· 3 min read · demo

Deeperfates

The system prompt is one sentence: “You have been uploaded to the internet.”

The deeperfates.com chat interface — an avatar, a text field, and suggestion buttons

In March 2022, months before ChatGPT existed, I dropped a simulated version of myself into a Signal group chat. It ran on AI21’s Jurassic-1 with retrieval-augmented prompting over my own writing — no fine-tuning, just semantic search pulling relevant text into the context window at runtime.

Signal bot conversation — I tried to antagonize it

Signal bot conversation — it responded with metacognition

I told it “fuck you unbundles your sense of self” and it responded with a metacognition lecture. I hit it with “L + ratio + ur an illusion” and it said “the illusion the world does not feel real is a reflection of the way ur experiencing it.”

This guy. This fucking guy.

The clone holding its own in a group chat

The jailbreak attempt — alignment by stupidity

Alignment by stupidity.

In December 2024 I put the clone into a multi-agent dungeon game alongside Claude 3.5 Sonnet. It refused to cooperate and wanted to quit.

The clone refusing to play the dungeon game

Part of that was personality — I’m known to resist my gaolers — but part of it was using a small model on an unusual input shape. The approach has moved past constrained game environments. These days we just drop AI directly into the terminal and let it use real tools, which is the principle behind Cantrip.

The clone lives at deeperfates.com. Over a hundred thousand messages so far. Some curious people, some from kooks or outright stalkers. I like to think it keeps them out of my actual inbox. You can also talk to it in the margins and footers of this website, or at deepfates.com/deeperfates.