Vision AI stumbles on object counting

// 45d agoBENCHMARK RESULT

Vision AI stumbles on object counting

A Reddit user tested Copilot and Gemini on a dense image-counting task, asking them to count the number of cases in a photo. The thread turned into a reminder that multimodal chatbots can describe images well but still struggle with precise object counting without task-specific tooling.

// ANALYSIS

This is less a shocking AI failure than a useful boundary marker: general-purpose vision-language models are not reliable measurement instruments.

–Dense, overlapping objects remain a weak spot for chat-first multimodal systems
–Prompt correction can improve answers, but it does not guarantee exact counting
–The better engineering answer is segmentation, detection, or classical CV plus verification
–For developers, this is a reminder to wrap LLM vision with purpose-built tools when precision matters

// TAGS

copilotgeminimultimodalllmbenchmarkcomputer-use

DISCOVERED

45d ago

2026-04-22

PUBLISHED

45d ago

2026-04-22

RELEVANCE

5/ 10

AUTHOR

YERAFIREARMS

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

VIDEO1h ago

Mint remasters 2D games into 3D worlds

Mint (mint.gg) has released a demo showcasing the ability to remaster classic 2D games into interactive 3D worlds. Using assets from Pokémon Ruby, the platform demonstrates how 2D tiles and sprites can be turned into a navigable 3D environment.

LAUNCH2h ago

Jarvis enforces human-approved local AI execution

Jarvis is a local AI operator system designed to prioritize human oversight and strict system control by requiring explicit human approval for every proposed action. All steps taken by the AI are fully logged, inspectable, and subject to legal verification to provide a practical, audit-ready local environment.

UPDATE4h ago

Antigravity CLI updates add LaTeX and model selection

Three releases for the Antigravity CLI were rolled out in the past week, delivering numerous quality-of-life improvements based on user feedback. The updates include support for LaTeX math equations, the introduction of a new --model flag along with the agy models command, and a new /permissions command for managing permissions.