The following are Zhou Hongyi's original words! You can take a look.
Computing power is the problem Centralizing the computing power of all AI companies in China may be one of the ways. First of all, our computing power is stuckThen the domestic Huawei will have a maximum of 500,000 this year, and Haiguang Information will have a maximum of 200,000. Cambrian is up to 200,000 sheets, plus the others are 1.1 million.
At the moment when SORA is popular, how do domestic large-scale model companies catch up with SORA, and what are the difficulties?
Zhou Hongyi, founder of 360 Group, believes that if SORA's technical route is open-sourced, China will be able to catch up quickly, but when catching up with SORA, computing power may become the thresholdPooling the computing power of all AI companies in China may be one of the ways. is a big threshold, because the computing power is seriously insufficient.
First of all, the consumption of computing power by the Dynamic Incentive Plan SORA analysis in February should be far more than that of the 100 billion model. Now the largest model of iFLYTEK is only 100 billion. Therefore, when domestic graphics cards are now stuck, computing power may be a problem. "At the moment,Meta already has 500,000 GPUs and may buy a million next yearMicrosoft should also place orders in the million-dollar range. All the AI companies in China together may have 500,000 GPUs, but they are all scattered among various companies.
The above Lao Zhou's analysis is correct from the data, and then I correct it here, the computing power is enough, but it is difficult to concentrate on one according to what Lao Zhou said, because 1 million cards are 10 million yuan, whether it is Huawei or Ali Tencent, they can't take out so much money to invest in a large model similar to sora, so it is unrealistic to catch up with sora, the fundamental reason is that we don't have so much money "no money, no money, this is very important". It is estimated that the enterprises in China that can come up with 100 billion yuan in cash are also monopoly traditional enterprises such as China Shenhua selling coal. Others don't have this strength. So it's unrealistic to catch up with SORA this year.
So should we chase it? Let's redo the algorithm, if the computing power can't afford it, let's make the algorithm again by ourselves. This is the only way for the algorithm to make up.
A window of AI technologyWell-known tech bloggers! Interpret global AI news. The future is the era of artificial intelligence to improve production efficiency, and in 2024, AI GC, intelligent driving, quantum computers, Internet of Everything, virtual reality and other applications will explode, and we are currently experiencing the fourth scientific and technological revolution.