How much does it cost to develop a ChatGPT applet APP plugin or other software system?GPT4/GPT3.5
After the first half of 2023 Wenxin Yiyi, Li Kaifu sent a circle of friends to resolve to do China’s chatgpt, called project AI 2.0.
Today we’re going to figure out how much it costs to make a big model of a language like this.
The first is to form a team, algorithm engineers, senior algorithm engineers, the higher the better. Each team member’s monthly salary is at least 30,000, senior algorithm engineers even more, a team needs at least 10 people.
The second is the graphics card.
First of all, the training of AI need to use Nvidia’s highest-end GPUA100, a piece of 80,000 yuan, gpt3.5 with 30,000 pieces of the A100. we still remember that year to beat Li Shishi’s alphago? It used 176 GPUs, that GPU is still not as good as A100, so a rough estimate of at least 2.4 billion yuan. In fact, due to the AI model parameters are always increasing, it consumes an exponential amount of arithmetic power, 10 billion investment is very conservative.
The second is power consumption. chatgpt’s power consumption for one training session is more than 900,000 dollars, which is equivalent to the power consumption of 1,200 Chinese people for one year, just for one training session. Currently ChatGPT’s electricity bill is 50,000 dollars per day.
At the end of the day chatgpt is very wasteful, and it is not the researchers who are wasteful, it is not the programmers, it is the ordinary people. for AI to grow it has to have people to label the data. For example, to train a model to recognize cats, then you have to input all such photos and so on, and tell everyone that this thing is a cat, and then when he sees such a picture, he will tell you that this thing is a cat, and such manual labeling has to be millions of pictures at the very least, and even millions of pictures are not new. So chatgpt is taking all the internet information on this planet and manually labeling it, especially the harmful content of the internet, crime, violence, anti-humanity and so on and so forth. Now it is mainly Kenyans who help chatgpt to label harmful information, and it is very harmful to one’s mental health to have to label over 200+ paragraphs a day. So it is said that every progress in the West contains the blood and tears of the African people, including this chatgpt, according to estimates, as of 2022, more than 5 million people in the world’s data taggers. So to train a tool like chatgpt, you need a million data taggers at the very least.
In summary, AI language big model is not suitable for us SME entrepreneurs. If you really want to enter the artificial intelligence AI this track, the most suitable for the lowest economic cost at present is to access the API interface of GPT.