The post says LLMs are unusually good at making “software for one” and argues that the next platform shift is not just code generation, but software built from the start to be extended by AI-written plugins. The proposed shape is web-first, sandboxed, and capability-based. Generated code should only be able to do what explicit references allow, rather than getting broad ambient access to APIs, data, or credentials. The pitch is that this could let normal users create and safely run custom workflows without becoming developers or enterprise buyers.
The useful signal is that people did not really dispute the demand for custom software. Several gave firsthand examples of already using
LLM-built personal tools or seeing clients arrive with vibe-coded prototypes. What they challenged was the delivery model. A lot of the skepticism landed on security and maintenance. Sandboxes look good until the app needs sharing, access control, external endpoints, or browser networking rules that leak around content security policy. Once you let generated code handle real data for small groups, the problem stops being “can the model write the feature” and becomes “can the platform express permissions and data boundaries clearly enough that mistakes are survivable.”
The other strong theme was that LLMs do not erase software architecture. People who know older, smaller toolchains can still get tighter results than someone defaulting to
Electron and a mountain of dependencies. Others pushed the opposite direction and said cross-device access is exactly why the web still wins, especially for family or small-team software. Put together, the conversation points to a narrower but believable near-term future: not infinite end-user programmability, and not fully automated product development, but platforms where users can spin up custom apps or extensions inside a heavily constrained host that owns auth, storage, sharing, and rollback.
jeremymorrell.dev