TypeSafe's Jev Can't See. I Made It Guess What I Drew Anyway
Blog experiment making text-only Jev guess what a sketch shows from a text description of the drawing.
This summary is based on how curated lists describe the project; its README was not available to review.
Overview
This blog post by Bartosz Mikulski describes an experiment with a model that cannot see images. According to the list summary, it probes what Jev, which takes no image input, can still recover when a sketch is described in text, effectively playing a drawing-guessing game.
It is a playful look at the limits of a text-only decision model and at how much the description of state matters.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What does this described drawing show?(inferred) | — |
How it works
- Turn a sketch into a text description.
- Ask Jev to choose what the drawing is.
- Compare guesses with the actual drawings.
Strengths
- Shows how state representation affects a text-only model.
Caveats
- Known only through a list summary.
- Informal experiment.
At a glance
- Type
- Article
- Status
- discussion
Use cases and integrations
How curated lists describe it
Blog: a drawing-guessing experiment that probes what a model with no image input can still recover from a text description of a sketch.
yibie/awesome-jev
Frequently asked questions
Can Jev see images?
No, the experiment converts drawings into text descriptions first.
What does it show?
How much a text-only model can recover from a description of a sketch.
Who wrote it?
Bartosz Mikulski, on his blog.
Related projects
ChatGPT-style web demo where Jev answers by choosing from predefined or custom answer sets, such as Yes/No, Pirate or Tabloid.
AppprototypeTypeScriptAwesome list and guide to System One models (Jev, Laya and others) organized around a five-stage agent lifecycle, with comparison tables.
ProjectproductionOpinionated catalog of 84 Jev projects in 11 categories, each with a hand-written reason it is worth looking at.
ProjectproductionUnofficial collection of Jev state formats, question patterns, cookbook index and case studies, plus a ranked Top 30 of community builds.
ProjectproductionBilingual catalogue of 183 Jev applications in 11 categories, explaining how each works and what its demo actually proves.
ProjectproductionKorean course code for attaching Jev to Claude Code, Codex and OpenClaw: skills, hooks, routers and open-model comparisons.
ProjectproductionJavaScript