Decentralized AI event

Just attended Decentralized AI event at TalTech organized by Astra Tikas and it was an inspiring evening that made me think and take some notes.

First, I think main motivation and problem defined was ownership. We are afraid that LLM models are not transparent on their architecture, training dataset. We are afraid of monopolies serving AI models. Monopolies that can deprecate models, inject own biases, political will and be a more serious way to control population. So instead of search engines and social networks creating a bubble effect, we may be brainwashed by AI. Or worse, give control over education, infrastructure or military, to unknown LLMs that can backfire without proper evaluation.

Secondly, I saw there is a lot of confusion about definition of decentralized AI. Seemed like every speaker saw it as a piece of a puzzle and had own impression of it. I see decentralized AI as layers and each one of those can be confused for true decentralized AI

• distributed deployment - you have multiple devices instead of single machine
• partition tolerant - AI nodes can go down and come back online
• governance & ownership - it runs on machines owned by different people
• open - you can see the source code, model weights, detailed research paper, trained dataset (but it may have custom license!)
• offline-first and edge deployment - it can work offline on low-power device using local data
• distributed state - if AI needs a global storage (that cannot be deleted at whim of some corporation), it uses blockchain
• privacy-first - decentralized AI nodes can share confidential end user authentication, payment and health data without risk of leaks
• distributed operation payments - AI nodes sell their operation work and can bill end user with tiny costs (death by thousand cuts problem)
• agentic - AI nodes, are processes running in a loop, performing tasks from other agents using own or external resources

Picking one property does not make it a decentralized AI.

I doubt that we can put "trustworthy", "ethical", "self-aware" on the list. I think future is complex and we will see all mixes of architectures out there. I'm sure that there will be agentic military drones gone rogue, centrally owned corporate fleet of robots leaking video footage of day-to-day life at homes, AI health trackers selling data to corporations for ads, AI-powered cars having built-in training preference whom to kill (elderly or kids) depending on the government priorities and its population age structure diagram.

But at the same time, we build the future we want to live in. AI as new industrial revolution, will free up time of lots of people, improving efficiency and quality not only in coding, but also in physical world. Your 🐝 beehive will be a robot and you will be able to talk to it (invest now!)

Coming back to the meetup, apart from blockchain, DAO and micro payments, there were these topics raised:

• AI Tinder Agent advocating on your behalf with another agent to arrange a mutually benefitial interaction or partnership :) -> Innar Liiv
• EU regulations
• Decentralized AI is hard 🚀 and thus fun to build (I agree) -> Tevo Saks

Questions from the audience

• Why do we need crypto instead of fiat money (too small costs obviously, hard to move money)
• How do we connect agents to real world (MCP obviously)
• Practical solutions with business when we have race to the bottom (of cloud LLM prices)
  • marketing agents
  • awareness of decentralized small LLMs
  • enterprise = workflow-oriented solutions. KYC
  • web3 = gaming, creative AI solutions

Astra brought an example of a toaster not being an agent because it does not decide anything. But that made me wonder, what is real decision-making? Who has such freedom? Who is sentient and not only self-aware but capable of foreseeing the future of the decision? If LLM was trained on some data, and then put into a box to process tasks, is it really free? Are we?

At the end I wanted to ask about real examples of agents that panel uses, but we ran out of time. Maybe you can enlighten me instead in the comments?

Big thanks to Astra Tikas, the speakers from TalTech, Microsoft, AIRE, and everyone who contributed to the panel. Also, shoutout to SingularityNET and DeepFunding for supporting the movement towards open, trustworthy AI.

Feeling energized and hopeful about building a future where AI stays open, fair, and decentralized 🐝