// writing

Research & write-ups

Personal, non-commercial research notes — reimplementations and analyses that probe where modern vision and multimodal models break, and how to know when to trust them.

// image placeholder
VLM blindness probe

// vision-language models · reimplementation

Reimplementing "VLMs are blind"

A from-scratch reimplementation and extension of the probes showing that large vision-language models fail surprisingly basic visual tasks — counting, spatial reasoning, simple object relationships. Why do models that "see" so well in demos stumble on things a child handles?

VLMs evaluation failure modes

Read article →

// note Personal, non-monetized research notes shared for transparency and discussion. More pieces will be added over time; the article body below is a placeholder to be filled in.