AI Perspectives: How Humans Adapt and Language Models Work

Hello PyData People!
We are excited to announce our next event of 2026! This time, we will be hosted at Shopfully's office in Milan for an evening dedicated to AI, exploring how AI agents are reshaping the way we work and the evolving role of humans in AI-driven systems, as well as uncovering the inner workings of large language models and how they process and interpret language.
***
📅 When: Wednesday, October 21st, 2026, from 18:30 to 21:00
📍 Where: Shopfully, Via Giosuè Borsi 9, 20143 Milan
⚠️ Important: Spots are limited. Please keep your RSVP updated to allow others to participate if you can no longer attend.
***
🕒 Agenda

• 18:30: 🚪 Doors open & Check-in
• 19:00: 🎤 Talk 1: From Owning Tasks to Owning Outcomes – Raffaele Bongo
• 19:45: 🎤 Talk 2: Garden Path Sentences in LLMs – Luca Mattiazzi
• 20:30: 🍕 Networking & Aperitivo offered by Shopfully!

***
🎤 The Talks
1️⃣ From Owning Tasks to Owning Outcomes
How AI agents changed the way we work and our unique value as human
Speaker: Raffaele Bongo (Senior AI Engingeer @ Shopfully)
When AI arrived, everyone got promoted: from executing tasks to supervising the AI that executes them. As agents get better and cheaper, they produce more output than people can review. Research finds that overseeing agents is the most mentally tiring way of working with AI, and humans become the bottleneck of the systems they run.
This talk argues for a second promotion, from supervisor to orchestrator. That means changing how we define doing well, from "the task I checked is correct" to "the system I own produces good outcomes". It covers the people questions this raises: what stays uniquely human, who owns what, and how we keep growing seniors when junior work goes to agents. It then presents pillars for designing agent systems people can own.
About the Speaker:
Raffaele is a Senior AI Engineer at Shopfully. Raffaele has an extensive experience in startups and scale-ups, where he has specialized in building and deploying machine learning models and data pipelines on cloud platforms. Originally, Raffaele signed up for mechanical engineering but switched to computer science last minute because AI sounded cooler.
***
2️⃣ Garden Path Sentences in LLMs
Diffusion of language models human migrations
Speaker: Luca Mattiazzi (AI Engineer @ Pillar)
Garden-path sentences are sentences that initially lead readers toward one interpretation, only to require a reinterpretation as they continue reading. A classic example is "The old man the boat", where "man" initially appears to be a noun but is actually a verb.
In this talk, Luca will show us how transformer models process these sentences compared to similar, non-ambiguous ones. Using mechanistic interpretability techniques, he'll explore attention patterns and layer-level changes across Gemma 3 models of different sizes (1B, 4B, 12B, and 27B).
He'll share which parts of the models respond most strongly to syntactic reanalysis, how these responses evolve with model scale, and what these findings reveal about how LLMs handle linguistic ambiguity.
About the Speaker:
Luca has a degree in Biotechnology and has been working as a developer since 2016. Currently an AI Engineer at Pillar, he's passionate about data visualization, machine learning, and building things. He enjoys experimenting with new ideas and working on side projects, even if, as he likes to say, most of them never quite get finished.
***
See you there! 🚀
The PyData Milano Team