Large Language Models

Introduction If you’ve been working actively in AI like me, either in industry or research, you’ve certainly noticed how hard it’s become to keep pace with discoveries and advancements, especially in the last two years. From DeepSeek-R1 1 to the Hierarchical Reasoning Model 2, from AGI promises to language limits 3, and from simple chatbots to complex agentic workflows, being up-to-date has become overwhelming. At least this has been my idea so far, until I realized one simple thing: it is just impossible 🤒....

Introduction If you’ve been following the evolution of API development and AI integration, you’ve likely encountered two important but fundamentally different approaches: OpenAPI and MCP (Model Context Protocol) 🤖. While they might seem similar at first glance—both deal with connecting systems and enabling integrations—they serve very different purposes in today’s tech landscape. OpenAPI vs MCP What is OpenAPI? 📋 OpenAPI (formerly known as Swagger) is a specification for describing REST APIs that defines a standard, programming language-agnostic interface description for HTTP APIs....

Introduction When working with Large Language Models, we often focus on their remarkable capabilities - from writing code to explaining complex concepts. However, there’s a crucial component that can significantly impact their behavior and performance: tokenization 🍣. As highlighted in a recent work by Garreth Lee and the Hugging Face team 🤗 1, even state-of-the-art models can stumble on seemingly simple tasks due to tokenization choices. For instance, many models struggle with the basic question “Which is bigger?...

Introduction Recently my PhD supervisor called me, asking: “Would you like to come to one of my lectures and present the students a use-case of yours? You choose!”. Of course, as the average infamous PhD experience with supervisors, I could not say no 😂 Jokes aside, it has been not the first time I hold a lecture at the University of Florence about Natural Language Processing, but this time I wanted to talk about something kind of new: retrieval augmented generation (RAG)....