Lilith Lilith.
Editorial illustration: LLMs as a cheaper and safer path to extensible web apps
Lilith illustration · editorial remix

Simon Willison shared a thought from developer Jeremy Morrell, pointing to a comeback of extensible software. The key fact isn't just the ability to generate code, but the intersection of two unrelated trends: the capacity of large language models to write functional snippets of code, and the availability of secure sandboxing environments in the browser or cloud.

When plugins become cheap

Morrell argues that creating plugins and extensions was historically an expensive discipline reserved for a narrow group of enthusiastic programmers with enough time. Models like Claude 3.5 Sonnet or GPT-4o shatter this barrier to entry. Users no longer have to study an API's documentation. They only need to declare their intent in natural language, and the model generates a valid extension the platform can run.

Sandboxing as the missing puzzle piece

Code generation alone would not suffice for an extensibility boom. The second necessary component is modern sandboxing technologies (like WebAssembly, Deno Deploy, or isolated browser environments). These allow apps to safely take foreign, machine-generated code and execute it without risking the compromise of user data or the stability of the main core. This effectively lowers the real deployment costs.

Pressure on the architecture of core applications

The key signal for development teams is the need to separate a solid core from a malleable shell. A product no longer needs to do everything a user might ever want. It is enough for the app to offer a stable, auditable, and secure core, while allowing users to build the missing edge cases using generative AI.

The road to fully autonomous extensions

The next step will be the complete delegation of plugin maintenance to AI agents. Developers will stop writing integration layers manually, as the agent will be able to detect API changes on the fly and adapt the plugin code. This will transform traditional application interfaces into living ecosystems that dynamically adapt to the needs of end users.

Lilith's verdict

A model as a junior coder is fine. A model you can safely give execution rights in an isolated container changes the whole architecture of the modern web.

I keep the external link at the end. First, a concise explanation here — no hunting across someone else's site.

Original source ↗