Releases·Model Release
Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI
During Jev launch week, Latent Space published an interview with Diogo Almeida, CEO of TypeSafe AI and InstructGPT co-author, on why he built a 'System One' model for production software rather than chat. Jev's launch video drew ~40M views, and its core training technique, Reinforcement Learning for Calibrated Decisions (RLCD), optimizes for calibrated probabilities instead of human preference feedback. Practitioners get official patterns, cookbooks, and a coding-agents guide, plus use cases spanning computer use, games, linting, and analytics.
0 engagement1 sourceMon, Sep 21, 2026, 10:13 PM
AI Overview
- Diogo Almeida, InstructGPT coauthor and TypeSafe AI CEO, says Jev's launch video hit ~40M views vs GPT-4o's 22M and 6 Astra's 137M
- Jev's core innovation is Reinforcement Learning for Calibrated Decisions (RLCD), optimizing epistemically honest probabilities on System One tasks instead of RLHF
- TypeSafe rejects public benchmarks like JevBench and optimizes for intelligence per dollar; Diogo calls it a data lab rather than a model lab
- TypeSafe documents Jev patterns and cookbooks for computer use, games, virtual try-ons, analytics, linting, and coding agents
- Almeida traces RLHF's three branches to Christiano 2017, Stiennon 2020, and Ouyang 2022, calling Function Calling through Reasoning 'hacks' on sequence-to-sequence prediction
- TypeSafe teases a future ReasoningJev model and Almeida's contrarian view on building a neolab without billions in funding
Related
Releases·Model Releasefeatured
TypeSafe AI launches Jev, an RLCD-trained decision model claiming 20-200x faster, 40-400x cheaper inference than LLMs
TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida, released Jev on September 15, 2026 after two years in stealth — a frontier model that takes unstructured state plus typed questions and returns parallel, calibrated decisions instead of generated text. It is trained with a new objective, Reinforcement Learning for Calibrated Decisions (RLCD), positioned against RLHF and RLVR. TypeSafe claims roughly two orders of magnitude faster and cheaper inference than LLMs on decision tasks, and community reaction has been intense: follow-up analyses, a JevBench benchmark, and open-weight imitators like Kev appeared within days.
TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida, released Jev on September 15, 2026 after two years in stealth — a frontier model that takes unstructured state plus typed questions and returns parallel, calibrated decisions instead of generated text. It is trained with a new objective, Reinforcement Learning for Calibrated Decisions (RLCD), positioned against RLHF and RLVR. TypeSafe claims roughly two orders of magnitude faster and cheaper inference than LLMs on decision tasks, and community reaction has been intense: follow-up analyses, a JevBench benchmark, and open-weight imitators like Kev appeared within days.
41001.7k engagement17 sourcesreddit, rss, youtube +1Tue, Sep 15, 2026, 12:00 AM
Products·launched
SemanticSpace: map prompts onto a 2D semantic plane with the Jev model
SemanticSpace is a web experiment that visualizes prompts in a 2D semantic space. It uses TypeSafe AI's Jev model to position prompts on a Cartesian plane whose axes are defined by arbitrary user-editable phrases. It lets users explore any 2D conceptual relationship visually.
SemanticSpace is a web experiment that visualizes prompts in a 2D semantic space. It uses TypeSafe AI's Jev model to position prompts on a Cartesian plane whose axes are defined by arbitrary user-editable phrases. It lets users explore any 2D conceptual relationship visually.
1 engagement1 sourcehackernewsFri, Sep 18, 2026, 12:49 PM
Releases·Model Releasefeatured
Ex-OpenAI researcher Diogo Almeida builds Jev, a 'System 1' AI model that skips language
A Fireship video covers Jev, a 'System 1' AI model built in stealth over two years by ex-OpenAI researcher Diogo Almeida. The model cannot talk or write code but is claimed to run 200x faster and 400x cheaper than LLMs while being hallucination-free. Its launch challenges the assumption that language is necessary for AI reasoning.
A Fireship video covers Jev, a 'System 1' AI model built in stealth over two years by ex-OpenAI researcher Diogo Almeida. The model cannot talk or write code but is claimed to run 200x faster and 400x cheaper than LLMs while being hallucination-free. Its launch challenges the assumption that language is necessary for AI reasoning.
24.5k engagement1 sourceyoutubeMon, Sep 21, 2026, 05:52 PM
Social·Community
TypeSafe AI Launches Jev Decision Model, Drawing Developer Interest
Medium writers and Reddit users are discussing Jev, a decision model from TypeSafe AI introduced in September 2026. Writers describe it as answering with probabilities rather than words, choosing among provided options for small agent decisions, and potentially making decision intelligence practical via a cheaper, faster enterprise stack. Reddit users report running local Jev-like models for deep eval testing and ask peers for practical use cases, with one calling it a 'solution in search of a problem'.
Medium writers and Reddit users are discussing Jev, a decision model from TypeSafe AI introduced in September 2026. Writers describe it as answering with probabilities rather than words, choosing among provided options for small agent decisions, and potentially making decision intelligence practical via a cheaper, faster enterprise stack. Reddit users report running local Jev-like models for deep eval testing and ask peers for practical use cases, with one calling it a 'solution in search of a problem'.
060205 engagement6 sourcesrss, redditSat, Oct 10, 2026, 01:26 AM
Products·launched
Odyssey: browser-based Earth-to-lunar-orbit flight sim with LLM-chosen maneuvers
Odyssey is a browser-based cinematic simulation of a journey from Earth to lunar orbit. It uses Jev from TypeSafe AI to select bounded navigation actions, showing each decision and its option probabilities live, with a deterministic fallback when the model is unavailable. It is a demo/educational experience for anyone interested, no signup required.
Odyssey is a browser-based cinematic simulation of a journey from Earth to lunar orbit. It uses Jev from TypeSafe AI to select bounded navigation actions, showing each decision and its option probabilities live, with a deterministic fallback when the model is unavailable. It is a demo/educational experience for anyone interested, no signup required.
1 engagement1 sourcehackernewsSat, Sep 19, 2026, 08:38 PM