Recently, Wang Jianshuo published an article titled "Don't take researching AI news as researching AI", the point of view of this article is very appropriate. Most of the time, I didn't do in-depth research, I just got a superficial understanding, which might be a little better than what he said, and I occasionally moved my hands. Many people just stand on the sidelines and watch, and don't really put their hands into practice. News is short-term, but skills are long-term. Spending time on relatively constant things can help us understand things that change every day.
I recently tried to install Stable Diffusion on a local server, but ran into various issues and had to give up after a long time. Even if there seems to be a very simple tutorial, the installation may fail due to the different system environment of each person. This is the case for every solution. If you don't practice it yourself, you can't know whether it is feasible. I have a friend who tried to build Tsinghua’s ChatGLM private server. He worked hard all day, stepped on a lot of pits and filled a lot of pits. He once wanted to give up in the middle, and finally successfully completed the deployment by following a video tutorial (almost no non-professionals) may be done).
Building some private servers of open source AI products requires not only technical strength, but also a certain amount of hardware investment. After all, general application servers cannot meet the demand and require super GPU computing power. Depending on the purpose, such as training, inference, computation, or image generation, the required cost varies. If you want to carry out secondary development on the basis of these open source products and package them into products for sale, the cost will be different. The application architecture is similar, but the architecture of the underlying AI capabilities and the application architecture are quite different.
