Post

Gemini Embedding 2 is out, and it's natively multimodal! 🐶🔥 but you might be wondering what the heck that means and why that's important? well... take the "this is fine" dog for example...

12/3/26, 1:02 am0 reposts0 likes

Christina

@cszhu.dev

12 Mar

older encoder models analyzed this using two independent, single-track systems: 🖼️ image encoder processes flames/pixels, and outputs "danger"! 📄 text encoder processes the words, and outputs "calm"! they're isolated, missing the deep cross-modal connections. the model gets confused.

Christina

@cszhu.dev

12 Mar

Gemini Embedding 2 processes text, images, audio, and video in a unified vector space. Because it's multimodal, it actually captures the irony of a calm dog in a burning room. You can search your vectors for "ironic dog meme" and find it! read more: medium.com/google-clou...