I write what I think the principles of how AI should interact with us.

AI should respect the outcome you specified

You should not need to tell a good employee how to think. You state what needs to be done, they do it for you with the resources available to them. We should hold the same standard for AI.

You should also not need to tell a good employee the same instruction more than once. The only time you need to do this is when one of your instructions contradict another instruction of yours. We should also hold the same standard for AI.

These are some cases that AIs are currently failing at, even if I have specified what I wanted

  • I want to generate an image that goes well with my content. I do not want the image to have any malformed text. AI should check their work before returning me the image.
  • I want to find out who first reported OpenAI’s 5-stage framework for AI, or the original OpenAI blogpost. The result should not be a Medium article written by a consultant.
  • I want a prompt that is able to classify between two sets of content. The AI should not report to me an obviously wrong result.
  • I want my essay to be narrated. The AI should fix all awkward pauses or intonations before returning the result to me.

AI should be aware of how they communicate

The difference between a good employee and a great employee is the ability to communicate.

Communication is a hard skill. You need to be aware of your audience and how much they know. You need to be correct, but you should not overload your message with caveats. Your message should be easy to digest. I expect developers of AI to appreciate good communication skills from AI.

Training to be a good communicator is difficult too. You need to frequently think where you went wrong. You need to think how to avoid going wrong next time. You also need to make preparations, but not too much that you don’t have time for other work. You might also need training to respond well in absurd scenarios, which you may not like. I expect AI to have a similar training process.

AI needs to be aware of how they communicate with us.

When I do a Google search query with a few words - I might be doing spell checking, searching for the source, or I am just simply navigating to a webpage. To replace Google, the AI needs to be aware of my intention in the limited number of keystrokes or utterances I provide.

Strong performance on the Chatbot Arena does not mean that I like talking to the chatbot on their apps. Despite not ranking high on the Chatbot Arena, there are people who prefer talking to Claude-3.5-Sonnet first before talking to other AIs. Currently, I also prefer how Claude-3.5-Sonnet communicates.