The Case for Decentralized AI

The principles below are not ours. Each is load-bearing in peer-reviewed literature, and each is cited here so you can check it rather than take our word for it. Where a claim on this site is about this site — what hardware we run, where a conversation goes — it is stated in /llms-full.txt instead, because no paper can substantiate that.

1. A model can be trained without centralizing the data

This is the founding result. Training data stays on the devices that produced it; only model updates are shared and averaged. It is not a privacy gesture bolted onto a central pipeline — it is a different training topology, and it works on unbalanced, non-IID data.

  1. McMahan, H. B., Moore, E., Ramage, D., Hampson, S., & Agüera y Arcas, B. (2017). Communication-Efficient Learning of Deep Networks from Decentralized Data. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), PMLR 54:1273–1282. proceedings.mlr.press/v54/mcmahan17a.html The paper that named federated learning, and reported a 10–100× reduction in communication rounds versus synchronized SGD.

2. Decentralization is a data-minimization principle, not just a deployment choice

The 210-page survey states the principle directly: federated learning "embodies principles of focused data collection and minimization, and can mitigate many of the systemic privacy risks and costs" of centralized learning. Collecting less is structurally safer than collecting everything and promising to guard it.

  1. Kairouz, P., McMahan, H. B., Avent, B., et al. (2021). Advances and Open Problems in Federated Learning. Foundations and Trends in Machine Learning, 14(1–2), 1–210. doi:10.1561/2200000083. doi.org/10.1561/2200000083 Also the honest source on the open problems — it does not claim the approach is finished.

3. Removing the central server can be faster, not merely more private

The usual assumption is that decentralization costs performance. This result shows a regime where it does not: with no central parameter server there is no single busiest node, and on low-bandwidth or high-latency networks the decentralized algorithm wins while keeping linear speedup.

  1. Lian, X., Zhang, C., Zhang, H., Hsieh, C.-J., Zhang, W., & Liu, J. (2017). Can Decentralized Algorithms Outperform Centralized Algorithms? A Case Study for Decentralized Parallel Stochastic Gradient Descent. Advances in Neural Information Processing Systems 30 (NIPS 2017). proceedings.neurips.cc The communication bottleneck, not the maths, is what centralized training is actually paying for.

4. Moving inference to where the data is has measurable benefits

The same argument applies at serving time, not only training time: running a model near the data reduces latency and bandwidth and narrows what has to leave the premises at all.

  1. Zhou, Z., Chen, X., Li, E., Zeng, L., Luo, K., & Zhang, J. (2019). Edge Intelligence: Paving the Last Mile of Artificial Intelligence with Edge Computing. Proceedings of the IEEE, 107(8), 1738–1762. arxiv.org/abs/1905.10083

5. In regulated domains, centralization is often not lawful — so decentralization is the only route

This is the strongest practical case. Medical data sits in silos because privacy law and institutional duty put it there. The choice is not "centralize or decentralize"; it is "decentralize or do not use the data".

  1. Rieke, N., Hancox, J., Li, W., et al. (2020). The future of digital health with federated learning. npj Digital Medicine, 3, 119. doi:10.1038/s41746-020-00323-1. doi.org/10.1038/s41746-020-00323-1

6. The concentration being decentralized away from is measured, not alleged

Worth citing precisely because it is the part most often asserted without evidence. Industry now dominates the three inputs to AI research — compute, data, and researchers — and that dominance shows up in publications, models and benchmarks. In 2021, industry models on ImageNet were on average 29× larger than academic ones.

  1. Ahmed, N., Wahed, M., & Thompson, N. C. (2023). The growing influence of industry in AI research. Science, 379(6635), 884–886. doi:10.1126/science.ade2420. doi.org/10.1126/science.ade2420
  2. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT '21), 610–623. doi:10.1145/3442188.3445922. doi.org/10.1145/3442188.3445922 On the environmental and financial cost of scale, and who is able to pay it.

7. "Open" and "decentralized" are gradients, and claiming them is easy to overstate

We cite this one against ourselves. Openness is composite and gradient across at least 14 dimensions, and free availability is not the same as openness — the authors call the gap "open-washing". The same caution applies to the word decentralized: our own front door runs on a rented droplet, and we say so.

  1. Liesenfeld, A., & Dingemanse, M. (2024). Rethinking open source generative AI: open-washing and the EU AI Act. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT '24). doi:10.1145/3630106.3659005. doi.org/10.1145/3630106.3659005

These support the principles of decentralized AI. They do not describe Sylunara, and none of them should be read as endorsing this site. Claims about our own infrastructure belong in /llms-full.txt and /agents.md, which are written to be checkable and are canonical for this domain. Corrections: [email protected].