TypeSafe 首席执行官详解 Jev：面向软件的 System 1 模型是什么？为什么可靠性比聊天基准更重要？

Summary

In this Latent Space interview, TypeSafe AI CEO Diogo Almeida explains why the company built Jev as a new class of “System 1” or large programmable model rather than another chat-optimized language model. He says Jev is designed for code to consume its outputs, with the company optimizing for intelligence per dollar and aiming to make intelligence a dependable software dependency. Its central training idea is Reinforcement Learning for Calibrated Decisions (RLCD), an unpublished approach intended to produce epistemically honest probabilities on System 1 tasks instead of optimizing primarily for human preference through RLHF or programmatically verifiable answers through RLVR. Almeida argues that RLHF can encourage mode collapse, sycophancy, hallucination, refusals, and overconfidence, while RLVR can produce highly capable but jagged behavior when optimized around narrow verifiable tasks. He presents calibration, robustness, and reliability as more important than strict determinism: similar inputs should produce similar outputs, while exact repeatability is only valuable for some workflows and can reduce intelligence per dollar. Jev’s API exposes programming-oriented primitives such as choices, scores, and Noulli probabilities, which map to switches, thresholds, sorting, and conditional logic. Almeida recommends passing structured state rather than large system prompts and decomposing workflows into small, measurable semantic decisions that can be tested and escalated by confidence thresholds. The interview also covers model versioning, the possibility of limited-term support for popular versions, internal rather than public evaluation, and the company’s reluctance to train on user data because it could overfit present-day usage. Early use cases discussed include coding agents, dark-data analysis, real-time intelligence, computer control, games, smart software, and verification of other model calls. Almeida says TypeSafe plans to explore additional model shapes, modalities, and smaller or larger points on the intelligence-cost frontier, while arguing that future coding agents should manage explicit state, sub-tasks, memory, and multi-agent coordination instead of being constrained by a single growing KV cache. He also describes the path from his work on InstructGPT to TypeSafe and says the company views itself more as a data and task-design lab than as a conventional model lab.