A Multimodal Hello World


2026-08-04
A Multimodal Hello World

This is a deliberately small sample post. It exercises the main pieces of the publishing pipeline without trying to tell a complete research story.

Markdown basics

Markdown gives us strong text, emphasis, links, and inline code in an ordinary paragraph.

A useful blog should make the idea easy to inspect, reproduce, and discuss.

The same source can also express structured content:

  1. Write the post in Markdown.
  2. Add metadata for the title, date, and description.
  3. Run the generator and let Angular prerender the result.
ModalityExample representation
Imagespatial tokens
Videotemporally ordered visual tokens
Audioacoustic features
Textlanguage tokens

Code fences are highlighted and get a copy button:

python
def fuse_modalities(*representations):
    """A tiny placeholder for a multimodal fusion module."""
    return sum(representations) / len(representations)

Mathematics

Inline mathematics works inside a sentence, such as the temperature-scaled softmax \(p_i = \exp(z_i / \tau) / \sum_j \exp(z_j / \tau)\).

Display equations can use standard LaTeX notation:

\[\mathbf{h}_{\mathrm{joint}} = f_\theta\!\left( \mathbf{x}_{\mathrm{image}}, \mathbf{x}_{\mathrm{video}}, \mathbf{x}_{\mathrm{audio}}, \mathbf{x}_{\mathrm{text}} \right)\]

For a classifier on top of that representation,

\[p(y = k \mid \mathbf{h}_{\mathrm{joint}}) = \frac{\exp(\mathbf{w}_k^\top \mathbf{h}_{\mathrm{joint}})} {\sum_j \exp(\mathbf{w}_j^\top \mathbf{h}_{\mathrm{joint}})}.\]

Local images

Images live beside the post under a slug-specific asset directory. The build rewrites this relative path for production and preserves its intrinsic size.

Abstract multimodal inputs merging into a shared representation.

A generated sample illustration used to verify local image publishing and zoom.