Problems of Artificial Intelligence Enterprises

1. Artificial intelligence products need to train models by neural network calculation, andthe data model training process needs to consume a large amount of computingresources. Also, Artificial intelligence products want to achieve better product index, inaddition to the algorithm. That is, there is a need for massive data to train, but moredata, in the case of equal computing resources, means longer training, say over a week or even a month to several months. If there are incorrect parameters in the trainingprocess, repeated training is needed. Long training time is extremely disadvantageous totheenterprise product's iterative updating, increasing the product's likelihood to fail intheindustry's competition. This leads to the fact that many manufacturers have to invest alot of money to purchase GPU, FPGA, and other hardware resources, directly causingthe artificial intelligence chip provider's, e.g. NVIDIA's, share price to rise rapidly. For most small and medium enterprises, more than one million of capital investment isahuge burden.

2. AI products still need to be decoded by a neural network after launching. The larger thenumber of users, the greater the amount of calculation required, hence pushing upthecost. Consequently, the user access frequency in different time periods will alsochange, and one-time purchase of a large number of computing resources will inevitably result inidle resources.

3. The three elements of artificial intelligence are computing power, algorithm, anddata. The amount of data is an important factor affecting the index of any artificial intelligenceproduct. Companies that design artificial intelligence products need to continually annotate low-quality data or directly purchase high-quality data, but most data usageinvolves the issue of user privacy, and data providers can only hope that the datawon't be duplicated. They just sell the right to use the data, but not its ownership, whichisalmost impossible to do.