Ecologies and Economics of Language AI in Practice

The presentation titled Ecologies and Economics of Language AI in Practice by Jade Abbott explores the challenges and frameworks for building language models under constraints in non-Western environments, particularly in Africa. The talk reflects on:

Introduction

• Introduction of Jade Abbott as CTO of Lelapa.AI and co-founder of Masakhane.
• Focus on language technology development for African languages under significant data constraints.

Main Themes

• Sustainability Concerns: The impact of global language model development on ecological and cultural sustainability, and the commoditization threats posed by Big Tech.
• Data and Linguistic Justice: The challenge of creating AI tools in a landscape with over 2000 African languages and the issue of linguistic justice due to economic drivers influencing tool development.
• Resource Constraints: Insights into building smaller, efficient language models that suit specific needs without diminishing performance despite limited resources.
• Economic Models: The need for sustainable business models over reliance on international funding for African NGOs, emphasizing investment as a sustainable approach.

Technical Insights

• Utilization of data as a raw material likened to minerals, advocating for a shift from extraction to data creation tailored for specific populations.
• Introduction of the AETU framework for sustainable governance in low-resource languages, promoting data creation and shared ownership as central strategies.
• Efforts in model distillation and the use of smaller models that maintain efficiency while enhancing economic and ecological viability.

Challenges and Reflections

• Issues with translating oral languages into written forms and questioning the necessity of textual representation in predominantly oral cultures.
• The role of AI in perpetuating hegemonic views and the importance of preserving cultural diversity through language.

Conclusion

• An emphasis on rethinking AI development to prioritize sustainability, cultural relevance, and fairness.
• The importance of smaller, more efficient models to improve accessibility and sustainability in the development of AI technologies.

The presentation concludes with a call for action towards creating human-centered and sustainable language AI solutions that respect environmental and cultural contexts.

This is the end of the AI-generated content.

Lessons learned from building language models in Africa: under strict data constraints in non-western environments.

The global proliferation of language models has predominantly followed a Western-centric approach that neglects cultural and environmental sustainability. This talk draws on firsthand experiences building language models in Africa under significant data constraints to challenge prevailing assumptions in AI development, while demonstrating how Big Tech's commoditization threatens both ecological and cultural sustainability.

Through case studies and practical examples, this presentation will show that the "bigger is better" paradigm is reaching diminishing returns while providing attendees with actionable frameworks for creating models that respect both environmental limitations and cultural contexts. Participants will gain practical insights into developing language AI that is more effective, economical, and sustainable, offering a roadmap for implementing human-centered AI solutions that better serve diverse global communities.