Agreed. My news feed is blowing up with “AI agent” stories. Including the one that started this thread “ChaosGPT”.
What’s making them explode, I think, is that with the rollout of GPT-4, these agents are actually producing interesting things.
The magical thing that they do, is to “produce something” based on a few small inputs.
I have an example on this thread, where I only input “Become a machine learning expert.” as the Objective, and the Initial task: “Learn about tensors.”, and it starts churning out data like crazy!
It will go continuously and feedback its generated tasks based on previously embedded results. It becomes an echo chamber of “AI thoughts” and it is continuously trying to satisfy the initial Objective that you give it. Since it is breaking down things in steps (tasks) it leverages the power of “Chain of Thought” (CoT) that these LLM’s seem to tremendously benefit from.
So the feedback, CoT, and now the expansiveness of GPT-4 make these “super agents”, at least compared to previous efforts. I think this could be the next big thing, especially since it directly taps the CoT superpower, that makes any LLM seem like it is on steroids!