Shorts

When AI doesn’t work, give it an example

Writing a working AI prompt doesn’t need to be annoying.

How often did you complain that AI gets worse because it ignores your instructions? No matter what you write, it always makes a mistake.

Show, don’t tell. Stop describing the outcome you need. It’s sufficient to describe a general idea to point a direction and provide examples to clarify the details:

Do you need to classify a text? Tell AI to classify the text as X, Y, or Z and provide at least one example for each class. Are you summarizing a text? Provide an example of what AI should retain in a summary and what isn’t important.

Providing an example has an additional benefit. You must know what you want. Often, when AI “doesn’t work,” we expect it to guess what we need because we don’t know it ourselves.

“I know it when I see it” never works.

As with everything in machine learning, the simple idea of giving AI some examples has a fancy name: in-context learning. Depending on how many examples you provide, it may be a one-shot in-context learning or a few-shot in-context learning. You may need to know it during a job interview… and never again.

Not only chatbots

Chatbots are only one of the possible interfaces for AI.
What’s the best interface? No interface at all.
AI works best when it’s hidden.

Large Language Models can transform data, make (constrained!) decisions, and search for relevant information in databases. Do the users need to type commands to make it work?

Is 2024 the year of proactive AI? Will we see more things anticipating users’ needs, like GitHub Copilot, but in a broader context?

So far, I have tested AI creating an action plan to resolve a customer support case using similar support requests from the past, AI adding reminders to your calendar if you promised to do something during a Zoom call, and a Gmail plugin retrieving relevant clients data from a CRM while you write an email.

None of these ideas require a chat interface. In fact, making it a chatbot would get in the way of getting the task done.

What do you overlook because you assume AI must be a chatbot?