TL;DR:
- Singapore allocates SGD70 million (USD52 million) for the development of a large language model (LLM) tailored to Southeast Asia.
- The National Multimodal LLM Programme (NMLP) is a collaborative effort involving IMDA, AI Singapore, and A* Star.
- The NMLP aims to create a 30-50 billion parameter LLM with speech and text capabilities.
- It builds upon AI Singapore’s SEA-LION model, expanding it into a multimodal speech-text model.
- Objectives include nurturing AI talent, enhancing LLM understanding, and creating a trusted AI environment.
- The initiative promotes local AI specialists and aims to establish Singapore as a global AI hub.
Main AI News:
Singapore has allocated a substantial SGD70 million (equivalent to USD52 million) to spearhead the creation of a groundbreaking large language model (LLM) tailored to embrace the rich tapestry of Southeast Asia’s diverse cultures and languages. The initiative, known as the National Multimodal LLM Programme (NMLP), is a collaborative effort between the Infocomm Media Development Authority (IMDA), AI Singapore, and A* Star, with an ambitious two-year timeline for development. The NMLP aims to give rise to an LLM boasting between 30 billion and 50 billion parameters, seamlessly integrating both speech and text capabilities.
Leveraging AI Singapore’s Southeast Asian Languages in One Network (Sea-Lion) model, an open-source LLM previously trained on 11 languages native to the region, this endeavor will expand SEA-LION into a multimodal speech-text model. The NMLP is set to achieve several critical objectives, including the cultivation of proficient AI talent within Singapore, deepening the comprehension of LLM functionality, and, notably, breaking away from the prevailing trend where LLMs are primarily developed by major tech giants in the Western and Chinese spheres. This venture aims to establish a trusted ecosystem for the utilization of AI while concurrently nurturing an AI industry focused on LLM-powered solutions.
“This national endeavor underscores Singapore’s unwavering commitment to evolve into a global AI hub. Language serves as a pivotal catalyst for collaboration,” emphasized Ong Chen Hui, the Assistant Chief Executive of the Biztech Group at IMDA. He added, “By investing in talent and harnessing the capabilities of large language AI models designed for regional languages, we aspire to foster cross-border industry cooperation and propel the next wave of AI innovation across Southeast Asia.”
In a LinkedIn comment, Laurence Liew, the Director of AI Singapore, highlighted a noteworthy statistic – a staggering 75% of the engineers involved in developing SEA-LION emerged from the AI Apprenticeship Programme, all of whom are Singaporean nationals. Many of these engineers were self-taught prior to joining AIAP, where they further honed their skills, acquired production-ready coding expertise, and mastered deployment skills. Liew underscored the significance of investing in local AI specialists, affirming, “Singapore boasts a wealth of talent.”
It’s important to acknowledge that the endeavor of training LLMs constitutes a substantial financial commitment, and the allocated USD52 million, while substantial, may represent just the tip of the iceberg. For instance, Google’s recent unveiling of the Gemini AI model comprises three distinct models of varying sizes. Nevertheless, the creation of a localized LLM model is a commendable stride forward, paving the path toward more inclusive and pertinent AI solutions that can benefit a broader spectrum of users throughout the region.
Conclusion:
Singapore’s significant investment in developing a region-specific, multilingual AI model demonstrates its commitment to becoming a global AI hub. This initiative not only fosters local AI talent but also seeks to break the dominance of Western and Chinese tech firms in LLM development. By creating a trusted ecosystem for AI, Singapore aims to drive innovation and provide more inclusive AI solutions for the Southeast Asian market, positioning itself as a key player in the evolving AI landscape.
