From Code to Prompts: My Journey Across the AI Divide

As I have been playing around with agentic AI, I can’t help but think of how much more accessible AI has become not just for consumers, but for producers of applications as well. Given the multimodal and multilingual nature of tools like ChatGPT, one can pick a natural language of their choice to program GPTs and Agents to automate a variety of tasks. It’s no exaggeration to suggest that natural language—especially English—is becoming the dominant “programming language” of the future.

Learning ML the Hard Way

There is now a clear demarcation between pre- and post-ChatGPT world when it comes to learning, coding, and generating content and the shift has been nothing short of revolutionary. Reflecting on my personal experience, an interesting parallel is how I ramped up on Machine Learning (ML). It wasn’t too far back. Just about 8 or 9 years ago.

To develop the right intuition for ML, I had to get reacquainted with Matrix formulations and transformations, brush up on Optimization, refresh my knowledge of partial differential equations etc. This is besides coding in R, Python, and MATLAB. Then I moved onto using higher level abstractions like TensorFlow and Keras which made building complicated neural networks simpler. With just a few lines of code and invoking various libraries, I could build a Neural Net. But programming language was still abstract and every now and then I got stuck. I had to get on stackoverflow and post code snippets in the hopes that someone would answer my queries (which they did in some instances). The learning was slow with fits and stops. I wished I had an omniscient, omnipresent tutor with me whenever I got stuck!

The Simplicity of Agentic AI Today

Fast forward to today, now “programming” GPTs in plain English and building custom agents has become so simple. No complicated math. No programming skills needed. With just a few natural language prompts, you’re off to the races. As the bar for learning lowers, AI is becoming more democratized, opening doors for non-technical and non-geeks to emerge as programmers and tech producers. We can now use the likes of ChatGPT not just to build apps but even create test cases for these apps using the same GPT that created the app in the first place. Somewhat of a lazy builder’s dream come true!

The Other Side of the Coin: Hallucinations and Risks

The other side of the story is that these agents are highly adaptive. As they say hallucination is not a bug but a feature. What is hallucination in one instance is creativity in another. Given the outputs of the GPTs are non-deterministic, our vigilance in building, testing, and securing the agents will have to be up by several notches. We have to be thoughtful in designing GPT based apps thinking of all possible misuses of the agents we create and biases they can introduce. Else, we will expose ourselves and our organizations to bigger risks. Look no further than the recent story of Chicago-Sun Times publishing an AI generated Summer Reading List of books that did not exist, which became a major embarrassment for the publication. When it comes to Agentic AI, my current mode is “Trust but verify”. Human in the loop is integral to building trust.

ML and Optimization Still Matter... More than Ever

Where does this place techniques like ML and Optimization? They live on, and for example, in the world of supply chains, these mathematical techniques are more relevant than ever given the increasing intensity and pace of disruptions. A supply chain is a deeply interconnected machine that requires sophisticated math to decide on what, where, when, and how much to make, move or source for example. The agents will bring natural language and multimodal interactions so we can extract the intelligence buried deep within these mathematical algorithms that are currently the realm of sophisticated supply chain engineers and business analysts, making them accessible to a broader range of users, including executives. We’re just beginning to unlock the true potential. We are still on book 1, page 1. The future of agentic AI is exciting but cautions I issued earlier need to be considerations.

Reskilling and Human Differentiation

Of course these tools change the nature of human work. In the pre-ChatGPT world, a good chunk of my time was spent on studying the world events, connecting the dots with my knowledge of the supply chains, apply reasoning and synthesizing various recommendations and hypotheses. Everything I said here is what ChatGPT now is exceedingly good at and in fact surpassing me in some ways. So where does that leave me? I’m actively reskilling—learning to use these tools more effectively so I can automate parts of what I do (promise you won’t tell my boss 😉). My differentiation now comes from my ability to draw on my decades of experience that no GPT is trained on (not yet any way!), the networks and friendships I built over time and the ideas and intelligence I can tap into that is trapped in the heads of these outstanding individuals. Last but not the least, it allows me to spend more “human time” with my customers, partners, and colleagues solving problems that deserve to be solved, with some heavy assist from AI of course!

Recommendations to Learn Agentic AI

If you’re interested in sharpening your skills in Agentic AI, I recommend Agentic AI and AI Agents for Leaders, a 3-course specialization. No programming skills needed. If you are adventurous and are a math nerd like me looking for advanced computational AI, check out Deeplearning.ai for a variety of courses on Machine Learning. My favorite is Machine Learning Specialization.

Final Thoughts

The future looks bright and exciting. It is time to adapt. Learning is becoming easier than ever. Let us arm ourselves with some superpowers! Would love to hear of how you are reskilling.