The A.I. Daily
Jev
TypeSafe AI's frontier model returning typed probabilistic decisions for software automation.
Jev
40 storiesJev is a lightweight decision model from TypeSafe AI, built with input from a ChatGPT co-inventor and released in limited form on September 15, 2026; it constrains outputs to run directly and is claimed not to hallucinate. Recent coverage includes Vercel data showing 35% of paid teams adopted it within two weeks, a roughly $870 million Andreessen Horowitz-led round at a $7.5 billion valuation, Cloudflare's Clef and Clairic competing models, open releases such as Decision 3.0 and Vega, and a study finding embeddings plus simple classifiers beat Jev when labels are available.
jevals: replacing LLM judges with typed Jev decisions for agent evals
jevals is an open-source tool that replaces LLM judges with typed Jev decision models for agent evals and guardrails. All evals for a trace are sent as one HTTP request costing a few thousandths of a cent and returning in a few hundred milliseconds, suitable for every trace or inside the agent loop. It supports Jev, Kev, Laya, and regular chat LLMs.
Four AI Models Tested on Doom Control Benchmark
An experiment let Jev, Laya, finetuned ModernCE and finetuned Qwen3.5-4B play ViZDoom autonomously, comparing kills, survival time and latency. Finetuned Qwen3.5-4B averaged 3.63 kills, the best, while Jev survived 5.63s on average with p95 latency near 200ms. No model received Doom-specific training.
ProgramAsWeights Compiles English into Local Neural Programs
Researchers at the University of Waterloo released ProgramAsWeights, an open-source project that compiles English function descriptions into reusable neural programs that run locally. A finetuned Qwen3-4B generates LoRA adapters for a frozen Qwen3-0.6B interpreter, separating compilation from inference so programs can be saved, distributed and composed with ordinary code.
Open-source 395M 'System One' model Von released
A developer released Von, an open-source 395M parameter 'System One' model that serves as a drop-in replacement for TypeSafe's JEV. It runs entirely on CPU with 1–2 GB memory, responds in 25–300 ms, and beats JEV on all benchmarks. The developer is open to job offers from AI research labs.
CUA-S1 released: A System One model for computer use
The Cua team released CUA-S1-FORMS, a narrow decision model for form interactions with only 706k parameters and a 2.8 MB checkpoint, trained in under 30 minutes. Inspired by TypeSafe's JEV and Kahneman's dual-system theory, it predicts whether to use, check, click, or skip given elements, without generating new values or considering screenshots.
Developer Says His Non-Autoregressive Decision Model Was Rebranded a Breakthrough
A developer says he released an RL-based non-autoregressive decision model, Laya, back in March 2025 via an arXiv paper, Hugging Face weights, and an open dataset—sub-35ms inference across 100+ languages. He criticizes frontier lab TypeSafe AI, founded by a former OpenAI member, for launching a similar product, Jev, in September 2026 as a brand-new breakthrough without papers, open weights, or training data.
Community Debates Whether Jev Is Just a Generalized BERT
Reddit users discussed the newly launched Jev, seeing it as an intelligent classifier with custom prompt and criteria instructions, and questioned how it differs from fine-tuned BERT or whether it is just a generalized BERT. No official technical details were provided.
Reverse-engineered Jev-like model released
An open-source repo releases a Jev-like model that scores multiple text options in one pass, with Doom and chess examples. Reported results: Doom checkpoint averaged 0.60 kills and -97.50 reward over 10 episodes; chess checkpoint scored 4 wins, 46 draws, 0 losses vs a random mover but 0 wins, 2 draws, 48 losses vs Stockfish level 0.
TypeSafe launches Jev, a faster cheaper frontier model
TypeSafe announced its System One models and Jev, claiming Jev is 20–200× faster and 40–400× cheaper on structured decision tasks, using a new architecture, parallel sampler, and RLCD training; Jev is available in early access.
That is everything