How nicely can AI fashions clear up (and create) rebus puzzles?

What does it imply for an AI to be artistic?

Final 12 months, I wrote an article about measuring creativity in Massive Language Fashions (LLMs) utilizing a number of word-based creativity assessments.

Since then, AI has developed quickly and is able to processing and creating each textual content and picture. These fashions, typically known as “Multimodal Massive Language Fashions” (MLLMs), are extraordinarily highly effective and have superior skills to grasp advanced textual and visible inputs.

On this article, I discover one technique to measure creativity in two common MLLMs: OpenAI’s GPT-4 Imaginative and prescient and Google’s Gemini Professional Imaginative and prescient. I exploit rebus puzzles, that are phrase puzzles that require combining each visible and language cues to unravel.

Creativity is extraordinarily multi-faceted and troublesome to outline as a single trait. Subsequently, on this article, I intention to not measure creativity typically, however to guage one very particular facet of creativity.

Word [modified from my earlier article]: These experiments intention to not measure how artistic AI fashions are, however relatively to measure the extent of artistic course of current of their mannequin generations. I’m not claiming that AI fashions possess artistic pondering in the identical method people do. Moderately, I intention to point out how the fashions reply to explicit measures of artistic processes.

A rebus puzzle is an image illustration of frequent phrases or phrases. They usually contain a mixture of visible and spatial cues. For instance, under are six examples of rebus puzzles (solutions are on the finish of the article).