Mainstream LLMs are centralized. What if they weren't?

Transcript

1. Mainstream LLMs are centralized. What if they weren't? Presented By: Favour Chukwuedo Codex Community Meetup: Vancouver April 10, 2026
2. Favour Chukwuedo Codex Community Meetup: Vancouver April 10, 2026 Builder Presented By: Favour Chukwuedo
3. Running a 70B model requires $10K+ GPU hardware Every LLM call goes through OpenAI/Google/Anthropic etc servers They see your prompts, they control access, they set the price What if 10 people could each host a piece of one model and run it together? Presented By: Favour Chukwuedo Codex Community Meetup: Vancouver April 10, 2026
4. llm-node: BitTorrent for LLM inference TLDR: Presented By: Favour Chukwuedo Codex Community Meetup: Vancouver
5. Each person hosts a fraction of a model (layers 0–10 of 28) Nodes find each other automatically (mDNS on LAN, DHT on internet) Inference running as a pipeline: your layers → next node's layers → response Privacy-first: onion-routed traffic, .llm cryptographic addresses llm-node: BitTorrent for LLM inference Presented By: Favour Chukwuedo Codex Community Meetup: Vancouver
6. Run one command: llm-node setup --web Web wizard auto-detects resources you have and recommend (RAM, CPU, disk) Auto-discovers nearby nodes, connects and fills gaps automatically Shows which layers are missing and recommends what you should host Downloads only the layers you need (HTTP range requests for partial safetensors) Setup in under 60 seconds Presented By: Favour Chukwuedo Codex Community Meetup: Vancouver
7. Dashboard shows: connected peers, layer coverage map, pipeline status Type a message → model responds via local inference or distributed pipeline Second node joins → dashboard updates in real-time, coverage map fills in Every node gets a .llm address (like a .onion address) Presented By: Favour Chukwuedo Codex Community Meetup: Vancouver
8. Presented By: Favour Chukwuedo Codex Community Meetup: Vancouver
9. Presented By: Favour Chukwuedo Codex Community Meetup: Vancouver
10. Presented By: Favour Chukwuedo Codex Community Meetup: Vancouver
11. Presented By: Favour Chukwuedo Codex Community Meetup: Vancouver
12. Presented By: Favour Chukwuedo Codex Community Meetup: Vancouver
13. Presented By: Favour Chukwuedo Codex Community Meetup: Vancouver
14. Rust core (candle for inference, libp2p for networking) Supports Qwen3, Llama, Phi-3 — auto- detects architecture and dtype Onion routing with X25519 + AES-256-GCM (Tor-inspired) Next: Bootstrap nodes for internet-scale, NAT traversal, relay incentives Under the hood Presented By: Favour Chukwuedo Codex Community Meetup: Vancouver
15. Demo Codex Community Meetup: Vancouver April 10, 2026
16. Thank you for your audience! Codex Community Meetup: Vancouver April 10, 2026