New AI model 'Jev' cuts agent loop latency by skipping text generation for decisions

·4 views

A common inefficiency in AI agent loops is that language models generate verbose prose just to produce a simple decision — such as which element to click — wasting seconds and hundreds of tokens per step. TypeSafe released a model called Jev four days ago that addresses this by accepting typed questions (yes/no, multiple choice, ratings) alongside a state input and returning structured answers with confidence scores in a single forward pass, without generating a token stream. A developer built an open-source plugin called jev-use that routes decision-only steps to Jev while keeping text-generation tasks with a standard LLM like Claude or Codex. Multiple questions about the same state can be batched into one call, keeping latency flat regardless of question count. Currently Jev is API-only with no self-hosted option, meaning state data such as page DOMs or command outputs must leave the user's machine to be processed.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Fake job interviews used to deliver malware and drain crypto savings

Microsoft's Defender Experts team documented a cyberattack campaign called Contagious Interview, active since at least December 2022, in which threat actors impersonate recruiters from crypto or AI firms to target job seekers. The attack is designed to reach the technical assessment stage, where candidates are socially conditioned to run unfamiliar code without suspicion. Victims are typically instructed to clone and execute a malicious NPM package from seemingly legitimate platforms like GitHub or GitLab. In one case reported by BBC World Service on 4 September 2026, a UK job seeker lost £18,000 in cryptocurrency savings within hours of completing such an assessment. Microsoft advises that no legitimate employer requires candidates to run unknown code on personal machines, and recommends using isolated virtual environments for coding tests.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Contributes Recent-Papers Section to Harvard's Open-Source ML Textbook

A developer successfully merged a pull request (PR #1962) into the Harvard CS249r Machine Learning Systems open-source textbook, hosted on GitHub with over 28,400 stars. The contribution addressed an existing issue requesting improvements to the StaffML section of the book. The developer built a new React component to display a curated list of recent papers on the About page, matching the project's existing design conventions. The pull request passed review by two project maintainers before being accepted. The textbook is used by thousands of machine learning students and practitioners worldwide.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Critical GitLab CVE-2026-85706 Exposes Over 1.2M Instances to File Read Attacks

A maximum-severity vulnerability, CVE-2026-85706, was discovered in GitLab's repository commits API, allowing unauthenticated arbitrary file reads with a CVSS score of 10.0. GitLab released patches on September 10, 2026, covering versions 19.1.8, 19.2.6, and 19.3.2, with CISA adding the flaw to its Known Exploited Vulnerabilities catalog the next day. Internet scans via ZoomEye identified over 1.26 million publicly reachable GitLab instances as of September 19, 2026, though the figure reflects exposure scope rather than confirmed vulnerable systems. Security teams are advised to inventory all self-managed instances, apply available patches immediately, and restrict public access where upgrades cannot be deployed right away. Additionally, service account credentials accessible through potentially exposed instances should be audited and rotated as a precautionary measure.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Stale READMEs Drive Away New Contributors — Here's an Automated Fix

Outdated setup instructions in open-source READMEs frequently cause new contributors to abandon projects before making a single contribution, according to a developer essay published on DEV Community. The author argues this 'README rot' is one of the costliest silent bugs in software, since frustrated users simply leave rather than filing issues. While the docs-as-code movement improved documentation review processes, it still relies on human diligence to keep content current — a dependency the author calls structurally flawed. To address this, the author built ReadmeForge, a GitHub Action that monitors code pushes, analyzes diffs, and automatically opens pull requests proposing README updates for human review. The tool is free for public repositories and requires no API key configuration, aiming to make documentation accuracy a continuous, automated side effect of normal development work.

0 comments Read more at DEV Community