Andrew Warner’s roundup on The Next New Thing stitches together clips from other creators’ Jev videos, with his own commentary between them. His through-line: Jev is not a writer. It is a fast, inexpensive decision step that returns probabilities over constrained choices, so it belongs inside larger agent workflows rather than in place of them. The clips are demos, not benchmarks, and nearly every number below is a creator’s own single run.

Decisions, not essays

  • An opening clip sets the frame: “It’s not a chatbot. It can’t even write a sentence.” What it does is make decisions faster and cheaper than general-purpose models.
  • Warner’s analogy, which he calls oversimplified: Jev turns a task into a multiple-choice test, where a model like Opus writes an essay.
  • Setup, via a clip from Lukas Margerie: on typesafe.ai, copy the quickstart agent prompt into Codex, Claude Code, or GrokBot, then generate an API key.
  • Warner says sign-up has flipped between open access and a waitlist more than once, and points to OpenRouter as a way in without the waitlist. That is access status at recording time.
  • Several featured creators say “Jeff,” and the auto-captions follow them; Warner remarks on it on air.

The pattern behind a voice-controlled browser

Moritz asks an agent to build a web app that listens to him and drives a browser in real time, then navigates Wikipedia by voice. He later said it had run faster at other times and suspected his connection during the recording.

His explanation of the loop is the most transferable idea in the video:

  • The dashboard page transcribes speech while he is still talking.
  • For each new fragment, a Node server lists up to about 100 elements on the controlled page and sends that list, plus his words, to Jev with a fixed set of multiple-choice and yes/no questions: what the user wants (navigate, search, click, type), which element, which website, whether the sentence is finished, whether he is talking to the browser or to someone in the room, and whether the action would buy, delete, or send something.
  • Plain code acts on the returned probabilities with thresholds: ignore the fragment if “is this a command” is below 0.5; wait for more words if “is the sentence finished” is below 0.6.
  • The model answers typed questions; ordinary code owns the decision rule. The clip does not show what the buy/delete/send question gates.

Sorting, routing, and triage

  • Email: an opening clip runs 500 emails through Jev, categorizing each, flagging which need a reply, and ending in a pie chart. Warner sees Jev as the tagging step, handing replies to a stronger model to draft.
  • Model routing: one creator compares about 12 prompts routed by Jev against sending every prompt to a top model. He reports roughly 70% savings because nine of the twelve never needed the top model. That is a small sample; savings depend on the task mix. He also suggests a toggleable skill, such as /jev, to switch routing on and off.
  • Customer support, from Mayank Aggarwal: a “third time asking” complaint about a broken export goes straight to the human queue. In a race against Kimi, the on-screen figures read 1.16 versus 8.38 (7.2× faster) and a 16× cost advantage. Warner says Mayank used Jev for the routing and, he believes, Kimi to draft the replies that were needed.
  • Zapier, in a sponsored segment: Warner says Zapier now has a Jev integration and gives filtering calendar spam as an example. He says he got the link directly from Zapier because it did not yet show up on Google. This capture has not checked the integration.

Scoring leads and searching memory

  • Lead qualification, from David Ondrej: a Typeform-style form with an admin view that re-scores the applicant after each answer—highly qualified, qualified, mediocre, or disqualified—so the app can decide on the spot who sees a Calendly link and who joins a newsletter drip.
  • Ondrej’s rationale: an LLM that reasons for 10–20 seconds is too slow at that moment, while roughly 100 milliseconds is not noticeable. In the demo the tier climbs from qualified to highly qualified as he changes his answers (founder/executive, “20,000 plus,” as soon as possible). He notes he never supplied explicit criteria; the model inferred them.
  • He also mentions hiring as a use. Ranking job applicants with an unvalidated classifier is a higher-stakes decision than the demo tests.
  • Agent memory, from Moritz: his “Claudia OS” is a folder of Markdown memory files, including daily notes. Before, the agent guessed which file held an answer from file names and keywords, read whole files, and, when writing, usually appended new facts to the daily file.
  • With Jev, the code sends the file list, the section list, a new bullet, and a batch of small typed questions (yes/no, pick one, score). Jev answers them in one request in a few hundred milliseconds; the code handles counting and chunking.
  • He reports one recall using 2,756 tokens instead of about 13,000—almost 80% fewer—at a cost of $0.00297. That is a single query. Warner says Moritz built it in Cursor with Fable 5.1 and had the model explain the design afterward.

Clips, lights, and an always-on agent

  • Clip scoring, from an interview on Greg Isenberg’s channel: a word-level transcript goes to Jev with several classifier questions, and it scores 17 candidate moments in about three seconds. Warner’s complaint is that the clips are never played back, so their quality cannot be judged from the video.
  • Smart home: through the Home Assistant MCP, “Turn off the top bulb” completes in about 300 milliseconds, where an LLM would emit a tool call and then wait on it. Warner nonetheless thinks this is not the right job for Jev and points to Needle, an on-device model he says works similarly; he has interviewed its founder.
  • GrokBot, from Eric Siu: an AEO/SEO content bot proposes 40 ideas drawn from his content and calls. Jev checks each against what is already live on his site and shortlists 10, with justifications, before the bot drafts them. Siu calls it a first pass with no guidance.
  • Warner’s point: an agent that runs all day accumulates spend—his own bill spiked the week before—and a cheap decision step in front of expensive calls helps.

Where it failed

  • One creator wired Jev to a Bitcoin signal making buy, hold, or sell calls every minute. It performed poorly, and his conclusion was to keep it away from portfolios and use it for routing-style decisions.
  • Warner thinks a day of tweaking right after launch was not a fair test, but keeps the lesson: Jev makes decisions against criteria—usually ones your agent supplies—and is not a forecasting tool.

Source clips

The video description links each featured video: Nate Herk (video), Lukas Margerie (video), Mayank Aggarwal (video), David Ondrej (video), Jay E (video), Moritz (video), Greg Isenberg (video), Matthew Berman (video), Syntax (video), and Eric Siu (first video, second video).

This post is an outline of the video’s full auto-generated transcript, checked against its title, channel, description, chapter list, and original publication timestamp. Clips are credited only where the narration names the creator. Speed, cost, and savings figures come from the creators’ own demos and have not been reproduced here.