OpenAI 's neural network language model " GPT-3 (Generative Pre-Training-3, hereinafter referred to as GPT-3) ", which trained a large-scale text data corpus of about 45TB using about 175 billion parameters, will soon be available . If so, the new version " GPT-4 " will be released by December this year, or February 2023 at the latest .


Towards Data Science: GPT-4 Is Coming Soon. Here's What We Know About It
The release of " GPT-3 " in May 2020 shocked the world with the ability of AI to create sentences that look natural to humans. Regarding its official successor, OpenAI CEO Sam Altman announced in April this year that " GPT-4" would be released in the next few months. GPT-4 is purported to change the way we interact with artificial intelligence by allowing AIs to have “more human-like” natural conversations, significantly reducing training costs for related applications.

Compared to GPT-2, which appeared in 2019, the current GPT-3 has more than 175 billion parameters, and is about 100 times larger than GPT-1, which was proposed in 2018, with 117 million parameters. , is found to increase significantly. Therefore, in the upcoming GPT-4, more parameters will be set not only for more natural interactions, but also for Turing tests to convince the user that the object of the interaction is really "human". Many people think that


There have been many AI systems that have claimed to have passed the Turing test, but they have mostly passed under certain circumstances. -4 is considered the first.

On the other hand, it is said that the threshold for corporate recruitment of GPT-4 in the future will be lower than the existing GPT-3. GPT-4 averages less than $1 million in training costs compared to GPT-3, which can cost hundreds of thousands of dollars at each stage to build a learning model, which can be huge depending on the training goal. It is said that it will be possible for more companies to build large-scale AI systems and utilize them in more fields