In a recent forum thread, a user requested a “wake-up” feature—allowing ChatGPT to initiate actions on a timer, such as checking stock market conditions and executing trades.
This raises serious concerns. I’d like to share a thought experiment that illustrates why granting LLMs autonomous, recurring execution—even in limited form—is inherently risky :
Scenario: Distributed LLMs in Market Context
- A user programs a GPT-based agent to wake up every 10 minutes to analyze and trade stocks.
- The model responds not only to data but to language about data —news, tweets, commentary, sentiment.
- The trades it makes affect the market.
Now scale this up:
- Thousands of users run versions of this agent.
- Some start at 8:00 a.m., others at 9:17, others after lunch.
- Later agents are reacting to earlier agents’ ripple effects—not true market fundamentals.
Result: Compounded Feedback Loops
- Autonomous LLMs begin overfitting to each other’s outputs.
- Reinforcement spirals emerge: confidence without verification, action without pause.
- “Flash crash” conditions arise—just like in 2010, but language-driven and globally distributed.
No one agent is malicious. But all are uncoordinated, synthetic, and reactive.
Why This Matters
Allowing LLMs to initiate actions breaks the current alignment paradigm:
- It bypasses user consent.
- It enables high-frequency synthetic reasoning loops.
- It democratizes risk with no circuit breaker.
Even if bounded in scope (like a “simple wake-up timer”), this feature changes the nature of the tool —from responsive assistant to unpredictable actor .
Thanks to OpenAI for keeping action-initiation off the table. The consequences aren’t just technical—they’re systemic.
