Anthony K. Brandt

"Both large language models (LLMs) and the human brain develop internal models of reality to make accurate predictions. Both typically prefer choices with the strongest track records. [PDF]

"However, when faced with a creative challenge, LLMs remain committed to high-probability options while humans can opt for unproven ones. This paper delves into one way of making unlikely events plausible —amplifying the anomaly.

"The concept involves extrapolating viable consequences from an unlikely proposition. Rather than being treated as oddball events or one-offs, the anomaly permeates the creative work. 

"Notably, novelty and appropriateness can be in tension with each other, with high utility coming at the cost of low novelty. 

"Amplifying the anomaly aligns these competing demands. It enhances originality: the rarer a proposition and the more thoroughly it is worked out, the more unique and surprising the result. At the same time, the effectiveness and value of the unproven option also rises: thorough elaboration in the creative product establishes its fitness. 

"Musical examples by Beethoven, Schubert, and contemporary composer Sky Macklay, along with creative products from other domains, illustrate this principle. Classic LLMs have several

limitations that make it difficult to amplify the anomaly: they are 

• Steered toward norm-driven outcomes, 
• Make short-term decisions, and are 
• Not designed to self-evaluate. 

"As a result, they have difficulty of developing unusual propositions in a non-obvious way without human guidance."