Training Data and its Impact on Cultural Understanding & Ethical Values in Giant AI Models
summary
Aleph Alpha built its own data center (ranked Place 72 in Europe) and trains models across five languages, claiming much stronger German/French/Spanish results than GPT‑3 and surprising Bavarian‑dialect completions; GPUs break regularly but training runs succeed. The models do translation (Angela Merkel → Hamburg), culturalized answers (soccer field vs NFL), a steerable conversational wrapper, and multi‑modality that reads images, OCRs signs, explains jokes and infers unstated events (boy drowned), highlighting dataset‑dependent values and accessibility tradeoffs.
AI-written summary · about
chapters (11)
- 0:00AI and cultural values
- 1:02Hand-built to AI-driven systems
- 3:09European AI infrastructure
- 4:34World models and language prediction
- 6:21Training data shapes AI
- 8:49Cultural context in AI
- 10:36Ethics and bias in AI
- 12:23AI knowledge amplifier limits
- 13:50AI interpretation nuances
- 15:34Multimodal AI text and images
- 17:38AI humor and context
related talks
02h 50mUnderstanding and Using AI Language Models
01h 03mIt's All About AI
30mAI Application Architecture: Building for Scale and Efficiency
48mDeploy large language models responsibly with Azure AI
52mBeyond the AI Hype: What's Real, What's Next
51mOpenAI and the AI that may overtake humans
06mThe Dark Side of AI - Environmental & Social Costs Developers Must Know
28mAI’s Next Great Challenge: How to create AI that truly understands us?