These sources explore innovative strategies for enhancing multimodal AI performance by repurposing existing technologies and optimizing instructions without intensive retraining. One paper introduces Scrapyard AI, a framework that treats obsolete AI models as a frugal, high-utility resource for researchers facing compute constraints. This concept is applied through Project Nudge-x, which utilizes these legacy systems and satellite data to interpret the environmental consequences of global mining operations. A second paper investigates evolutionary prompt optimization, a method that uses survival-of-the-fittest algorithms to discover advanced reasoning strategies in vision-language models. Through this iterative process, AI models independently learn to utilize external tools, such as Python scripts, to decompose and solve complex visual tasks more accurately. Together, these works highlight a shift toward computational parsimony and sophisticated inference-time adaptations to achieve state-of-the-art results. This research collectively suggests that the future of artificial intelligence lies in the creative reconfiguration of existing assets and the refinement of human-machine
