The Filipino qualities we need to build world-class language technologies

Last Saturday, I gave a talk at the Analytics & AI Association of the Philippines (AAP) on the topic of Philippine language model evaluation, specifically on FilBench.1 The purpose of the meetup was really to figure out (1) whether we can build a “Philippine LLM” and (2) what doing so would entail. I had some experience building open language models back in my previous work, and my research on multilinguality is directly related, so I was able to share my thoughts and experiences. To preface: the answer to the question of “can we actually build Filipino-centric LLMs?” is definitely YES. The tools are open-source, the recipes are publicly available, and there’s definitely a lot of talent. However, the challenges are two-fold: Building Filipino-centric LLMs does not entail applying Silicon Valley approaches to our local contexts. The capital difference is vast and the ecosystem is different. It’s better to look into how our neighbors (geographically and economically) are doing it, such…