🐶 What Telegram quietly built on TON

Six months after Pavel Durov announced it in Dubai, COCOON is live.

Strip the buzzwords. It lets you run an AI model on someone else's GPU without that GPU's owner being able to read what you asked. ChatGPT logs your prompts. COCOON cannot. Nobody in the middle can.

Use cases: third-party Telegram translation, AI bots that don't leak (Mira's Private Mode is the working example), autonomous AI agents on TON that can't be front-run.

I run cocoon.doge.tg, the dashboard that decodes every transaction.

Full deep-dive, including the April outage that revealed Telegram is running more than one inference path:

→ Read on Telegraph