Jev: System One Models for Prod, Not God - with Diogo Almeida, CEO, TypeSafe AI

TypeSafe AI CEO Diogo Almeida released an AI model system called Jev. According to the article's account, the launch video drew about 40 million views; for comparison, the figures given in the article are 22 million for GPT4o, 15 million for Fable 5, 74 million for Navier Stokes, and 137 million for 6 Astra. These view-count figures all come from the article author's account, with no first-party data to corroborate them, and should be treated as source claims rather than confirmed facts.

Jev is positioned as a System One model for software rather than chat scenarios. One of its core goals is to "disappear into the background," that is, to be as unobtrusive as a regular expression. This positioning stands in sharp contrast to the current mainstream path of large models pursuing general chat and reasoning capabilities.

Jev's core innovation is "Reinforcement Learning from Calibrated Decisions" (RLCD), a technique not yet publicly published, whose optimization goal is to make System One tasks give "cognitively honest probabilistic answers." The article says RLCD differs from RLHF, which is based on human rating feedback—the latter leads to hallucinations, sycophancy, and permanent dependence on humans; it also differs from RLVR, which uses programmatically verifiable outputs and scoring criteria—the latter can solve Navier Stokes but exacerbates "jagged intelligence" and is hard to integrate with other software.

RLCD is described as a "technique not yet publicly published," and its specific methods and effects cannot be verified from the source. This means that although the article presents it as Jev's core innovation, readers can for now only treat it as a source claim rather than an independently verified technical breakthrough.

TypeSafe refuses public benchmarks, instead optimizing "intelligence per dollar." The article mentions that Diogo Almeida publicly rejected the idea of doing a general JevBench benchmark. This stance is rather rare in the current AI competitive environment, because public benchmarks are usually the main way to prove model capability.

The article also mentions that TypeSafe sees itself as a data lab rather than a model lab. This self-positioning echoes its stance of refusing public benchmarks: if the core advantage lies in data and task selection rather than model scale or general capability, then traditional benchmarks may indeed not be the most appropriate measure.

In the essay "The Bitterest Lesson," Diogo Almeida argues that "you get what you optimize for, and the bitterest lesson in ML is that the most important part of it isn't ML at all," meaning that choosing the right North Star metric matters more than compute. The article further explains that his view is: choosing the correct North Star metric—for example, whether to optimize for user preference votes or for being integrated into tool calls—makes everything else fall into place.

This claim is logically consistent with Diogo Almeida saying he would not spend $1 billion on pretraining: if the most important part isn't ML, then pouring huge sums into pretraining compute may not be the optimal path. The article mentions that he also hinted at his contrarian views on scaling laws, and used this to lead into how to build a modern new lab without the billions of dollars of a big lab.

Diogo Almeida was formerly a member of OpenAI's post-training team and a co-author of the InstructGPT paper (Ouyang et al 2022). The article says that for years he believed API-available frontier models were on the wrong path, from alignment to refusal to the reliability perspective, and that ChatGPT's enormous success led the industry to abandon all paradigms other than autoregressive chat fine-tuned LLMs.

Diogo Almeida says that between 2023 and 2024, due to underestimation at both the personal and organizational levels, he failed to successfully train a model that could solve what he considers the fundamental problem of LLMs as software—reliability. This narrative is a personal account, with no third-party corroboration, and should be treated as a source claim.

The Jev use cases listed in the article include speed-oriented scenarios such as gaming and computer use, voice plus browser control, a Doom demo, in-game driving, Excalidraw, virtual try-on, intelligent NPCs, text-message guided replies, coding agents, linting, tool-call compression, analytics replay and user journey review, dark data, entity resolution, natural language search, acting as a judge, and more.

The actual effects and maturity of these use cases are not given verifiable evidence in the source. Therefore, despite the broad list of use cases, readers cannot judge from the available information which are already mature and which are still at the demo stage.

The article mentions that TypeSafe's future models may include ReasoningJev, but it appears only in interrogative form, with no specific plan given. Therefore, whether ReasoningJev will actually be released appears in the source only in interrogative form, with no clear plan, and can only be treated as inference.

Diogo Almeida published a note on "the tyranny of the KV cache" as follow-up reading material for the topic of Jev and coding agents, because he believes "cache decides everything." The existence of this note indicates that TypeSafe's thinking about coding agents is not limited to the model itself, but also involves constraints of the underlying infrastructure.