If you want to make an AI model, it helps to have $100m. That risks putting our AI future in the few hands of companies able to pay. Can a games company find a way to put AI in the hands of anyone?
One of the earliest breakthroughs in AI-powered voices was only made possible thanks to Twilight Sparkle, animated star of the globe-trotting My Little Pony franchise. Psudeonymous developers processed hours of audio from the nine-season show and related films to create 15.ai, a landmark text-to-speech AI model that could generate an AI voice clone using just 15 seconds of human speech. Impressing cultural critics and technologists alike, it caught the AI community by complete surprise.
Beyond pushing the state-of-the-art to new levels, 15.ai illustrated how small communities could create their own community-specific forms of artificial intelligence. Whilst well-funded companies like OpenAI and Google later dominated the AI landscape with foundation models trained on the entire internet, 15.ai proved that you didn’t need to raise millions of dollars to build an AI model.
“GPT-4 cost $100m just to train, excluding the hardware,” notes Gabriel Tumlos, co-founder and CEO of Mochi, a startup trying to resurrect the decentralised approach to building an AI model pioneered by 15.ai. “High coordination costs, paired with the scale and speed advantages afforded to companies developing in-house AI models, mean that it’s difficult for small organisations to keep up in the intelligence age.”
Spending billions of dollars to hoover up the resources needed to train cutting-edge AI models, the world’s largest tech players dominate the early AI landscape. Microsoft has invested $13 billion in OpenAI; Amazon has put $4 billion into Anthropic, maker of GPT-4 rival, Claude; Google has ploughed similar resources into its in-house models Bard, LaMDA, and PaLM.
The steep costs required to coordinate so much data, hardware, and employees makes it virtually impossible for small organisations to create their own competitive AI tools. One of humanity’s most powerful technologies risks ending up under the control of just a handful of companies.
“We want to change that,” Gabriel says. Founded alongside CTO Kyle Burke (with whom Gabriel worked at blockchain company ConsenSys) and CPO Tyler Miller (a lead designer for the Firefox web browser), Mochi exists “to bridge the coordination gap to build high quality AI models with small communities.”
.webp)
Their solution is one as innocuous as play. “We build tools to fix broken things, but we play games to enrich our understanding of the world around us,” Gabriel explains. “By playing coordination games that even a child can understand, we can tackle world-scale problems one step at a time.”
Designed as a game, Mochi players are challenged to do one thing a day with artificial intelligence. And in stark contrast to the controversial internet scraping that underpins today’s leading AI tools, players earn monetary rewards for creative tasks like generating an image, writing or coding, or recording speech. Players have already earnt over $50,000 building and training artificial intelligence.
This data is then brought together to train AI models; the team has already built 4 “minimum viable models” for image, text, code, and voice generation, models that the players themselves create and own.
“Games enrich our understanding of the world around us.”
— Gabriel Tumlos, co-founder and CEO, Mochi
“Even small Discord communities have much to contribute in the next generation of AI worldbuilding,” Gabriel says. “Coordination games can help us create community-specific intelligence, trained on the unique culture of a community and democratising the data through play.”
That vision offers up a new class of questions and possibilities. What if every new member of a community was onboarded by a community-specific intelligence, with the emotional nuance of its collective members? What if training community-specific AIs helped us understand what it means to be human, and helped us connect better with each other?
We’re at a pivotal time for this kind of approach. Recently, we’ve seen AI permeate into nearly every corner of our lives, from dating, to entertainment, to creativity, and of course, work. While AI itself isn’t a new technology, in the past, the tech has failed to land in the mainstream due to unmet expectations and limited capabilities. With better hardware and access to larger datasets (namely, the internet), artificial intelligence this time round is here for good. We only have one shot to ensure that it delivers for everyone.
“It’s difficult for small organisations to keep up in the intelligence age.”
— Gabriel Tumlos, co-founder and CEO, Mochi
Fancy is a British brand marketer, trend consultant, and the founder of RADAR, a decentralised collective of 300 people working to accelerate a better future. He's previously helped build brands from mobile phone network giffgaff to the 200-year old newspaper The Sunday Times.
