Working with AI

Yesterday, in the thinkdeli whatsapp community we had an interesting discussion around using AI for writing. A couple of us voiced their fears about losing their voice. One of us showed how he was using AI to improve his posts. One of us asked, “At what point does your article stop being your article?” We have started using AI apps in our every day lives. But so many questions have stayed unanswered. There is no escaping this AI zeitgeist. We need to embrace AI and make it work for ourselves.

To take advantage of AI, we need to understand how these applications work. We need to understand their limits. From my personal experience, most of the AI applications, work well as assistants. Even with thinkdeli, I never get AI to write the entire thing. I use it more to fix gaps and discover insights.

Since the last few months, I have been working on building AI applications for a living. thinkdeli, is one of them. Here is something I want to share with you. It is mind blowingly frustrating to work with the current state of AI for production level systems.

For me, the main reason that working with AI systems is frustrating is that the outputs are non deterministic. Meaning, you can’t be a 100% sure about the outcome every single time. Consider this scenario. You are trying to build an AI agent that performs appointment bookings for you. This seeminlgy simple problem statement has many nuances. Is the requested appointment time valid and in the future? Are there any conflicts? If yes, what alternate times are available? The user may ask “Is 3:30 pm available day after?” or “is anyone available today?” Point is, when dealing with real problem statements, there are just too many scenarios. For most developers coming from a transactional - request/response mindset - this requires a change in the way we think and write programs. As a developer, your job then becomes to build a system that follows a set of rules with guard rails to prevent it from going off track.

I think current AI systems work well when the problem statement is narrow. The system can be fine tuned to hell for such problem statements. So, in the above example, I can spend a lot of time to fine tune the appointment booking agent and have a working app ready to go. But I can’t expect that agent to book a flight ticket for me.

It is not easy to build AI apps. The ones that work well, do so inspite of the AI, not because of it. So it becomes important to understand how AI apps work and how you can make them work to your advantage. AI is a new tool. We don’t have to solve all our problems again because we have a new tool. There is still value to learning to write well, to code, to create art. AI can’t compete. 

In the community, we wrapped up the discussion with a very wise quip from one of the members.

AI can help you write for other people. But you are the only one who can write for yourself.