Below is a loose transcript of my presentation for the final project from 7 July 2026. If you want to skip the whole introduction, navigate directly to the prototype.
– :Fabian Morón Zirfas

Dissecting Discussions

is a research artifact about the »Toxic Triangle project by Tactical Tech«

• to navigate the noise
• by Fabian Morón Zirfas
• in Summer 2026

The Toxic Triangle, to put it in the words of Tactical Tech, is »a unique look into young people’s online experience on social media and the industry behind it.«

We started this project with a workshop at Tactical Tech. My first contact was through these discussion-starter cards they used to hand out to decision-makers. Sena, Shashank, and I came up with some ideas (which I can’t remember) in a short period that day.

I left social media quite some time ago (~2019?), and it directly captured my attention. What is going on there? What happened? I don’t want my kids to be exposed to this toxic environment.

Where did this start?

In the Eco Social Software project, Fabian Ehmel and I teamed up and came up with this formula for not leaving prototyping mode: $Fabian²\; \times \; Ehmel\; +\; Morón\; Zirfas\; \times \; 2026$. We asked the following questions with our project:

• What if you want to know who is involved in the discussion?
• What if the discussion is too fuzzy to know who to listen to?
• How do we get a bird’s-eye view of the different actors?
• How do we know who is affiliated with whom?

We created these semantic networks and maps to visualize the relationships between different actors. We used semantic similarity to position them in space and get a grasp of the discussion space.

As a takeaway, I see the usefulness of semantic search as a tool in retrieval-augmented generation tools and others, but it is an intermediate step. Seeing the similarity between two words, such as “age”, on the screen is possible. But spotting the similarity of two or more chunks of text is not feasible without closely reading and analysing them. Using it as a sorting method might not be the best solution.

Back to Obsidian

Therefore, I went back to where we were coming from: force-directed graphs from the Obsidian graph view.

So here comes the pitch again:

Dissecting Discussions

• Online discussions have become an integral part of our daily lives.
• But forming an opinion based on unknown sources is risky.
• Who is behind the source?
• To whom are they connected? What is their motivation?

And of course for every research artifact, there are many trajectories that lead nowhere!

Dissecting Discussions tries to map the discussion landscape around the social media ban along:

• Political Spectrum (Z-Axis)
• Stance (X-Axis)

Take a look at the prototype and play with it here.

There are some caveats:

• The Y-Axis is still TBD.
• This is nearly all synthetic data. To make this useful for the discussion, it needs a significant editorial effort.
• Interaction is brittle. Some things work; others don’t. It is a vibe-coded experiment.
• No sane defaults yet.
• Filtering/search through properties is still sparse.
• No semantic positioning (yet?).
• Chromatic aberration is eye candy only. It looks nice but does not provide value.

Thank you for your attention!