TypeSafe AI CEO Diogo Almeida discusses Jev’s design for reliable, composable AI systems on Latent Space

Diogo Almeida, CEO of TypeSafe AI, appeared on the Latent Space podcast to discuss Jev, a system designed for reliable, composable AI workflows rather than human-preference-aligned chat interfaces. Jev’s core innovation is a novel training technique called Reinforcement Learning for Calibrated Decisions (RLCD), which optimizes for "answers with epistemically honest probabilities on System One tasks" instead of human-rated feedback (RLHF) or programmatically verifiable outputs with rubrics (RLVR).

Almeida, a coauthor of the InstructGPT paper, argued that the industry’s focus on human-preference alignment has led to issues like hallucinations, sycophancy, and poor reliability for automation tasks. He contends that RLHF and RLVR are mismatched with the needs of composable, programmable AI systems, which require robustness and determinism rather than human-like responses.

TypeSafe explicitly rejects public benchmarks, stating that optimizing for standardized evaluations distracts from the company’s north star: reliability and intelligence per dollar. Almeida emphasized that the right task and data can matter more than compute, citing his essay *The Bitterest Lesson* to argue that developers should prioritize measurable, decomposable workflows over monolithic model performance.

Jev is designed to "disappear into the background" of software, functioning as unremarkable as regex. Its programming primitives emphasize structured state, small decisions, and shared state management—contrasting with the giant prompts and system messages typical of chat-first LLMs. Use cases highlighted include computer use, dark data analysis, real-time intelligence, and coding agents, with examples like driving cars in games, virtual try-ons, and smart NPCs.

Almeida also addressed the limitations of reasoning models and the KV cache, advocating for a shift toward multi-agent architectures and sub-agents to overcome the constraints of autoregressive chat paradigms. He suggested that many neo-labs are approaching AI incorrectly by overemphasizing scaling laws and pre-training costs, instead advocating for task-specific optimization and data-centric development.