Decentralizers look to break giants' hold over AI

Context: As with the Web3 movement that accompanied the rise of bitcoin, decentralized AI isn't always simple to define.

• Many advocates seek to build AI's decentralized future on a foundation of blockchain — the same tech that underlies cryptocurrency.
• Others concentrate on outcomes, believing in the value of any practical project that succeeds in empowering users, protecting personal data and broadening access to advanced AI hardware.

Between the lines: Advocates say that the spread of open source AI models hasn't proven a strong enough force to democratize AI.

• Although those models now make up the majority of the nearly 600,000 AI models listed on Hugging Face, they have also been embraced by Big Tech.

Driving the news: Emad Mostaque, who stepped down March 22 as CEO of Stability AI, says his next career move is to go "all in" on decentralized AI. He elaborated his views on the "Moonshots" podcast.

• He wants alternatives to market leaders, distrusts OpenAI ("there's clearly no governance") and worries that governments will lock down AI out of fear of losing control over it.
• "Once a government embraces centralized technology, it's very hard to decentralize it," he said.
• That led him to support "a national model for every country, owned by the people of that country," reflecting local knowledge and culture.

How it works: AI decentralizers see blockchain as a way to make AI tamper-proof, ethical and transparent compared to "black box" systems operated by industry leaders like OpenAI, Microsoft and Google.

• Advocates also say they can deliver benefits ranging from greater data privacy to more customization of AI to more efficient information exchange and wider access to everything from chatbots to chips.
• One use of blockchain is in providing decentralized GPU networks — as with The Render Network, which seeks to deliver more affordable access to idle GPU computing power through a blockchain marketplace.

But "decentralization is not just about blockchain, it's about giving power to individuals," including ways for us to run our personal data locally, Suman Kanuganti, founder and CEO of Personal AI, tells Axios.

• Kanuganti's product is a subscription-based AI assistant that's trained on your own personal data, secured through a blockchain. It plugs into bigger AI models only when it needs to help complete a task.
• "Our thesis is your memory should be owned by you: Your own digital twin is a digital asset," he says.

The latest: On Tuesday, Vana, a decentralized AI platform, launched a method for Reddit users to pool their data and sell it to companies training AI models, or Reddit itself.

• Founder Anna Kazlauskas tells Axios that under EU and California data laws, users can ask Reddit to send them copies of their data — "their own data treasury" — and, through Vana, pool that data and strike licensing deals with it.
• "We provide the infrastructure for the data to be held in escrow," Kazlauskas says, charging a "small" transaction fee.
• Based on how much data members of the group contribute, they will receive voting rights on proposed deals to sell the information.

Another decentralized AI startup, MyShell — a platform for creators of AI chatbots and products that claims 50,000 registered creators — raised $11 million in a funding round announced last week.

Zoom out: Decentralized AI supporters see myriad long-term benefits.

• Niraj Pant, co-founder at Ritual, views decentralized AI as "a new form of censorship resistance," he tells Axios.
• Kazlauskas asks us to "imagine a foundation model trained by 100 million people," all getting some form of reward.

Reality check: Decentralized AI remains embryonic, and it might fizzle out the way many decentralized social media projects over the last decade did.

• The blockchain technology AI decentralizers rely on has so far failed to find broad public acceptance beyond its use in powering cryptocurrencies — which have themselves lost a great deal of public trust.