Large language models (LLMs) are rapidly evolving, but adding visual capabilities often leads to them forgetting what they already knew about text. This "text-only forgetting" poses a significant challenge for building truly versatile AI. Imagine asking an AI assistant a series of questions, starting with text-based queries and then adding an image for more context. Ideally, the AI should seamlessly handle both, drawing on its textual knowledge even when processing images. However, many multimodal LLMs (MLLMs) struggle with this, performing worse on text tasks after being trained on image-text pairs. Researchers have discovered that this forgetting is related to how the MLLM’s “attention” shifts. When presented with an image, the model tends to overly focus on the visual information, neglecting the surrounding text. To address this, a novel approach called WINGS has been developed. WINGS acts like a balancing mechanism for the model’s attention. It incorporates two sets of "learners"—visual learners that focus on images and textual learners that focus on text. These learners work in tandem to compensate for the attention shift caused by visual input. Initial tests show promising results. WINGS outperforms existing MLLMs in both text-only and visual question-answering tasks, especially on complex, interleaved conversations where text and image inputs are mixed. The research suggests that the secret to building genuinely intelligent MLLMs lies in maintaining a balance between learning new modalities and retaining proficiency in existing ones. This work opens exciting new avenues for developing AI assistants capable of seamlessly blending text and images in more human-like conversations.