Can AI learn to be creative? New research explores the possibility

Two innovative studies are harnessing large language models (LLMs) to guide AI towards novel and useful ideas. For AI to understand the world broadly and navigate complex tasks, it needs to explore and prioritise new concepts autonomously. 

Usage of GPT-4 to select “promising” states during AI’s learning process, a significant advancement from previous hand-coded rules. This system was tested on tasks involving text processing and multistep solutions, such as solving arithmetic problems or navigating grid worlds. IGE outperformed other methods, demonstrating its potential to aid in discoveries like new drugs or materials.

OMNI-EPIC builds on the previous OMNI system by generating and evaluating new tasks within a physics simulator. It creates a diverse array of tasks, leading to a more comprehensive learning process for AI agents. The system has generated over 200 unique tasks, showcasing its capability to innovate beyond its initial programming.

While these developments are promising, they also raise concerns. The use of LLMs, trained on vast amounts of human data, introduces limitations and risks. Experts warn that open-ended AI could potentially diverge from human values, posing safety challenges. However, proponents argue that open-ended learning could balance power among AI agents, enhancing overall safety.

As AI research advances, the balance between innovation and safety remains crucial. The current focus is on refining these systems and understanding their broader implications, with the ultimate goal of creating AI that can explore and learn in ways that closely mirror human creativity.