Lutz Roeder

Meta researchers introduced an architecture that enhances LLMs with trainable memory layers — components that efficiently store and retrieve relevant factual information without requiring a massive increase in computation. By structuring memory keys as combinations of smaller “half-keys,” the team scaled memory capacity dramatically while maintaining efficiency. In tests, LLMs equipped with these memory layers outperformed unmodified counterparts on several question-answering benchmarks despite being trained on significantly less data. Learn more in The Batch: https://hubs.la/Q03n-jkg0

3

From Unstable Library to Production-Ready Memory: Solving Graphiti's Core Flaw -- Excited to share my latest deep-dive on a critical challenge we encountered while building advanced AI agents: managing complex, stateful memory systems in production. The `graphiti-core` library offers a powerful way to build temporal knowledge graphs, a huge step up from vector stores. However, as I detailed in my previuos article, its internal async handling can cause critical event loop conflicts when integrated directly into frameworks like Google's Agent Development Kit (ADK). My new article presents the solution we engineered: a decoupled, production-ready FastAPI microservice that wraps `graphiti-core`. This architecture completely isolates the library, exposing its power through a stable, scalable, and maintainable HTTP API. This is a must-read for any developer looking to build robust, production-grade AI agents with sophisticated knowledge graph memory. Key Takeaways: - Understanding the hidden dangers of event loop conflicts in async libraries. - The architectural advantages of a decoupled microservice for AI systems (stability, scalability, maintainability). - Complete, production-ready code for the FastAPI service and its integration with Google ADK. You can read the full technical walkthrough here: https://lnkd.in/eHcFPVBv Links to my previous article: https://lnkd.in/ew-VNVM7 #AI #ArtificialIntelligence #LLM #AgenticAI #SoftwareArchitecture #Python #FastAPI #KnowledgeGraph #GoogleADK #Graphiti

9

Just read the transcript of the NVIDIA earnings call, and the last question might’ve been the most insightful. Goldman asked Jensen if he still feels good about growth accelerating. It turned into a lengthy answer from Jensen laying out why he believes growth is still accelerating. His argument is that AI isn’t a small upgrade, it’s a step change. If tokens become the unit of value, demand for compute doesn’t grow linearly, it expands massively. We used to write code, compile it, then run it. Now software is generated in real time, adapts continuously, and improves through ongoing inference. That’s not iteration, it's a fundamentally different model, and it requires a lot more compute. It’s a shift from pre written software to generated software. From static execution to continuous inference. His rationale, was this isn't a cycle, it's the new standard for computing. Cloud companies are shifting entirely to this token-based architecture. Curious how others are seeing this play out, especially inside enterprise environments where this shift is already underway.

10

2 Comments

Want to deepen your understanding of vector embeddings? I turned my Pycon poster in a blog post: "A visual introduction to vector embeddings" https://lnkd.in/gCEuiSjw It compares three models (word2vec, ada-002, text-embedding-3-small), four different similarity/distance metrics, and two compression strategies (quantization/MRL).

102

4 Comments