DataKrk – Scheduling at Scale & Fragile AI Alignment

After a short pause, we’re rebooting Kraków’s engineering meet-ups with an evening devoted to Kubernetes scheduling and AI safety. Join us on Thursday, 10 July 2025, at 18:00 in the Tech Talks Space at the Google Office (Rynek Główny 12). Expect two deep-dive sessions, cold refreshments and plenty of hallway chat with fellow cloud-and-ML enthusiasts.

The first half of the night belongs to three engineers from Google’s Kraków office.

First, Michal Wysokinski will talk about AI Training at Scale: Industry Patterns and Trends. Training massive AI models often feels more like alchemy than engineering, especially when dealing with complex infrastructure and tight budgets. This session focuses on the practical side: the tools, techniques, and strategies needed to build efficient, performant, and manageable large-scale training pipelines. We'll move past the buzzwords and get into specific engineering challenges and solutions. We’ll dive into the crucial role of Kubernetes, explore the nuances of topology-aware scheduling and high-speed interconnects, discuss real-world resource optimization techniques, and delve into the politics and practicalities of fair resource sharing in large organizations. Expect actionable strategies, not just hype.

Then, Ania Borowiec (Software Engineer, GKE AI Training) and Maciej Skoczeń (Software Engineer & SIG-Scheduling Chair) will give an “Overview of Scheduling in Kubernetes.” In just thirty minutes they will unpack how the kube-scheduler decides where every Pod—and increasingly, entire AI training jobs—should run, even in clusters scaling past 65 000 nodes. You’ll hear what the open-source scheduling team is building next and why those ideas matter for anyone running large-scale workloads.

Right after the Q&A with the Google team, we’ll shift gears to AI alignment. Przemek Maciołek —currently a Staff Data Engineer at Airbnb, formerly a founder, architect, VP Engineering and ML/NLP researcher - will argue that “AI Alignment Is More Fragile Than You Think.” He’ll walk through experiments showing how a handful of malicious examples can derail a large language model’s moral compass via lightweight fine-tuning. A live demo will set the stage for an open discussion on what these findings mean for practitioners and end-users alike.

### Logistics

The entrance is through the main reception of Google Kraków—ask for the Tech Talks Space on Level 2. All talks and discussions will be held in English. We provide air-conditioning, refreshments and beverages. Attendance is free, but seats are limited, so please register early.

### Speakers

• Michal Wysokinski is a Software Engineering Manager at Google, where he leads the GKE AI Training OSS team. He's currently immersed in the complex challenge of ensuring Kubernetes is not just a place, but the definitive best place to run large-scale AI training workloads.
• Ania Borowiec has been a software engineer at Google since 2011; today she focuses on improving kube-scheduler for both GKE and the broader OSS community.
• Maciej Skoczeń is a software engineer at Google and co-chair of SIG-Scheduling, driving the roadmap for Kubernetes’ core scheduling engine.
• Przemek Maciołek is a Data Engineer at Airbnb with a PhD in ML/NLP, a track record of scaling start-ups and a knack for turning napkin sketches into production systems ten-times faster than you thought possible.

Important: The venue has a limit for 25 attendees, so first come, first serve.

See you at Google!