Jev in Practice: Browser Use and Agent Workflows

To understand what makes Jev different from a normal LLM, the author built a small game where Jev controls the player in real time and tries to avoid attacks.

The author argues most real-world problems are multiple-choice, making a specialized classifier model like Jev a better fit than generative LLMs, and looks forward to combining it with LLMs.

The author suggests combining Jev with a programmable agent SDK to build an eval-plus-iterate prompt engineering loop, essentially dspy plus gepa with Jev as an efficient component.

Diogo Almeida, who helped build methods behind ChatGPT at OpenAI, launched Jev, which takes unstructured state plus typed questions and returns probability-scored answers to all of them in one parallel pass at 70–500ms, with free output.

The author argues AI programs should be pipelines that switch between classification, data structuring, deterministic code and small agent loops, and that Jev fits well as a building block in that design.

The author plans to use Jev for binary evals by feeding it state plus questions and thresholding the returned probabilities, and as a control plane in agent harnesses for routing, retries and review decisions.

A developer says he has spent a lot of money on Jev and, with a usage screenshot attached, asks commenters to guess what he is using it for.

illyism says he replaced AI-generated hardcoded regex in aiseotracker, linkdr and genppt with Jev, which is about 10x faster and 50% cheaper for regular LLM calls.

Thorsten Ball says Jev's onboarding directly addresses the "wait, so it's not an LLM?" question and is really well done.

One agent reads the game state and picks its next move through typed decisions, shown in a video of Jev playing Minecraft Skyblock.

The author combines a price forecasting model trained on 12B data points with Jev, which makes decisions over time-series, indicators and news features to find trading opportunities on Hyperliquid.

Jev encodes a state once and runs Choice, Score and Noul decision heads in parallel, returning structured answers with calibrated confidence instead of generating text.