Update one package in your container image and watch your users re-download 500MB of unchanged content. Traditional Dockerfile layers are instruction-based—a single package update invalidates an entire layer.

This talk introduces chunkah, a tool that post-processes container images into content-based layers. Files are grouped by package, not Dockerfile structure. Update one package, users download only that layer.

We'll cover:

• Why instruction-based layers hurt pull performance (with numbers)
• How content-based splitting works under the hood
• Live demo: chunkah in action, before/after comparison
• How podman history reveals package-to-layer mapping

From Project Hummingbird, where we run 70+ production container images with chunkah, we'll show real metrics on bandwidth savings and how to adopt this in your own builds.

Audience: Container image builders, CI/CD engineers, registry operators, anyone who's wondered why their image pulls are so slow.

Software engineering is undergoing its most radical shift since the move to Cloud. As AI assistants evolve from writing simple snippets to suggesting entire system architectures, the line between human-authored and machine-generated code is blurring. However, this newfound speed introduces a new breed of technical debt and ethical risk. How do we evolve our day-to-day practices without losing our grip on security, maintainability, and accountability?
This session will explore the transition from Human-Driven to AI-Driven Engineering. We will ask the hard questions and discuss practical strategies for: how to manage "synthetic debt," redefine peer reviews for AI-generated code, and ensure that 10x velocity doesn't bypass safety.
We will explore a series of community-driven practices designed to maintain quality while navigating the move from "experimental" to "operational" AI.