JevMade Sign in
← Back to experiments

Playable demos & bots

surprise

Bookmark: surprise Keep this in your collection.
Leave a noteWhat would you try with this? : surprise

Only you can see your notes.

Source screenshot of surprise
SOURCE SCREENSHOTFull screenshot ↗

What it does

An agent that predicts what actions will do, notices confident mistakes and rewrites its explicit model of the world.

How you can use it

A developer can adapt this idea for a game character that learns from its mistakes. The included 3D playroom offers a good starting point. The character watches a block fall and asks an outside AI service to guess what happens next. If the guess is wrong, the character writes a new rule.

To use the outside AI, a developer must run the project on a computer. The AI service blocks requests from the public website. The developer gets an access key from TypeSafe AI and adds it to the software to connect them.

Maker-reported (not independently measured by JevMade): 3D playroom curriculum, 164 predictions: Brier 0.116 for the hand-written rules, 0.207 for Jev zero-shot, 0.126 for Jev with the learned rules in context (author-reported, generated 2026-09-20) · One playroom curriculum run with Claude Fable 5.1 accommodating: 123 predictions, 82% accuracy, Brier 0.137, five concepts acquired (author-reported) · Jev on that workload: 3,398 calls, mean 389 ms under concurrency, 1,437,895 input tokens, $0.060 total (author-reported)

Primitives
noul
Platform
TypeScript
Added
Project created
GitHub stars
0 (snapshot, not live)

Keep this for later

Sign in to bookmark experiments, guides and videos, and keep notes only you can see.

Continue to sign in

We’ll bring you back to this listing.