Imagine if you could see not with your eyes, but with the very structure of a language model. That's the intriguing idea behind a new research paper, "Visual Perception by Large Language Model's Weights." Traditionally, AI models "see" by processing images into a format understandable by language models, a process that's computationally expensive. This new research proposes a radical shift: representing visual information directly as the model's weights. Instead of feeding images as separate inputs, researchers convert visual features into "perceptual weights" that merge directly with the language model's existing structure. This allows the model to perceive visual information without the usual computational burden of processing separate image data. The results are promising. This new method, called VLoRA, performs comparably to state-of-the-art models on various benchmarks while significantly reducing computational costs. This breakthrough could democratize access to powerful multimodal AI models, making them more efficient and accessible for a wider range of applications. While challenges remain, such as optimizing the visual feature extraction process, this research opens exciting new avenues for how we integrate visual perception into the core of language models, potentially leading to more efficient and capable AI systems.