Diogo Almeida traces TypeSafe’s philosophy from his OpenAI work on InstructGPT to Jev, a model designed for code consumption and machine-native software. He argues that RLHF optimized AI for human preferences, while RLCD aims at calibrated, programmable decisions that can remove humans from the loop. Across the conversation, he explains why public benchmarks, giant prompts, refusals, and KV-cache-centered coding agents are poor foundations for dependable automation. Jev’s programming primitives, structured state, decomposition, and use of small measurable decisions are presented as ways to integrate intelligence into software. The discussion also covers dark data, computer use, coding agents, model economics, AI safety, TypeSafe’s data-first culture, and the future of shared-state multi-agent systems.