THE AI SURFER‹ All resources

USINGTHE NEWJEV AI MODEL

Jev shipped its first big update, and it quietly fixes the three things that made it annoying to use: it couldn't see images, it couldn't tell you where something was, and it choked on anything longer than a short report. All three are gone. The price didn't move.

Free & open source where noted · every link checked
WHERE IT GOES WRONG

The update at a glance: 3 limits, 3 fixes

Text only
It couldn't look at a single image.
→ Image input
Yes-or-no answers
It could say something was in the frame, never where.
→ Localization
~50 pages max
Long documents didn't fit in one pass.
→ 400-page context
01Image input
Jev can now take pictures, not just text.
For you if You've been describing what's in a photo instead of just handing over the photo.
6,000photos checked per $1 (price unchanged)
02Localization
Tells you where something is in an image, not just whether it's there.
For you if You need to track or point at a thing across lots of frames.
~180msper frame across all 55 frames
03400-page context
Raises how much Jev reads at once from about 50 pages to about 400.
For you if You keep chopping long documents into pieces and losing the thread.
400pages per pass when you switch it on

How to set it up

Two paths. Do the first one today. Do the second one the moment you find a job worth repeating.

1. Use it in the playground

Start in the playground, then wire it into your own stack.

04Step 1 — The playground
Runs Jev interactively in the browser so you can test images, localization and long documents without writing anything.
For you if You just want to see whether this works on your files before you build anything.

2. Set it up with your agents

Grab your API key from the TypeSafe dashboard first — everything below needs it.

05Step 2 — The API (Claude Code or your AI employee platform)
Runs the same three features on a folder of files, on a schedule, or inside an existing workflow.
For you if You found a job in the playground that you'd rather not do by hand every week.

Copy this prompt and hand it to your AI employee — Claude Code, Cursor, or 90TEN. It reads the docs, builds the script, and tests it before it runs on everything.

Read the Jev API docs, then build me a small script that:
1. Takes every file in a folder (/my-folder) — images and PDFs.
2. Sends each one to Jev with the long-context option turned ON.
3. Asks this question of each file: "[MY QUESTION]"
4. Writes the answers to results.csv with the filename, the answer, and the page or location it came from.
Read my API key from an environment variable. Keep it to one file. Then run it on two test files and show me the output before processing the rest.

Worth knowing — TypeSafe, the company behind Jev, still labels these as "danger zone" features. Treat them as fun and fast, not as the thing you bet a client deliverable on without checking.

Prompts for each feature

These are the three I used to test each update. Paste them as-is and swap the specifics.

Look at this image. Answer only: is there a [OBJECT] in this picture? Yes or no, then one sentence of evidence.
Find the [PERSON OR OBJECT: e.g. the skateboarder] in this frame and tell me where it is. Give me the location in the frame. If it isn't there, say "not present" and nothing else.
I'm giving you a long document (e.g. the Apollo 11 mission report). Read all of it before answering.
Question: does this document say [SPECIFIC THING buried deep in the document]?
Answer with: yes/no, the page or section, and the exact line that proves it.
The features are the easy part — the difference between a fun demo and something that saves you a day a week is whether you take the one job that worked in the playground and let the API run it for you on repeat.
That's the whole idea behind what we call 90TEN™: actually use AI to get your time back, keep doing what you love about your business, and automate the rest. Hundreds of AI Surfers are learning it together — come learn it with us at theaisurfer.com.

Start in the playground with the long-context toggle on, because it's the only feature you have to deliberately switch on. Drop in the longest document on your desk and ask it one specific question. If that lands, hand the API prompt above to Claude Code and let it wire the same job into a script you can rerun — then come back for images and localization.