The model trying to kill chat

Diogo Almeida, a co-author of the original InstructGPT paper, has launched Jev from TypeSafe with a single target: reliability on "System One" tasks, the fast, intuitive decisions where current models routinely fail. The launch video has racked up roughly 40 million views — ahead of GPT4o at 22 million and Fable 5 at 15 million, though still well behind Navier Stokes at 74 million and 6 Astra at 137 million. Almeida argues that every innovation since InstructGPT — Function Calling, Structured Outputs, Reasoning — is merely a patch on top of the text-sequence prediction paradigm, and that paradigm has run its course.

The core novelty is RLCD, Reinforcement Learning for Calibrated Decisions, a technique not yet officially published. Where RLHF optimizes against human feedback and produces hallucinations, sycophancy, and permanent dependence on human raters, and where RLVR optimizes against programmatically verifiable outputs with rubrics — solving Navier-Stokes but sharpening "jagged intelligence" and integrating poorly with other software — RLCD targets answers with "epistemically straight" probabilities on System One tasks. Almeida draws a line through three RLHF branches: Christiano et al. 2017 (the robot backflip demo), Stiennon et al. 2020 (summarization learning), and Ouyang et al. 2022 (InstructGPT). All three, he says, were the wrong north star.

The central thesis: a generation of training helpful human assistants damaged their ability to serve as a training mode for models. Almeida spent 2023-2024 trying to train a model that would solve the reliability problem at the heart of software, and failed due to personal and organizational underestimation. Jev is built to "disappear into the background," to be as unnoticeable as regex, not to hold a conversation. The use-case list is long: speed-dependent gaming and computer-use, voice-plus-computer-control, a Doom demo, driving in games, Excalidraw, virtual measurements, smart NPCs, guided responses in text messages, formal code agents, linting, tool-call compression, programming languages built on top of Jev, analytics and user-journey reconstruction, "dark data," and "smart software." There is also an official guide for using Jev as a judge.

Almeida publicly rejected building a generic JevBench, calling it low-effort work that misses the point. Theo raised a credible critique of "KV Cache tyranny," and Almeida published a detailed response arguing that "Cache Rules Everything" — a critical concern for code agents. The company hints at future models such as ReasoningJev, but has released no benchmarks, dates, or technical details beyond the RLCD principle. The full AIE talk video is available, and TypeSafe publishes official patterns and cookbooks for developers who want to build on Jev instead of replicating the API for the 55th time.