Bilge Yücel explains that true AI sovereignty is a technical challenge, not just a policy issue, built on four pillars: data, model, infrastructure, and operations. This talk explores the common pitfalls of retrofitting sovereignty—from performance regressions after swapping models to discovering deep vendor lock-in when moving on-prem—and presents a checklist for building genuinely sovereign systems.
Spotify is personalizing open-weight LLMs without full fine-tuning by combining three key components: foundational user embeddings from streaming history, 'Semantic IDs' that tokenize its 100M+ item catalog, and a 'soft tokenization' layer that projects a user's embedding directly into the LLM's context. This allows the model to autoregressively generate the next song or podcast as the next token in a sequence.
Erik Hanchett from AWS demonstrates how to accelerate the development of full-stack applications using AWS Amplify Gen 2. The talk focuses on its TypeScript-first approach for end-to-end type safety, its abstraction over the AWS CDK, and the integration of generative AI features through the Amplify AI Kit to connect with services like Amazon Bedrock.
Jacob Miller and Jeremy Mumford, authors of 'Architected Intelligence', discuss the enduring principles for building successful AI products and organizations. They cover why velocity is the only durable moat, why hallucinations are a data curation issue, and the proper progression from skills to workflows to agents, emphasizing a shift from focusing on models to focusing on process and speed.
Guillaume Vernade from Google DeepMind demonstrates a full generative media pipeline, using Gemini to read a public domain book and act as a master prompt engineer for other models. Imagen generates character portraits, Veo animates scenes into video, Lyria composes a unique soundtrack for each chapter, and a clever TTS trick creates a multi-character audiobook.
Explore advanced techniques for building long-running AI agents, moving beyond simple loops. Learn why self-evaluation fails and adversarial evaluators succeed, how to manage context with structured handoffs instead of just compaction, and how to use negotiated 'sprint contracts' and detailed rubrics to build and test complex, full-stack applications autonomously.