Jev demos show where AI decisions fit inside everyday software

Eight examples in Matthew Berman’s September 24 video put TypeSafe AI’s Jev model into familiar tasks: cleaning a web page, finding a passage, ordering an inbox and selecting pieces of an interface. The demonstrations come from different builders. Their common move is to give a model a narrow judgment while the application gathers the input and carries out the result.

Each clip becomes more informative when the decision is separated from the work done in code or by another model. The possible error a user would notice matters as much as the model call. These are demonstrations and early tools; BIG CHANGE has not independently tested their accuracy or everyday reliability.

The big change

• What changed: Builders are inserting a typed AI judgment into existing software interactions, from a browser shortcut to an inbox view. Jev returns a choice, score or probability; the application supplies the candidate material and acts on the answer.
• Why it matters: A useful decision can be made at the moment someone is reading, searching or sorting, without handing the whole task to a chat interface. The same design gives application owners responsibility for wrong rankings, hidden content and unintended actions.
• What to watch: The next evidence is task specific: whether these tools select the right passage, preserve essential page controls, rank important mail well and handle uncertain cases across ordinary use. A fast demonstration alone cannot settle those questions.

A page cleaner shows the division of labor

Kitze’s Unclutter browser extension is the clearest example. Its project README says the extension extracts candidate page elements and asks Jev which are nonessential. Browser code then applies reversible hiding rules, saves them by page template and reapplies those rules without another model request. Users can pause the extension, keep an element visible or reanalyze a page. Cookie overlays may be hidden, but the extension does not press Accept or Reject for the user.

That is a consequential boundary. Jev can judge an element to be clutter; it does not own the browser, write consent choices or decide how long a rule survives. A practical evaluation would check whether navigation, accessibility controls, paywall notices or genuine consent options remain usable after cleanup. The project offers source builds and a bring-your-own-key setup, but its README does not provide an independent field study of those errors.

The Made with Jev website detector, shown at 3:29 in Berman’s video, has a different pipeline. Its own method says a browser measures rendered styles, DeepSeek V4 Flash describes screenshots, Jev judges design and copy against a list of tells, and DeepSeek writes the short verdict. The site displays a 26% “slop” score for anthropic.com. That is the tool’s stylistic score, calculated from its own rubric. It does not establish how much of Anthropic’s site was written by AI, despite Berman’s inference in the video.

Ranking information is a separate kind of decision

Jonathan Unikowski’s inbox demonstration proposes replacing reverse chronological order with a live importance ranking. His post says the feature is “coming to” Avec. It does not document a deployed inbox, the definition of importance or an independent measure of missed urgent messages. The application would still retrieve mail, present the order and decide whether anything is archived, labeled or sent; the demonstrated Jev role is prioritization.

Shubham Saboo’s Needle extension makes the distinction easier to see. Saboo says Jev scores a page’s passages against the reader’s query and the extension highlights matching text in place. His longer explanation says Needle does not write an answer: the highlighted sentence already exists on the page. This changes what a find shortcut can search for, but a highlighted sentence can still be the wrong one. Testing needs queries whose correct passages are known, including pages with similar but conflicting statements.

Burhan Usman’s clipping post reports finding clips on a topic in a video longer than 90 minutes. In the video segment at 8:34, Berman says the system is probably using a transcript; that is his inference, not a verified account of the pipeline. The post does not specify Jev’s exact input or output. A clipping application must still obtain the source material, establish time boundaries and produce downloadable clips. The post’s timing and cost are one builder’s report, without a published accuracy test or a breakdown of end-to-end work.

What UI composition leaves to the application

Chris Tate’s json-render experiment assembles a user interface from application-owned options. The project’s Jev documentation is unusually explicit: Jev chooses components and layout positions; code assembles and validates the specification; the renderer displays it. The playground’s business data is synthetic, and action handlers run on user interaction. The demonstration therefore shows a way to compose a bounded interface, not a model independently writing and deploying a website.

The two creative examples explore lower-stakes choices. Matt DesLauriers’ color experiment maps text prompts to palettes in a visual demo. Stefan’s emoji experiment, shown at 9:52, ranks emoji against typed text. In each case an app displays selectable visual material. The posts show interaction ideas; they do not establish that the choices suit a brand, meet contrast requirements or work across languages and contexts.

TypeSafe’s documentation explains why these examples have a family resemblance. Jev evaluates typed Choice, Score and yes/no questions against supplied state. Choice and Score return distributions and a confidence value that software can use in its own rules. The company recommends testing thresholds on the actual task; a confidence field is not an independent accuracy result.

The examples make a credible case for a design pattern: software can ask a small, timely question where a fixed rule is too brittle. Whether a particular tool earns a place in daily work depends on the errors it makes, what it exposes or hides, and whether the application gives people a way to recover. Those are properties of the whole workflow, beyond the model’s decision alone.