30 ideas, open opportunities and challenges in decentralized & distributed AI

In this post I want to explore the most promising pieces of technology that will enable open source and/or decentralized AI and agents to successfully compete with existing products from OpenAI, Google and Microsoft.

Over the last year we’ve seen a surge of interest towards decentralized AI which mostly driven by these 4 factors:

  1. Fears of centralization and control over most capable AI models by a single or very few organizations or nation states
  2. Lack of experimentation and innovation in closed-source ML research and development
  3. Bias and lack of transparency in closed-source, closed-weights models
  4. Bags pumping, but not mine. Anything “decentralized” sells best if it fits the dominant narrative on the market (which today is clearly “AI”).

Last year we’ve seen a surge of simplistic, yet relatively successful (valuation-wise) decentralized AI projects. Most of those were focused on token-gates chatbots, GPU rental services and LLM-based agents capable of interacting with blockchains. Even though, those are amazing and sometimes super fun experiments, I would argue that the true value of decentralized AI lies a bit deeper.

AI is obviously useful to pretty much every industry in the world, including web3. However, more important question is what building blocks from cryptography, p2p systems, distributed fault-tolerant systems, consensus and mechanism design can be applied to solve the abovementioned problems.

Compute

Compute refers to the processing power needed to run and train AI systems. Decentralizing compute distributes the computational load, increasing resilience and reducing centralized control over AI capabilities.

End state

  1. Widely available trustless, permisionless and private compute no more expensive than current cloud offering.
    1. For a developer, there shouldn’t be any UX difference between running my agent code and inference on AWS, Replicate or a trustless alternative.
  2. Decentralized compute works for any kind of job, not just EVM code execution or small model inference.
  3. Efficiency: it should be cheap and efficient to run complex ML tasks in web3. For example {{{ADD training, inference, TEE}}} cloud using sharded inference, Deloco, petals
    1. Sparse coding and CPU training: https://www.thirdai.com/bolt-technology/
  4. Privacy: there should be (ideally, cryptographic) privacy guarantees
  5. Reliability: web3 cloud should provide some form of SLA or reliability guarantees to be competitive with AWS and GCP (not saying they don’t have outages but they do have a brand name).

Data

Data is the lifeblood of AI, essential for learning and decision-making. Decentralizing data improves privacy, reduces monopolies, and increases the diversity of input, leading to more robust and fair AI systems.