jevfieldnotes
FIELDNOTES / VIDEOS

Jev videos and demos

Play the Android demo here. The three official videos open on TypeSafe’s site because their player restricts embedding.

01 / COMMUNITY VIDEO DEMO

Mobile Jev: a task on Android

droidrun / Mobile Jev contributors · GitHub · droidrun/mobile-jev

▶ VIDEO / PLAY HEREOpen original ↗

Use the play control to watch here, or open the original video above.

Watch a task prompt, action log and phone screen as the agent enters an Uber route.

What to look for

Check each action against the requested goal and the visible phone screen.

The recording ends at payment selection; it does not show a completed booking.

02 / OFFICIAL COMPARISON DEMO

Jev and GPT: side-by-side decisions

TypeSafe AI · TypeSafe launch article

▶ VIDEO / OFFICIAL DEMOOpen original ↗

The creator’s Vimeo settings restrict playback on other sites.

Watch structured answers appear side by side in the official launch demonstration.

What to look for

Compare the same input and questions, then read the conditions in the launch article.

The publisher uses GPT-5.6 Terra with default reasoning and notes that the short, dense input favors Jev. This is not a Fieldnotes benchmark.

03 / OFFICIAL APPLICATION DEMO

Jev plays Doom

TypeSafe AI · TypeSafe launch article

▶ VIDEO / OFFICIAL DEMOOpen original ↗

The creator’s Vimeo settings restrict playback on other sites.

See the model making decisions inside a game loop.

What to look for

Follow how the application turns game state into actions.

The model receives structured text state, not the video frames. Read the publisher’s setup notes alongside the demo.

04 / OFFICIAL COMPARISON DEMO

Wikiracing: choose the next link

TypeSafe AI · TypeSafe launch article

▶ VIDEO / OFFICIAL DEMOOpen original ↗

The creator’s Vimeo settings restrict playback on other sites.

Follow models navigating between Wikipedia pages by choosing links.

What to look for

Observe each choice and the path taken to the target page.

The publisher uses different reasoning settings across models and a two-stage selection for more than 255 candidates. Results apply to that setup.

Browse official articles and evaluations →

Explore more community projects and demos →