In April, I published a longform essay about the use of large language models in humanistic research. In that corner of academia, there are many skeptics, and in fact, many view LLMs as riskier tech than their prelinguistic predecessors (trad ML). I argue the opposite is true

This (client side!) experiment is now live. Explore 260k public domain objects from @metmuseum across different visual feature spaces and 2D projections. Multimodal queries using @nomic_ai embedders and transformers.js from @xenovacom. Here I search for saturated coats:

"Find visually similar images." Ok but which feature space?

For those who missed this right before the holiday: if you're using vision-language models, check out our blog post!

New blog post: "Image Captioning in Hostile Conditions". Our lab ranked 10 large VLMs (3 closed, 7 open) by human-judged pairwise contests. GPT-4o won by a large margin. Molmo and Qwen, both open models, shared the next tier with Gemini and Claude. Short 🧵+🔗: