"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis Sovereign AI in Poland: Language Adaptation, Local Control & Cost Advantages with Marek Kozlowski
57 snips
Dec 6, 2025 
Marek Kozlowski leads Poland's AI Lab and spearheads Project PLLuM, focusing on localized language models. He discusses the importance of AI sovereignty through small, culturally adapted models to overcome biases in mainstream AI. Marek highlights challenges with existing English-centric models and the necessity of benchmarks that respect Polish language and culture. He elaborates on the principles of transparency and organic data usage in AI, while addressing the legal constraints that shape model behavior in Europe.
AI Snips
Chapters
Transcript
Episode notes
Localized Small Models Can Compete
- Poland can match larger models for local language and culture by training much smaller, language-adapted models.
- Marek Kozlowski argues these models offer control, privacy, transparency, and cost advantages over frontier models.
English/Chinese Data Dominates Performance
- Frontier models train on overwhelmingly English and Chinese data, causing language performance gaps.
- Marek explains models often map tasks to English internally, which harms niche-language fluency and cultural phrasing.
Benchmarks Miss Cultural Fluency
- Standard benchmarks bias evaluation toward extractive, multiple-choice tasks and miss cultural or long-form generation quality.
- PLuM created PLCC to measure grammar, culture, history and ambiguous language use in Polish.