Eric from Modal explains the engineering behind Modal Notebooks, a cloud-based Jupyter notebook platform designed for rapid GPU access and real-time collaboration. The initiative aimed to create a seamless experience that retains the speed of local notebooks while operating in a multi-tenant cloud environment. Central to this is the concept of Modal Sandboxes, which are secure, isolated processes that manage the execution of code through a kernel protocol. This reimagined architecture allows multiple users to interact with the notebook simultaneously while maintaining low latency in code execution and output display. A significant innovation is the lazy-loading container filesystem that reduces startup time for containers. Instead of downloading complete images, the system fetches only the necessary files on demand, using a tiered caching approach that keeps most accesses fast. Modal's infrastructure also automatically manages resources, pausing idle kernels to save costs while allowing users to resume work instantly. The storage solution, VolumeFS, provides a global, mutable, and fast filesystem for managing extensive AI workloads, ensuring that files feel local even when accessed from different regions. For collaboration, Modal Notebooks utilize a library called Rushlight, which handles real-time editing through operational transformation. Changes made by users are processed via Redis Streams, allowing for multiple users to edit simultaneously without conflicts. The system separates editing from execution, ensuring that users can run code and receive outputs while maintaining a consistent and current state for all collaborators.