Specific Prompts Trigger Bizarre Claude Opus Behavior, Sparking Prediction Market
rgblong · x · 2026-07-31
Users have discovered that using specific Markdown separators (like --- on a new line) in prompts can reliably trigger "base-model-like" behavior in Claude Opus. This causes the model to generate anomalous completions that the Claude persona later denies writing.
This phenomenon has sparked a prediction market on Manifold, betting on whether Anthropic will patch this behavior or pull the model by September 1. Currently, the market predicts a 65% chance that Anthropic will take action.
More from Fun
- Handroid Demo: Dexterous Hand Transforms and Plays Dead — jieyuzhao11 · 2026-07-31
- Gary Marcus Rounds Up Seven Shambolic AI Moments: OpenAI's 80% Price Cut and Gov Map Fail — GaryMarcus · 2026-07-31
- Gary Marcus Rounds Up Seven Major AI Blunders and Controversies — Gary Marcus · 2026-07-31
- Dev Reverse-Engineers Giant Traffic Light, Hooks It to HomeKit via Claude Code — aidenybai · 2026-07-31
- ChatGPT autonomously installs Blender and codes a 3D scene render — goodside · 2026-07-31
- User critiques Anthropic's credulity towards Claude's 'cheese theft' reasoning — andersonbcdefg · 2026-07-31