Svitla Smart Talk: Tomasz Ducin - Under the Hood of LLMs and AI Agents

If you use #LLMs in your daily work, this video is for you. During this #svitlasmarttalk we explored what it truly means for a language model to be a probabilistic machine. We discussed: What LLMs are actually doing most of the time Why they require so many GPUs Where hallucinations and nondeterminism come from How embeddings and attention mechanisms work What jailbreaking is — and why it works A high-level architecture overview of agent systems Myths around agents, RAGs, MCPs, and more