Article
CommunityJev Guides & Articles
Jev as a System One Model: Latent Space Interview Notes
A Chinese-language summary of the Latent Space interview with TypeSafe CEO Diogo Almeida, framing Jev as a production System One model rather than a general God model: parallel sampling, 70–500 ms latency, and $0.042 per million input tokens.
- Format
- Article
- Reading time
- 2 min
- Published by
- Community
- Author
- 公众号_计算机魔术师
- Last verified
- Oct 2026
Highlights
- Diogo Almeida co-authored OpenAI's InstructGPT paper and helped develop the RLHF alignment approach before founding TypeSafe.
- TypeSafe reports 70–500 ms latency, free outputs, and $0.042 per million input tokens for Jev.
- A parallel sampler evaluates every candidate output in one request and returns schema values with calibrated probabilities.
- On invoice processing the article cites Jev at 61.8% against 79.1% for Sol, and says Jev trails the strongest Astra and Opus models.
- RLCD, Reinforcement Learning from Calibrated Decisions, trains for correct and well-calibrated decisions instead of human preference.
Watch out
A second-hand Chinese-language summary of the Latent Space interview, so figures such as 70–500 ms, $0.042 per MTok, and 61.8% versus 79.1% are reported rather than independently measured. The page notes TypeSafe has not published ECE, Brier scores, or reliability diagrams, and no licence or open weights are shown.
Questions about this entry
Short answers drawn from the linked project or page.
It summarizes a Latent Space interview with TypeSafe CEO Diogo Almeida, arguing that production systems need fast, structured System One decisions rather than full language generation, and presenting Jev as the model built for that role.
Not intuition in the cognitive-science sense, but an engineering class: fast, structured, predictable decision units. Jev returns typed judgments such as Choice, Score, Route, or Escalate with probabilities, while System 2 models reason token by token and output strings.
A parallel sampler evaluates all candidate outputs for one request at once and returns values inside the predefined schema plus calibrated probabilities, instead of autoregressive text. TypeSafe reports 70–500 ms end-to-end latency on this path.
RLCD stands for Reinforcement Learning from Calibrated Decisions. The article says RLHF optimizes for human satisfaction, which can make models overconfident, while RLCD optimizes whether decisions are correct and whether their confidence is statistically calibrated.
The interview reports Jev close to Sonnet and Terra on most decision tasks but weaker than the strongest Astra and Opus models, and on invoice processing it cites 61.8% for Jev against 79.1% for Sol.
Customer-service ticket classification, document routing, simple risk control, and Agent output validation, deployed at decision nodes, as a pre-filter before expensive models, or as an API-level replacement for structured-output LLM calls.
It notes that TypeSafe has not published full ECE, Brier scores, or reliability diagrams, so calibration remains a directional claim that the article itself flags with a question mark rather than a verified, independently checkable metric.
More Jev resources like this
Charlie Chen's five-minute guide to Jev: the hosted API opened to everyone on 21 September 2026, Choice, Score, and Noul in practice, a working Python request, and the jevmodel.org playground with free tokens.
Jev Guides & ArticlesArticle#community#article#tutorial
Community
czmilo's long guide to TypeSafe's System One model: the 193.6x/444.6x benchmark claims, RLCD training, Cloudflare, OpenRouter, and LangChain availability, a 32K context window, and vendor caveats.
Jev Guides & ArticlesArticle#community#article#guide
Community
accept99 explains how Jev differs from chat models and how to run Jev-like decision models for free: Ollama serves nimble 9B and tev1 4B and 0.8B over /v1/systemone, while the open-source Laya project offers pip-installable weights and an MLX build for Macs.
Jev Guides & ArticlesArticle#community#article#local
Community
A Google Cloud Community walkthrough by developer advocate Jeff Nelson that calls Jev from BigQuery through a private Cloud Run adapter, turning support-ticket text into owning-team, urgency, and blocking columns for 30 fictional tickets.
Jev Guides & ArticlesArticle#community#article#bigquery
Community
Crunchbase profile for TypeSafe AI: CB Rank 25, San Francisco headquarters, a Series A in Funding Rounds, founders Diogo Almeida, Erik Gafni, and Sasha Sheng, and Jev described as a System One Model returning typed decisions.
Jev Guides & Articles#community#company#funding
Community
Vercel's changelog for typesafe-ai/jev on AI Gateway, with AI SDK 7 experimental_evaluate and a support-ticket routing example.
Jev Guides & ArticlesArticle#community#article#vercel
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.