home | .game | .club | .dao | .world

Spawning Community Intelligences with Mochi!

Table of contents


Since our last zine, Mochi has been quietly meditating on futures where crypto and AI collide.

“But Make It Kawaii” is our attempt to synthesize a decade of coordination + alignment research into something cute, playful, and digestible.

May this 💧 be the 〰️ that carries the 🌊 to the 🌏.


🦠 Coordinated 41 players across 17 teams doing 1 thing a day with AI

🦠 Curated 4 minimum viable datasets with image, text, code, and voice

🦠 Spawned 4 community-specific intelligences that live on consumer hardware


Artificial intelligence has all but captured human attention.

In the last six months, we’ve seen a proliferation of AI tools everywhere, many of which rely exclusively on proprietary models like GPT-4. Why?

Because attention 👏 ain’t 👏 cheap.

It’s a well-known fact that collecting robust datasets to train models from scratch requires a tremendous amount of time, energy, and capital investment. Some speculate that GPT-4 cost roughly $100M to train.

High coordination costs, paired with the scale and speed advantages afforded to companies developing their own in-house models, make it difficult for small organizations to keep up in the Intelligence Age.

Perhaps less appreciated are the online subcultures who’ve succeeded in training their own AI at relatively low cost.

In 2020, members of the My Little Pony community famously spawned 15.ai, a legendary text-to-speech (TTS) model capable of cloning any voice with only 15 seconds of audio.

By organizing contests in Discord, 15.ai (alongside the Pony Preservation Project) was able to intelligently leverage the nigh religious fandom of the “Bronies” to collect voice datasets and train a state-of-the-art AI voice model. This caught the AI community by complete surprise.

The project has since fallen into a mysterious state of neglect.