The inner workings of LLMs and AI Agents

• Philip Soerensen (University of Padova / INFN Padova)

Room: 1/3-1 - Sala R In this journal club, will attempt to explain how an AI agent works. The goal is that by the end of the session, you have a rough idea of what practically happens between the time when your tell your agent to "write me a physics paper", and the time when the .tex, with a complete (but certainly garbage) paper, shows up on your hard disk. In the interest of time and limited knowledge, I will not try to discuss *why* it works like this (the underlying theory) or *how it is built* (the training), which will be topics for another day. I will, however, try to cover the practical way that important concepts like "attention", "context", "reasoning", and "temperature" arise, so that you have an intuitive idea of why they matter within the world of agentic coding. The plan: I will talk about both large language models (LLM's, e.g. ChatGPT, Claude, Mistral) and the harnesses that are used to turn them into agents. I will attempt to explain the general architecture behind LLM's, the transformer, in (just) enough detail that it becomes reasonable that it actually works. Then, by evesdropping on the actual process, I will show how a harness (Codex, Claude Code, Mistral Vibe, etc.) turns the pure text generator into an agent capable of interacting with the (digital) world around it. Throughout the session, I will make use of a locally hosted LLM to demonstrate a few features that are less transparent in online chat.

Tutorial: Local LLMs

• Philip Soerensen (University of Padova / INFN Padova)

Room: 1/3-1 - Sala R After the main 1 hour session is finished, and the majority of you have probably left, I will stay for another half our to show those of you who are interested how you can install your own completely offline language model and agent, and demonstrate a bit what can be achieved with limited hardware like an offline laptop.