1. The Basics Matter More than I Think
When I first watched a few YouTube videos about how LLMs work, it was purely out of curiosity. People kept calling them a “black box,” and I wanted to see why. I expected complex math and scientific jargon—and yes, those did appear, and I probably understood less than 5% of them. However, I was pleasantly surprised to find that amidst the complexity, I began forming a rough framework of key concepts such as transformers, embeddings, tokens, vectors, context windows, and fine-tuning. This basic understanding became invaluable later, as it made absorbing other LLM-related content much easier. The dots started to connect, and I could better grasp how these models function. You don’t need to dive too deep; even a few short videos can provide the foundational building blocks you’ll rely on in the future.
2. Try Low-Code Instead of No-Code
If you have zero coding background, this step is likely to offer the biggest return on investment in the long run. I once searched for no-code solutions—there are many—but they never fully convinced me. On a friend’s recommendation, I took a basic Python course, investing no more than 20 hours total. I still can’t code an entire app from scratch, but that small investment has helped me "decode" and understand the structure behind the code. It also opened the door to other applications like deploying open-source local models, experimenting with APIs, and trying more advanced agentic applications—all without waiting for someone else to build me a nice user interface. In addition, the more I explored, the more I realized there are development tools and services you can’t avoid if you are serious about building useful applications, such as Docker, AWS, Hugging Face, Google Colab, GitHub, or Azure. Having at least a little knowledge of coding can significantly reduce your level of frustration. And don't forget that modern LLMs can actually write or debug code for you. You can make the models work better for you if you at least know how to put them together.
3. See Beyond the Cool Demos
LLMs can do almost anything on paper, but ultimately, our goal is to solve real business problems. Quite often, the gap between the demo and real-world application is much wider than estimated. Take the popular “RAG” approach as an example. Tutorials often showcase a handful of PDFs containing well-known content—such as widely known books or publicly available guides—which the LLM already “knows” from training. In reality, you might have thousands of domain-specific documents the model has never seen. Simply uploading (embedding) these documents isn’t enough. Data cleaning, indexing, and optimization are critical steps for achieving commercial-grade performance. These processes demand time, expertise, and domain knowledge. Additionally, the success of your RAG system depends on multiple factors. Database architecture plays a crucial role, whether it involves vector databases like Pinecone, traditional SQL, hybrid approaches, or even graph databases. Embedding methods—such as chunking, sliding window approaches, or sentence-level embeddings—are equally essential. Moreover, the choice of LLM is vital. Factors like parameter size, quantization, fine-tuning, prompt engineering, model temperature settings, and the quality of fine-tuning data all significantly influence the system’s performance. Without a clear understanding of the real problem to solve, it’s easy to get lost in technical complexities and fail to address the actual issue. Your business needs should guide the applications. Sometimes, a well-organized file folder or a clear table of contents might be all you need to achieve your goals!
I’d love to hear about your high-ROI discoveries as well. Feel free to share them in the comments below!