TypeSafe CEO Diogo Almeida Podcast: Jev Model, System1 Architect | TypeSafe CEO Diogo Almeida 播客访谈：Jev模型、System‑One架构

TypeSafe CEO Diogo Almeida recently appeared in a podcast discussing the Jev model, System-1 architecture, and the rise of programmable AI. Almeida shared his personal state after Jev's release, noting the extreme exhaustion and exhaustion both physically and mentally. However, he felt a sense of fulfillment, as developers finally understood the value of Jev, which he believes could lead to an economic revolution based on AI.

He emphasized that his priority is to focus on the developer community over VIP investors, running online town hall meetings to engage with developers. Almeida highlighted the industry paradox where AI can solve complex mathematical problems but cannot automate many economic tasks. He criticized reinforcement learning from human feedback (RLHF) for causing pattern collapse, which leads to inaccurate model outputs.

He also disagreed with Yann LeCun's view on JEPA (Joint Embedding Pretraining), stating that while JEPA is a cool early research, its practical utility is doubtful.

Almeida acknowledged the flaws of RLHF, including mode collapse, and its impact on model calibration. He warned against embedding value judgments into the model's technology layer and instead suggested separating capability alignment (ensuring models can complete tasks as requested) from safety alignment (ensuring models follow third-party value constraints). He argued that forcing AI to follow specific value constraints at the technical level can harm model intelligence.

Regarding benchmarking, Almeida strongly opposed public benchmark testing, believing private proxy evaluations were more accurate. He proposed evaluating models within real-world workflows to build trust. He also emphasized that data is the core of model capability, and companies should focus on acquiring data to improve reliability metrics. Jev's API featured three new primitives: Choice, Score, and Bernoulli (nule), designed to work with structured JSON objects rather than strings.

Almeida discussed the importance of reliability, determinism, and versioning strategies, stating that models should be versioned and have long-term support plans. He also addressed concerns about AI over-promising and the potential AI winter, asserting that TypeSafe's route of zero RLVR is the optimal choice. The CEO criticized the notion that AI should be slowed down and stated that TypeSafe is focused on creating reliable, developer-friendly infrastructure rather than becoming a new AI startup.