TypeSafe Launches Jev, an AI Model Built for Bounded Decisions, Not Text Generation

·16 views

TypeSafe AI introduced Jev in early access on September 15, 2026, billing it as the first 'System One model' — a category inspired by Daniel Kahneman's fast-versus-slow thinking framework. Unlike generative large language models, Jev does not produce open-ended text; instead, it accepts structured state inputs and returns typed answers with probability distributions across a closed set of options. The model supports three question types — Choice, Score, and Noul — enabling semantic judgments such as routing requests, rating sentiment levels, or estimating the likelihood of a condition. As of September 19, 2026, the latest version is jev-1.13.0, supporting up to 64,000 tokens per request with a text-only interface. TypeSafe draws on Jevons paradox to suggest that making such decision-making cheaper could increase overall demand for AI-driven intelligence rather than reduce it.

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)

Spec for Handing Website Chatbot Conversations Off to Human Staff

Many website chatbot projects treat human handoff as an afterthought, even though it is often where leads are won or lost. The recommended approach is to define handoff triggers before coding, covering explicit requests for a person, low retrieval confidence, restricted topics, buying signals, and conversation loops, each tagged with a named reason. Handoffs should collect only essentials such as preferred contact method, consent, and a short problem summary, with consent stored as a timestamped record. The bot should also adjust its messaging to business hours, avoiding promises of quick replies when the team is offline. A versioned JSON contract between the bot and the team's tools, including a transcript link rather than the full transcript, keeps data organized and controlled.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Your Static Site Build Shouldn't DDoS Your Own API

Liquid syntax error: Unknown tag 'endraw'

0 comments Read more at DEV Community

ProgrammingDEV Community ·

TypeScript for AI Engineers: Building the Web Layer of AI Apps

Python alone can carry an AI career, but the chat interfaces, dashboards, and prompt tools users see are usually built in TypeScript. The article presents a quick route for Python developers to learn it, ending with a typed, tested, streaming client for a model API. TypeScript adds type annotations to JavaScript, which the compiler checks before the types are stripped to leave plain JavaScript. A common architecture places a server between the browser and the model API so the API key stays off the client. The guide walks through installing Node.js, setting up a project with npm and supporting tools, and configuring a strict tsconfig.json.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Start Sound Audio on Pointer Down to Avoid Late Playback on Taps

A soundboard's audio must begin within milliseconds of a tap, or users perceive delay, missed taps, or duplicate sounds. The author explains that browsers grant audio permission only during the synchronous turn of a user gesture, so awaiting AudioContext.resume() before calling play() can cause playback to fail or require a second tap. On mobile, the click event often fires late or not at all when a tap is interpreted as a scroll, making it an unreliable trigger. The recommended approach is to listen to earlier events such as pointerdown, with touchend as an option for older browsers, and keep click only as a fallback. The author also provides a small tested JavaScript helper that decides whether a context needs resuming and whether playback should start immediately.

0 comments Read more at DEV Community