A bittersweet goodbye: Why I’m moving on after 5 incredible years on Bubble

[randomanon]

@boston85719

after that, I think they will hit the wall that is the Context limitations and Memory limitations of LLMs combined with size of JSON file for an app

Yeah which is why they should just do this:

@randomanon

Lastly, and this is addressing the 100,000 pound herd of elephants in the room: I think the entire AI agent is misguided and not even technically feasible. Bubble is legos. You’re trying to combine “3d-printed” code with legos that have very specific tolerances and ways of fitting together. It’s never going to be safe to edit an entire application. This is ignoring the fact that no context window is going to large enough for some of the larger apps’ JSON files.

The AI editor needs to be an “airlock.” You click on an Action, Thing, or Element and say “Edit in Assistant.” This sends it over to a sandboxed environment where now the chat-based assistant can edit it freely without worrying about breaking the entire app. You can’t let the LLM decide where to place the code snippets in the JSON. You also can’t let the LLM have access to the entire JSON. Instead you create specific element templates with strict guardrails where the chat assistant inserts dynamic data into the static template. That way you can safely create whatever you need in the validated environment of the sandboxed assistant, and then copy/paste it directly. The user decides where it goes and if it’s good enough. This also has the nice effect of not depending on the buggy undo/redo feature to roll back multiple changes.