The conversation centers on the creation of Composer, a specialized foundation model for software engineering. Instead of pre-training from scratch, Cursor and Fireworks used a top-down approach, starting with a strong open-source base and applying mid-training and reinforcement learning (RL) to specialize it for coding. This strategy allowed them to get a useful model into users' hands quickly. The RL training process was particularly innovative, using an asynchronous pipeline across four global clusters to maximize GPU utilization. They leveraged off-peak production inference GPUs and developed a lossless compression algorithm to efficiently ship model updates between clusters. The training incorporated both offline simulated rollouts and online real-time RL with actual user data, using self-summarization for long-horizon tasks and verifiable rewards for evaluation. The key takeaway is that the most powerful RL environment is a company's own production system, properly isolated.