The Harsh Reality of Small-Scale AI Research: A Personal Tale of Frustration and Limitation

In the fast-paced world of artificial intelligence, where tech giants dominate headlines with their massive language models, there exists a less glamorous reality for individual researchers and small teams. This is my story — a tale of dedication, hope, and ultimately, frustration.

For the past week, I poured my heart and soul into creating a dataset. Hours upon hours of meticulous work, carefully curating and preparing data that I hoped would push the boundaries of AI in my own small way. With my dataset ready, I dove into the next phase: model training.

Three days. Twelve precious hours of compute time. Each day, I dedicated four hours to fine-tuning and training, optimizing every parameter I could. The anticipation built with each epoch. What breakthroughs might I achieve? What new insights could my model uncover?

But as the final training run completed, the harsh reality set in. Despite all my efforts, I found myself stuck in a limbo of limitations. My GPU, with its 12GB of memory, simply wasn’t enough to handle the models I aspired to create. The 7B and 8B parameter models — the current “small” standard in the field — remained frustratingly out of reach.