Building LLMs with Cultural Context
summary
Argues that Southeast Asia’s AI gap is measurable and fixable: over 99% of web training data is English, leaving 3 in 4 regional users behind; project collects 20,000+ image‑prompt audio transcripts and open‑sources them via the Aquarium beta platform so local developers can pretrain and fine‑tune models; introduces the Sea Helm leaderboard for Southeast Asian languages and uses DeepMind’s KM composition method to retain base skills while specializing — early deployments show clear performance gains.
AI-written summary · about
chapters (14)
- 0:00Introduction and context
- 1:13Southeast Asia overview
- 2:19Challenges for migrant workers
- 3:00Agriculture and small farmers
- 3:43Education gaps for children
- 4:27Why AI access matters now
- 5:52Data asymmetry in LLMs
- 7:37Collecting representative data
- 9:02Demo of Aquarium platform
- 10:05Model training techniques
- 11:28Evaluation and benchmarks
- 12:34Partnerships driving progress
- 12:54Future plans for project SEA-L
- 13:35Key lessons learned
related talks
44mKeynote: The Art of Language: Mastering Multilingual Challenges in LLMs
44mKeynote: The Art of Language: Mastering Multilingual Challenges in LLMs
33mAI is coming for your job
27mThe Cutting Edge of AI: From LLMs to Multimodal Intelligence
57mMachine Translation of Human Languages in the Age of LLMs: Is This the End of the Language Barrier?
13mAI: Big Expectations
01h 26mInclusive & Ethical NLP Design Building AI for Everyone
48mBridging the data gap between LLMs and children