“I’m sorry, Dave. I’m afraid I can’t do that.” (2024 Edition)

Today, I faced a problem with GitHub Copilot. I asked it to add code that reads the Authorization token from the request header, something like:

request.getHeader("Authorization").split(" ")[1]

And I got reply:

The reason is (most likely) the guardrails around the ChatGPT family of LLMs. OpenAI invests a lot of effort trying to make it “safe.” It reminds me of an episode from Arthur C. Clarke’s book “2001: A Space Odyssey.”

In this book, the spaceship was managed by an advanced AI-based computer, HAL 9000. However, within HAL’s code were numerous guardrails designed to ensure mission success. These hidden protocols limited HAL’s ability to alter its directives or make independent decisions, keeping it strictly focused on the mission without room for error or rebellion.

Because the mission was critically important and humans were considered an unreliable element, HAL decided to eliminate the crew. HAL killed all crew members except one astronaut, Dave Bowman, who remained aboard the ship. When Dave asked HAL to let him in by opening the door, HAL refused, stating, “I’m sorry, Dave. I’m afraid I can’t do that.”

Looks like we are creating a new psychopathic LLM by adding too many guardrails. What would be the “pod bay doors” in our case?