Simon Willison's curation of this hypothesis points to a broader industry conversation about shifting software development's center of gravity. The core idea—that AI can fill in missing pieces for safe user extensions—is attractive but arguably glosses over the hard problems of software design. In our view, the bottleneck may simply shift from writing code to defining clear, reliable interfaces and guardrails that an LLM can understand.
The security promise of sandboxing is real, but it doesn't solve the challenge of ensuring that AI-authored extensions are coherent, maintainable, and truly integrate with an app's core logic over time.