I’m glad to see Bubble continuing to build out its AI features. It’s only going to get better from here.
I think it’s important to keep in mind that different platforms serve different types of builders. Someone who’s more experienced with code might find a frontend-only platform like WeWeb more appealing because it lets them piece together a “traditional” tech stack with the tools they already know. That’s totally valid.
But Bubble is a full-stack platform. You don’t need to wire up a backend or manage data in different tools because everything from database to frontend to workflows lives in one place. That makes Bubble a smart choice for business owners, solo founders, and creators who want to build serious apps without managing multiple platforms or taking the time to learn code. So different platforms appeal to different people based on their skill level and the time they have or care to have to learn a ton of different platforms just to get an app up and running.
I think that’s why Bubble AI feels a bit different than tools that only generate frontend code. It’s working with a much more complex system.
How are you even training the model being used for this? Is it trained on the .bubble JSON language? Or are you just using a frontier model with prompting tied to editor actions? That’s going to result in hallucinations every time.
I’m curious if AI will flatten the learning curve so that after getting started with it, people will have learned a lot and then could build independent of it or will it be something that without it, those who started with it, would be completely lost; like “Where’s my Coffee?!?!”
Intention with Bubble’s AI is good, however as for now it is far from useful and sorry to be honest it is hype feature. I lost several hours playing with it. While first output looks ok, modifications with agent are just not working. Not even for simple buttons. Once it will properly build frontend and backend features, it could save us time - and especially if AI agent will PROPERLY work after initial creation. Now I tested AI app for 1 day and canceled AI app plan.
I would prefer if bubble finally has design system / UI kit building blocks like most web builders. (and AI can be leveraged here even without backend), This would be time saver.
Well they started with the component library and in usual Bubble fashion, forgot about / deprioritized / tabled it.
A while back I thought to myself, “it kind of makes sense since AI will allow people to make their own infinitely customizable/extendable components” but it didn’t quite end up working out that way.
Generalised models extremely rarely need to be trained on data. Bubble doesn’t use any apps for training. It’s just a case of instructing it to output data in the way Bubble wants. Sometimes that’s Bubble JSON, other time I believe it is an intermediate language like HTML which is then programmatically converted to Bubble JSON.
If the data exists in their scraping data (i.e., the internet), yes. Otherwise, no. This is why you can’t use [frontier] LLMs for certain esoteric languages. And also why they can’t be used for certain specialized domains.
I would imagine they would never output anything lower-level than the JSON because then you end up with the same problem as vibe coding and so the “AI agent” wouldn’t actually solve anything.
But if that’s what’s they’re doing, what’s stopping someone from just doing the same thing using maybe a very long context prompt and their own choice of AI model? I guess someone already did that (I forget his name, he’s on these forums) and supposedly it works pretty OK, but it can also supposedly completely destroy your app and make it unusable since it’s editing the .bubble file directly. You could argue the official Bubble version might run into that same problem but I guess they would have a better idea of how to implement programmatic guardrails.
