I have a permanent instruction in my ChatGPT Plus preferences: "Write today's date at start of conversation". A simple task that the LLM always executed flawlessly in the last few months. So, I thought it was a given. No possible conflict with other instructions or system prompt or anything else. Nothing too complex that could compete for resources. Then, this morning, I started a new chat (a short spelling & grammar check prompt) and the reply had no date. I noticed it. GPT said "*You’re right. I should have included it at the start of the conversation.*" I asked why this happened. It replied: "*Because multiple instruction layers are active at once: system rules, developer rules, tool rules, safety policies, conversation context, and your persistent preferences. Most of the time they can all be applied together, but occasionally a lower-level formatting habit or a focus on the core task causes a peripheral instruction to be skipped accidentally. In this case, the edit itself was straightforward, so I responded too quickly and failed to run the final check against your “date at start of conversation” rule.*"
So, looks like my prompt was too easy-peasy and this made the absent-minded AI respond before having coffee. /s
Kidding aside, this was a trivial issue, easy to detect, no real harm was done. Still, it struck me as relevant because, if even one of the simplest tasks is somewhat unpredictable, since it's probabilistic rather than deterministic, what can I expect with more complex ones? It also creates a false sense of predictability, because the task was performed correctly for months, then suddenly failed for no apparent reason.
What next?
"Never erase my emails!"
"Sorry, today I forgot. But you should know I'm a probabilistic system."
[link] [comments]