At Rise of AI 2019 in Berlin, Trent McConaghy, founder of Ocean Protocol, argues that AI can be aimed at the UN's 17 Sustainable Development Goals, which he calls a de facto KPI for humanity, but that scattered AI-for-good projects will not reach global scale without better infrastructure connecting problem owners to problem solvers. He proposes an 'AI Commons': a public utility network, built on blockchain technology, that connects data owners, storage and compute providers and AI researchers the way ImageNet or the UCI repository connected researchers to well-defined problems. Using Bitcoin as an example of a public network no one owns or controls, he describes using token inflation to pay suppliers, such as telecom companies, for data like cell-tower signal changes useful for weather prediction, so that data stays free for users while suppliers are still compensated. He also describes fully decentralized federated learning, where hospitals train shared models without a central party controlling the data or the result, as a way to pool medical data such as DNA samples for cancer research without violating privacy. He cites pilot projects with hospitals in Munich and Singapore and a Grow Asia project on fertilizer allocation for farmers.
Key takeaways
- Trent McConaghy frames the UN's 17 Sustainable Development Goals as a de facto KPI for humanity that AI projects can target.
- He argues the main bottleneck for AI-for-good is a gap between problem owners with data and problem solvers with AI skills.
- He proposes an AI Commons: a blockchain-based public utility network connecting data, compute and researchers, modeled on Bitcoin.
- Token inflation could pay data suppliers, such as telecoms collecting weather-relevant signals, while keeping the data free for users.
- Fully decentralized federated learning lets hospitals train shared models, such as for cancer, without any party controlling the combined data.
Detailed notes
AI aimed at the UN Sustainable Development Goals
McConaghy frames the UN's 17 Sustainable Development Goals, agreed in 2015, as a KPI for humanity with roughly 180 measures tracking progress. He gives examples of AI already applied to them: satellite imagery predicting deforestation, tracking livestock movement as a predictor of conflict, micro-insurance for crops, smart traffic lights, and epidemic modeling that could trace an outbreak like mad cow disease to a specific herd. He notes the AI for Good Summit began in Geneva two years earlier with about a dozen UN agencies, grew to roughly 500 people and 20 to 30 UN agencies the following year, and was expected to reach more than 60 agencies and over 1,000 attendees that year.
The gap between problem owners and problem solvers
He argues the obstacle to scale is a gap between 'problem owners', people facing issues like deforestation or disease outbreaks who cannot formulate them as AI problems, and 'problem solvers', AI researchers who could build models but lack access to well-defined problems and data. He compares this to how ImageNet, created around 2011, or the decades-old UCI repository, turned raw problems into shared benchmarks that thousands of researchers could work on.
A blockchain-based public utility network
McConaghy proposes a 'public utility network' modeling the AI Commons on the electric grid or water system, but built with blockchain technology. He cites Bitcoin as proof such networks can scale globally with no single owner, growing from one node to more than 10,000. He argues blockchains let a network connect data owners, storage and compute providers and problem solvers into a shared substrate, similar to how a travel-booking system unifies many airlines' data behind one interface.
Paying for free data through token inflation
To supply data that is free for users but still worth collecting, McConaghy proposes rewarding suppliers with newly minted network tokens, the way Bitcoin miners earn roughly 12.5 bitcoins every ten minutes for securing the network. As an example, he describes telecom companies adjusting cell towers to gather signal data useful for predicting weather in Africa, where sparse weather stations create a data 'black hole'; a network that pays for this data through inflation could close that gap.
Decentralized federated learning for medical data
For sensitive data such as genetic information, he describes fully decentralized federated learning: a cancer researcher he knows has DNA data from only about 100 patients, each with thousands of variables, so instead of pooling that centrally he could update a shared model hospital by hospital, or even device by device, without a central party ever holding or controlling the combined dataset or the resulting model, avoiding the privacy risk of a middleman that could be reverse-engineered to leak data.
Early deployments in Munich, Singapore and Grow Asia
He cites a project between hospitals in Munich and Singapore modeling Parkinson's disease without moving GDPR-restricted patient data between them, and Grow Asia, a World Economic Forum spinoff using AI to help farmers allocate fertilizer, plus Singapore's government using the approach to balance data access and privacy in policy.
Quotes
Cattle don't like bullets, so they run.
Trent McConaghy
Do you want to sell ads, or do you want to change the world?
Trent McConaghy
Questions answered in this talk
What is the AI Commons?
According to Trent McConaghy at Rise of AI 2019, the AI Commons is a proposed blockchain-based public utility network that connects data owners, storage and compute providers and AI researchers so that AI-for-good projects can scale globally instead of remaining isolated pilots.
How could data stay free for users but still be paid for?
McConaghy proposes that a public utility network could mint and pay tokens to data suppliers, similar to how Bitcoin rewards miners, so companies like telecoms are compensated for gathering useful data such as weather-relevant cell-tower signals while users access the data for free.
How can hospitals share medical data for AI research without violating privacy laws like GDPR?
McConaghy describes fully decentralized federated learning, where models are updated hospital by hospital or device by device with no central party controlling the combined dataset, citing a Munich-Singapore hospital project on Parkinson's disease as an example.