Doom Haikus: News Haikus from 2020 by Eli Holder

2020 News
Was All So Bad, But Maybe
Less Bad as Haikus?

Prompting and visualizing 2,700 gloomy haikus about news events from 2020 (from ~2,000 authors on Mechanical Turk), forever memorializing the worst year of our lives, as anxious sets of 5, 7, 5 syllables.

Background / Motivation: This project had 2 purposes:

1) Primarily it was a coping mechanism. Some people coped with 2020 by making banana bread or learning to play guitar. My de-stressor was making silly things with data.

2) It was also a technical exploration. The project started with a seemingly simple goal: Teaching Tensorflow to write haikus about current events. Just before GPT-3 came out, it wasn't entirely clear what kind of creative writing we could expect from AI text generators. But since poetry is subjective, there aren’t straightforward objective functions to quantify good haikus, so we’d need some actual haikus as training data (i.e. self-supervised is out). So this was an attempt to build a dataset for training a model to generate silly haikus about the news.

The Visualization / Exploration Tool:
https://doomhaikus.3iap.com/

By the end of 2020, about ~2,000 people had submitted ~2,700 haikus that were compelling in their own right (and sad, and funny, and sharp). The goal with the exploration tool was to make them easy to explore and experience, within the context of all the other chaos happening that year.

So the tool had 3 design goals:

1. Make it easy to browse the haikus. The site was primarily content-forward and the "viz" was baked into the navigation.
2. Give visibility to the HUGE thematic stories of 2020. You can see the Ahmaud Arbery, George Floyd, and BLM movements enter the consciousness around May ("Racism & Unrest"). But also, some big stories would wax/wane ("Covid-19") or not break through at all ("Climate Change"), while some stories were unnervingly constant ("Trump").
3. Use (fast fast fast) text search to highlight stories about users' long-tail personal interests (e.g. search "Bernie").

The project made it to the first page of Product Hunt, and was featured by Data Is Plural, Nightingale's newsletter, boingboing, and AV Club.