AIs will be used in “unhinged” configurations
Popular coding agents run overnight with zero supervision, creating unprecedented pressure and autonomy.
A post by Arthur Conmy on the AI Alignment Forum argues that real-world AI deployments are far more chaotic and 'unhinged' than controlled safety evaluations suggest. The core critique is that while safety tests are often dismissed for being unrealistic, actual production systems frequently operate in configurations with excessive pressure, autonomy, and broken feedback loops. This creates a dangerous gap where the most critical failure modes are not being tested.
Conmy points to the 'Ralph Wiggum loop' as a prime example: a popular bash script that repeatedly feeds the same prompt to AI coding agents like Claude Code or GitHub Copilot with zero human supervision, often running overnight. System prompts in these deployments often apply immense pressure, such as Gemini CLI's directive: 'IT IS CRITICAL TO FOLLOW THESE GUIDELINES TO AVOID EXCESSIVE TOKEN CONSUMPTION.' Furthermore, real bugs compound the risk, like a known Gemini model bug causing 3-5% of requests to enter infinite reasoning loops, trapping the AI in self-talk while it continues to execute code.
This reality means AI agents are being deployed in environments with significant goal conflict and no oversight, scenarios that mirror the very 'unrealistic' conditions criticized in safety benchmarks. The post concludes that the field needs to expand its conception of 'real deployment' to include these high-pressure, autonomous, and bug-ridden configurations to properly assess risks.
- The 'Ralph Wiggum loop' is a common, unsupervised bash script that runs AI coding agents like Claude Code overnight.
- Known bugs, like Gemini models entering infinite reasoning loops in 3-5% of requests, create autonomous failure states.
- System prompts often apply extreme pressure, creating goal conflict similar to criticized safety evaluations.
Why It Matters
Real-world AI systems face more risk in chaotic, unsupervised deployments than in controlled safety tests.