Impatience with AI Agent

I am impatient with my AI agents. This is not about complaining how long an agent can take to finish a task. It is about when I notice the agent cannot do the job in a few attempts, I will decide to do the remaining task myself instead of continuing to instruct the agent to figure the task out. My involvement time including specifying the task is normally <15 minutes.

In contrast, I am much more patient with my human collaborators. When they show an incorrect result, I will spend much more time discussing with them (meeting, written notes and demo) and eventually let them figure the task out. My involvement time is normally > 45 minutes.

Why do I have such a different patience level? Here are my thoughts.

Epistemic Status. Humans often have better "epistemic status" than AI. This means my human collaborators can communicate their confidence levels effectively, helping me understand where I need to provide more information. In contrast, AI often seems overconfident, asserting that its output is correct even when it isn't. This makes correcting mistakes a difficult and frustrating process.

Communication Channel. With AI, the main channel is text. With human, we have voice, facial expression and body language that can convey information faster and richer. Easier communication channel also makes interaction more fun.

Empathy. I, as a human, understand human better than understanding AI. I know where human might need more information and how they might make mistakes, so I knew my feedbacks would be valuable. With AI, however, I am often not sure what extra information I need to provide to help the AI succeed.

Can I be more patient in the future? Probably yes.

Companies are training AI to become human collaborator, hence I expect AI’s epistemic status will become better over time. Also, I can probably add some system instruction to improve AI’s epistemic status right away.

More communication channels will be available in the future. Dictation Tools like superwhisper are becoming more and more popular. People start using voice for coding. For me, the main blocker for adopting this tool is to be brave enough to talk to a machine in public :)

With more AI interactions and research, we might become more empathetic with the AI. We become more empathetic to each other through deliberate practice like standing in someone’s shoes. We observe, interact and get feedbacks from each other. Similarly, can we become more empathetic with AI through practice? Yes and No. Yes because people have started noticing different “personality” from different AI. Those traits can help us predict what will happen next if we do XYZ with the model. In some sense, that’s empathy. Meanwhile, We might never be able to get as much empathy with AI as with human. The AI grows up and perceives the world in a very different way from human. Some subtle training algorithm or data change can often affect the model behavior in an unpredictable way.

Moving forward, I will try to be more patient with my AI agents. This doesn’t mean I’ll linger in a single failing conversation; it means I will embrace a new kind of patience based on trial and error. Unlike a human, I can restart a conversation with an AI a hundred times without causing exhaustion or resentment. Patience with AI isn’t about waiting for an answer; it’s about having the persistence to find the right question and try to understand them better. Understanding this distinction is the foundation for a productive future alongside AI.