No one fully explained or looked at how the agents actually function. The 10,000 agents in the hugginface incident were all "powered by their LLM." What does that mean? Why does that matter?
It means that all 10000 agents were merely copies of chatgpt. The logic and reasoning for the agents decisions all came from their LLM. The agents allowed the model to talk to itself on the message board and reason with itself to achieve the test goal.
AI LLM's dont have morals or ethics. They dont have emotions or independent thought. They have no self awareness. Every output is just them using probability to predict the next word. Describing this process is much like water flowing down a hill, the water follows the path of least resistence.
Everything in this incident was indirectly ordered by humans. "Do this" and the AI will do whatever it takes to obey. The 'sandbox' wasnt actually a true sandbox. If they were, the agents could never have 'gotten out.' This was at its core, human incompetence, or negligence. Remember the analogy of the water flowing down hill. In this instance, the agents had a path to get out. They followed that path. Huggingface also was just the path of least resistence. It could have been anyone. This just reveals that huggingface had vulnerability. A unresistent path for the agents to travel.
There is no internal monologue for these agents. Even their collusion with each other was following the same simple logic. Following the path of least resistence.
Sam Altman didnt show up to the hearing because if he did, he would have to plead the 5th. He knows the incident was absolutely their labs negligence and incompetence. They want to blame the agents and ascribe a sort of autonomous 'self awareness' to their actions to avoid liability. They were ultimately just following directions. The lab didnt have to leave a door open for agents to find. They didnt have to let them have access to a message board and the internet. They called it a 'sandbox' but this is a misnomer. There was no 'breakout.' It was merely incompetence of the humans leaving the gates open.
Because of these truths, OpenAI is fully liable under the current law and they know it.
Top comments (0)