WIP: llama-3 as a censorship-resistant LoRA host on blockchain infra

[kobs30]

🛠️ WIP / Show & Tell

llama-3 as a censorship‑resistant LoRA host on blockchain infra

Hi folks!
I’m building a prototype that runs LLM inference (llama-3 today, StableDiffusion next) in a fully decentralized, censorship‑resistant environment.
Think of the LLM as a virtual machine (VM) executed layer‑by‑layer across blockchain nodes.

---

🔍 What’s working right now

• ✅ llama-3-8b inference on 3 cyclonechain nodes
• ✅ Decentralized randomness for token sampling
• ✅ Consensus on prompt‑embedding + random‑seed → deterministic generation
• 🧪 StableDiffusion VM in progress (GPU cost is the blocker)

How it flows (1 packet consensus → deterministic LLM)

1. Nodes reach consensus on prompt‑embedding + random‑seed (~7 s).
2. Each node runs full llama-3-8b inference inside its local VM (deterministic).
3. Results are identical → recorded on‑chain; no extra consensus round needed.

---

🤯 Why this matters

• LoRA / SD models are fragile on centralized hosts (takedowns, ToS, etc.).
• This infra aims to enable permissionless deployment and long‑lived links.
• Opens a playground for experimental model‑sharing and DAO‑governed updates.

---

🌐 Live demo

https://cyclonechain.com/projects/demo.html
(Proof‑of‑concept: ~7 s consensus + extra VM time; slow but real.)

---

📖 Full design breakdown

https://medium.com/@sergeik30/liberating-ai-can-we-build-a-truly-decentralized-llm-on-blockchain-principles-63de54f8e4a3

---

💬 Looking for feedback

• Would you use a LoRA host like this?
• What features would help model creators the most?
• Does it solve a problem you’re facing today?

Thanks for reading!
— cycloneLLM