The hardest part of scaling AI

Peter Steinberger reflects on the real bottlenecks when scaling software development with AI.

Overview

In this clip from an OpenClaw fireside chat, Steinberger argues that while automation can take on more of the mechanical work (like testing and parts of review), the hardest constraint is still human attention and judgment.

Key points covered:

• Automation helps, but doesn’t remove the bottleneck

  • Testing and review activities can be increasingly automated.
  • The limiting factor becomes the availability of human attention.

• Taste and thoughtful design remain essential

  • Scaling output isn’t just about producing more code.
  • Quality depends on design judgment and “taste” that automation doesn’t replace.

• Closing the loop with automated verification builds trust

  • Automated verification is positioned as a way to validate AI-generated changes.
  • This feedback loop helps teams trust AI-generated code in real workflows.

• Open source AI development context

  • The clip is framed as part of a broader conversation about open source AI development on the GitHub channel.