Abstract

The efforts to evolve AI have focused on gigantification for improved performance, innovation of basic technologies like learning methods, and solution of its root issues, all to aim for artificial intelligence for a wider range of purposes.

In natural language processing, an increase in the number of parameters, a component of AI, has been proven to improve AI performance. GPT-3, the latest AI, has 175 billion parameters, and its writing ability has improved to an extent where test subjects fail to identify AI-generated passages. Additionally, AI can perform a great many passage-generating tasks like translation and summarization without new learning for each task, with its results already in wide use. The development race for gigantic AI is expected to become intense in the future, while the monopolization of the development knowhow by just a few companies seems to be on the rise.

In the field of AI imaging, advancement in learning methods has been changing the way AI is used. In the past in AI learning for images, AI needed instructor-provided data supplied by humans for supervised learning, The new self-supervised learning, however, does not need instructor data; AI looks for its own clues from a given image. Self-supervised learning will likely solve the AI utilization issue of having to manually prepare huge amounts of instructor data.

Efforts to solve one of AI's root problems, not being able to perform outside the data it has learned, are beginning to produce results. Approaches like trial and error, performing tasks outside the data it has learned, and inferring causal relationships, have been adopted toward the solution of this problem in an effort to create a more human-like AI that possesses general-purpose intelligence. AI will probably keep evolving in the future, creating innovations that become the basis for all technologies.