Descriptions:
The Cognitive Revolution podcast features Marek Kozlowski, head of the AI lab at Poland’s National Information Processing Institute, discussing Project PLUM — Polish Large Language Models — a national initiative building sovereign AI infrastructure for a country of 38 million that cannot compete with US or Chinese frontier labs on raw compute or data scale.
Kozlowski explains why frontier models trained predominantly on English and Chinese data increasingly fall short for Polish speakers, and why the gap is widening as labs prioritize coding benchmarks over multilingual support. His team’s approach layers language adaptation techniques on top of Llama and Mistral base models, supported by several hundred billion carefully curated Polish-language tokens — well below the trillion-token scale of frontier training runs, but sufficient when combined with targeted domain fine-tuning. EU regulations preventing large-scale web scrapes force a more deliberate curation strategy compared to what American labs can do.
The conversation explores why government agencies and national enterprises are often better served by smaller, locally-controlled models fine-tuned for specific tasks than by massive generalist models accessed via foreign cloud APIs — offering data privacy, cost advantages, and political sovereignty that global providers cannot guarantee. Kozlowski also addresses the strategic realism behind Poland’s approach: not trying to beat OpenAI or Google, but securing a meaningful, trustworthy AI capability that serves Polish language, culture, and institutional needs on Poland’s own terms.
📺 Source: Cognitive Revolution “How AI Changes Everything” · Published December 06, 2025
🏷️ Format: Interview
