on China's rapid rise in the field of AI.