Is anyone actually training AI models from scratch anymore?

When I was a third-year student, I built a chatbot with an Amharic language option using TF-IDF vectorization, and later trained a CNN to detect coffee plant diseases. I loved getting my hands dirty with data and training loops. Fast forward to today, and building AI usually just means plugging in a third-party API. It made me wonder if it is still worth training your own models, or if we should just let APIs do all the heavy lifting.

I believe custom training is still a true superpower for a few key reasons.

First, there is a major localization gap. Massive global models still struggle with low-resource languages like Amharic. Custom tokenization and local embeddings often give much better, more accurate results for local contexts.

Second, there is the edge advantage. My coffee disease CNN was built for real-world agricultural use. In remote farming areas with zero internet connectivity, a cloud API is completely useless. A custom, lightweight model can run fully offline on a standard smartphone.

Finally, the smartest approach today is to hit the hybrid sweet spot. It does not have to be API versus custom. You can use APIs to prototype fast, but train or fine-tune smaller, open-source models to save money, protect data privacy, and master niche tasks.

APIs are great for quick, general fixes. However, if you are solving deep, localized, or offline problems, knowing how to train your own model is what sets you apart.

What are your thoughts on this? Are you team API-first, or are you still training your own models? Let us discuss in the comments.

#AI #MachineLearning #SoftwareEngineering #TechInAfrica #DataScience