The post discusses predictions about the future of multimodal AI, particularly in the realm of image generation combined with language understanding. It argues for a shift from monolithic AI models to modular systems that utilize pre-trained components effectively. Key predictions include: 1. **Modularity Wins**: The future will favor systems built from the best pre-trained models rather than creating entirely new ones from scratch. 2. **Pre-trained Everything**: Leveraging existing state-of-the-art models and developing efficient connections between them will accelerate innovation. 3. **Understanding over Generation**: The quality of the control signal derived from large language models (LLMs) will be critical for successful image generation, meaning that understanding and reasoned instructions will primarily drive the creative process. The post emphasizes the need for better interfaces and training data while outlining the potential of this new modular approach in refining image generation capabilities.