Overlay extracted coordinates to find hidden errors

Published by Simon Willison

This credits the original publisher. Better Loop membership or a shared assessment is not implied.

The public work

Willison published bounding-box outputs and visual tools for locating objects in images. When boxes appeared rotated, he investigated image orientation metadata, tested a different image representation, and documented a preprocessing fix.

What to notice

Overlay extracted coordinates on the original image and check representation details before attributing errors to model accuracy.

Keep the context

This is an author-documented prototype with sample outputs and code, not a standardized object-detection evaluation. Historical API behavior and browser security were not validated here, and no production readiness is claimed.

AI use: Documented in the source.

The source explicitly uses Gemini 1.5 Pro through its API, Claude 3.5 Sonnet with Artifacts, and GPT-4o with ChatGPT Code Interpreter for investigation.

A useful public example is not an assessment of a reader, a publisher or a Better Loop member.

Authored practice suggestion

Try the idea.
Check your own work.

Use material you are allowed to work with. This suggestion is preparation; it does not record a completed task or an improvement.

A prompt to adapt privately

Choose a public image with several objects. Request labeled bounding boxes in an explicit coordinate order and scale. Overlay the boxes, check missed or misplaced objects, and inspect orientation metadata if locations look rotated. Record corrections before accepting the extraction.

A check to adapt

Boxes align with the intended objects under the stated coordinate convention, and missed objects or unresolved errors are recorded.

Open a related practice brief

Private practice needs no account. Your configured model provider may process the evidence you select.

See the full skill-to-sharing flow

Another way into the work.

Compare audio and visuals before summarizing video

OpenAI’s notebook processes a real DevDay recap into frames and audio, then prints summaries and answers from visual input, audio-derived text, and both together. The differing outputs make missing context visible.

Published byOpenAI Cookbook / Juston Forte

Curated public exampleGuide

Math & science

Selecting peaks in a recorded ECG signal

The SciPy example applies height, spacing, prominence and width conditions to a recorded ECG segment. It displays selected peaks, interval arrays and property values, making each filtering choice inspectable.

Published byThe SciPy community

Curated public exampleNotebook

Everyday work

Control summary detail without confusing length and quality

OpenAI summarizes a Wikipedia article about artificial intelligence using different chunk counts and a detail parameter. The notebook prints summaries and their lengths, including a recursive variant that supplies earlier summaries as context.

Published byOpenAI Cookbook