[Keynote #2] Architectures of Ambiguity Mapping the Technical Hurdles...
Talkat PythonAsia 2026 · on Saturday, 2026-03-21· at 10:05· Teresa Yuchengco Auditorium (Main Hall)·
summary
Claims that LLMs read words but not the world, failing Filipino sarcasm about nine times out of ten, drive two concrete projects: a 10,000‑message benchmark called PhilSarcasm with triple annotations and cultural scaffolding to teach meaning flips, and LIGTAS, a toxicity benchmark that labels eight harm categories to probe off‑the‑shelf safety filters. Results show cultural distortion from western training data and safety rules misfiring; recommended actions are localized evaluation, honest benchmark scope, and dataset sharing and tooling for regional languages.
AI-written summary · about
chapters (8)
- 0:00Nuance gap in LLMs
- 3:18Cultural context in AI responses
- 10:10Python developers and ambiguity
- 12:22Four architectures of ambiguity
- 16:19Phil Sarcasm and LIGTAS projects
- 23:58Improving inclusivity in Python
- 28:17Culturally grounded AI systems
- 30:45Q&A
related talks
50mAI: A Tool for Software Architecture?
33mOpening Keynote: LLMs, between fantasy and reality - the poin...
54mKeynote: Can you trust your (large language) model?
47mHarnessing LLMs for Next-Level Security Analysis
21mThe Coherence Trap: Why LLMs Feel Smart (But Aren’t Thinking)
01h 49mAI engineering with open access LLMs that lie, curse, and steal
55mLessons Learned From Implementing Generative AI Solutions
01h 00mAI Safety through Programming?