A multi-fold approach to AI

When I first came across the term "prompt engineering", my judgement was way out of mark. I equated it to having good human language skills. Why would someone need engineering for asking questions using good grammar and getting anwers in a chat window? But that view has changed, as I have finally begun on the path of developing some serious expertise around AI and ML.

Better late then never!

My relationship with AI/ML domain is not new. I remember attending a mini-bootcamp 8 years ago, where I was first introduced to the ML concepts and I was fascinated by it. But that was also the time I was heavily divested into picking up infrastructure and automation skills - another fascination back then. Over the period, while I gained expertise in infrastructure and automation, I could not do justice to learning AI/ML as there were no real "opportunities" to cater to. The ecosystem was small, and there was only so much I could build. However, every now and then, especially after Nov of 2022, I have been in light touch with this domain, and always thought there would be a day when I would conquer this domain. I think, the time has come.

A lot has happened in the last few years in the field of AI/ML. A couple of months ago when I started looking for a "real" starting point, I was left completely overwhelmed and confused looking that the amount of resources, books, papers, references, and content around this topic. What bogged me further was the sheer amount of influencers providing a spectrum of perspectives and approaches about "getting started" with confidence. When I decided to stick to one approach, there was always a contradicting approach standing around the corner which made equal sense.

So I decided to keep things simple.

As mentioned earlier, I have been using these tools to transform my own job and life, mainly as chat inference and automation, and to build apps like Budge. Given the sheer amount of entry points, I thought why not start with the outermost layer and something that I use effectively on a day to day basis - the chat and LLMs. Going deeper this way made sense to me, as I think I can "figure out" what may happen under the hood.

However, I soon realised I was far from that understanding. Plus I discovered aspects like production deployment, data retrieval, agents and their orchestration, etc. which actually help make business value to teams. Of course we always have the academic angle where we learn about developing the model from scratch. But at this point and time, I am more interested in the applied side a.k.a. getting things done side, rather than being a professor of how do develop any model. So that helps reduce my focus area and my anxiety. May be, I could keep those topics for leisure reading like I always do.

Another point that made more sense to not go the model development route is - it takes a lot of data, time, and energy for training a new model from scratch. Neither do I have those resources, nor the compute power to make this happen. From all the genuine sources I have read so far, it is always advised to fine tune currently built foundational models to make them lean for our own use case.

So am I really keeping it simple?

For now I have a very clear broad sense of direction on what NOT to do. Having said that, the scope otherwise is already quite huge. At the time of writing this, I am finishing a book on prompt engineering, taking the LLM Course from Huggingface, and next step planned is to understand transformers. There are quite a few questions and curiosities on my mind while I do this, but I have carefully sequenced them to be quenched, in the coming weeks. This may not be the simplest approach, but I will get there and come back with a report on "what would I do to learn AI/ML in 20xx year". ;)

Being an advocate for hands-on learning, I did have experience with prompt engineering while building Budge, and another app which is work in progress. In fact, that is the exact reason where I faced the wall and thought about several use cases I could implement in these apps beyond prompt engineering using RAG techniques, support agents, agent orchestration, multi-modal architectures, etc.

If you would like to take a look, here is a reference I built to build prompt layering engine in any app that uses AI. Feel free to check it out.