Millions of dollars spent annually on AI training in world - analysts
According to Stanford University's Artificial Intelligence Index Report for 2024, the United States leads in the number of known machine learning models - 61 units,El.kz reports with reference to Ranking.
A machine learning model is an application of artificial intelligence (AI) that can identify patterns and generate predictions or content. The main part of the learning process occurs due to the use of computing power and without the participation of a programmer, so it seriously helps to save time. Machine learning can be used in any large business today.
After the United States, China is 15 in terms of the number of known machine learning models, and France also has a significant number with 8 models.
In order not to start developing AI from scratch, basic models are used as a starting point for creating machine learning models. They are characterized by huge amounts of data for training. Their development in recent years has reached a high pace, and their use has good prospects.
For the period from 2019 to 2023, Google leads in the number of basic AI models created, accounting for 40 models, followed by OpenAI (20 units) and Meta (19 units). A significant number are also accounted for by Microsoft (18 units), DeepMind (15 units) and Tsinghua University from China (7 units).
Meanwhile, the costs and expenses of AI training are growing exponentially, forcing universities that have traditionally been centers of AI research to stop developing their own advanced basic models.
Epoch AI, an AI research institute, conducted an analysis of the cost of cutting-edge models. The duration of the training was analyzed, as well as the type, quantity and utilization rate of the training equipment, taking into account information from publications, press releases or technical reports related to the models in the period from 2017 to 2023.
The estimated cost of training Google's basic Transformer model in 2017 cost only $930. (hereinafter - taking into account inflation). The following year, training the BERT-Large model from the same Google cost 3.3 thousand dollars. More than three times as many, but still relatively few.
But now the training of the GPT-3 model from OpenAI, created in 2020, accounted for $4.3 million, and the next GPT-4 model, created in 2023, is worth $78.4 million. Google, in turn, has already spent an estimated $191.4 million to train its AI model, Gemini Ultra.
Despite the ever-increasing cost of training basic models, such an investment is considered profitable because the models can be used for a very long time. Based on them, it is faster and cheaper to develop new machine learning models, and accordingly to develop and create more and more advanced AI.
Subscribe to our Telegram channel and be the first to know the news!