Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI

TypeSafe AI CEO Diogo Almeida appeared on the Latent Space podcast to discuss Jev, his company's newly launched model. A launch video for Jev has been viewed roughly 40 million times, according to the podcast. Almeida coauthored the InstructGPT paper.

Almeida described Jev as the first of a class he calls "System 1" or "large programmable" models, designed for code rather than chat as the consumer. He said Jev is optimized for intelligence per dollar, and is the first model to carry that name. TypeSafe rejects public benchmarks, per the podcast.

Jev's core innovation is what Almeida calls Reinforcement Learning for Calibrated Decisions. He described it as a novel, unpublished technique that optimizes for answers with "epistemically honest probabilities on System One tasks." Almeida contrasted RLCD with RLHF, which he said causes hallucinations, sycophancy and permanent reliance on humans, and with RLVR, which he said solves hard problems but exacerbates jagged intelligence and integrates poorly with other software. He traced RLHF to three papers: Christiano et al 2017, Stiennon et al 2020, and Ouyang et al 2022.

Almeida said that from 2023 to 2024 he struggled unsuccessfully to train a model addressing reliability, which he described as the core problem with making LLMs the heart of software. He also discussed his essay on the Bitterest Lesson, saying the most important part of ML is not ML itself, and that choosing the right task and data can matter more than compute. He said TypeSafe considers itself a data lab rather than a model lab.

Almeida also published a note titled the Tyranny of the KV Cache, recommended as a follow-up. Other topics in the podcast include Jev's use cases, the inverse SaaS-pocalypse, and why Almeida said he would not pre-train with $1 billion.