DeepSeek officially launched its app on iOS and Android marketplaces last month. On Jan. 20, the company quietly open-sourced its inference model, DeepSeek-R1. According to the National Business Daily, as of Feb. 9, the DeepSeek app had surpassed 110 million cumulative downloads, with weekly active users peaking at nearly 97 million.
On Thursday (Feb. 20), NVIDIA CEO Jensen Huang made his first public remarks on the rapid rise of DeepSeek. He described the DeepSeek-R1 inference model as "impressive" and said that the previous market sell-off of NVIDIA stock was based on a misunderstanding.
Huang explained that the market's sharp reaction was driven by investors misinterpreting the situation. While R1's development appears to reduce dependency on computational power, the AI industry still requires immense computing resources to support post-training processing methods. These methods enable AI models to perform inference or prediction after training.
From an investor's perspective, they think of the world as being divided into two stages: pre-training and inference, where inference is simply asking AI a question and getting an immediate response. I don't know where this misunderstanding came from, but it's clearly incorrect, Huang said. He stressed that while pre-training remains critical, post-processing is the "most intelligent part" and "the key step in teaching AI models to solve problems." Huang added that improving AI inference capabilities represents the next frontier for the tech industry, which will still depend on massive computational power.
Related News:
Hong Kong should proactively embrace AI while ensuring safety precautions
Comment