Good morning/afternoon/evening everyone 
For the past few months we’ve been running a monthly AMA with Bubble’s co-founder Emmanuel live on YouTube. It’s very casual and Emmanuel just reads and answers questions from the chat.
Here’s the link to the one starting at 12pm EST today 
@emmanuel at ~33:00 you mention using a mix of claude and chatGPT for the appgen/agent respectively. Does this actually make sense given these models aren’t trained on .bubble JSON? Wouldn’t the agent work better if you fine-tuned your own model on thousands of cases? I just don’t see how it would be possible for a frontier model to not hallucinate on a language that’s not in the training set, even if you had a giant system prompt with lots of examples.
Watched the whole thing, my takeaways,
Performance improvements = Internet will advance no need for development
AI improvements = AI will advance, don’t need to rush
SEO improvements = Internet advances so don’t need SEO, AI search will take regular searches thus regular SEO’s place
For beginners = Vague
My expectations from Bubble from now on is just it survives, and mobile improves.
For me one of the most striking (and disappointing) things was confirming this at 17m34s:
No longer an unstated ethos.
Sounded specific to AI log reviews, but overall a solid plan by bubble to leverage OPM.
Hallucinations do not occur only through lack of training data…hallucinations are inherited to LLMs and the fact they make a guess as to what should come next, and due to costs of operations, not built to self check and verify validity of first output. No amount of training data can stop an LLM from having hallucinations
We’re doing a lot of things behind the scenes, and yes, fine tuning is definitely something we’re looking into to improve quality
Building consumer products to a thankless crowd is tough gig.
From a user interface perspective, I do agree with this:
@emmanuel this is my biggest pain point with Bubble. I want pre-built UI components, similar to Apple’s UIKit framework. I am tired of spending hours trying to get pixels to look perfect.
For example, when building an iOS app, you just add a tab bar, which has a pre-set design. You don’t have to determine the tab bar height, icon size, item spacing, selected state, etc. You don’t have to customize responsiveness. It just works.
Look at Bootstrap - if you put a header on a page, it is pre-designed, collapses on mobile, has a mobile menu, and requires zero customization to work.
I’d love a fixed UI framework that allows you to build an admin dashboard for a middle-market business, without having to think about design customization.
We need native components; Bubble hasn’t released native components in years. Believe it or not, the last major release was a button with an icon? In 2026 we’ll still have the same components from years ago, leaving us reliant on plugins. This is already outdated.
dark mode enjoyers are curious if there is any plans for editor dark mode @emmanuel
@randomanon This isn’t a training problem. It’s an execution layer problem.
In the last couple of days I built a constrained Bubble agent that operates directly against the editor write protocol. It inspects the full app structure, scopes changes into discrete issues, executes through deterministic write scripts, and verifies the diff after every pass.
No fine-tuned model. No massive system prompt.
Just controlled structural access and disciplined writes.
The result is stable multi-step edits and real lifecycle logic that’s numbers of magnitude beyond what the native agent is currently doing and works out of the box on non-AI apps (I don’t see why those would be a whole separate feature?). It currently adding features and fixing issues on two live apps…
Case study coming soon…
Agreed. That part was extremely disappointing, especially in a world where people like @georgecollier are actively filling major gaps with tools like BuildPrint.
When a third party can repeatedly build essential infrastructure that becomes widely adopted, that’s a signal it belongs in the core product. Relying on plugins and 3rd parties indefinitely feels like deferring ownership of critical capabilities.
Another moment that stood out was the framing that visual building is the future and the alternative is “semicolons and if/else statements.”
That’s a false dichotomy (and a dangerous one).
The alternative to visual building today isn’t raw syntax. It’s structured systems and tooling that abstracts syntax away entirely. Serious builders haven’t been thinking in punctuation for years. They think in architecture, state, and deterministic behavior.
^^^ this
The part bubble is missing is due to its builder it will always need user interaction to understand it, agent flows require just understanding English. Building production level apps vibe coded requires understanding of how to guide AI and what questions to ask but today it does it and it does it very very well.
Bubble is a skill.
AI vibe coding to a production level is a skill.
The majority of people building on bubble have bad optimization, awful security practice, awful UI, etc. it’s not production ready either.
Vibe coding is faster to learn, easier to understand, and the skill gap between building production and hobby is far smaller.
Having production apps on both sides and hearing bubble executives/staff justify bubble it seems like they themselves may not understand what they’re up against.
That’s ironic. That’s not what Vibe bros usually say.
Skill comes in keeping it running in prod and I have not come across a single vibe coded market viable app that is surviving production. Feel free to prove me wrong.
There will always be a place for a visual builder like Bubble. Bubble AI has far more relevant than vibe coded code because it already serves the segment that vibe coding attracts.
Bubble has already been designed to be a bridge between non-technical users and production level apps. It takes away friction of building and keeping an app in production.
Anyway there will soon be a market full of vibe coded apps that need to be redone.
Vibe bros don’t own vibe coding and there are allowed to be other approaches than theirs and there are also allowed to be highly skilled ppl who also leverage AI coding because of it being a scalable force multiplier and allows them to spend their time doing what they are best at.
Where are you looking? vibe bros may not have any (i wouldnt know) but there are thousands out there (including dozens of bubble apps recently migrated). And then there’s an app you probably use all the time CoWork…
Never heard of CoWork. I use a similar service; Deskimo, which is pretty crap already. Any other names?
There’s no denying that SWEs that add LLMs to their toolbox are benefitting greatly from them. What needs to be understood better in this forums is that there is a distinction between vibe coding and using LLMs as another tool. They are 2 very different definitions.
Vibe coding is very much it’s name sake. The other still requires expertise and experience. They do not produce the same quality of work.
I’m extremely frustrated with Bubble’s AI builder… Bubble as a platform is great. I build real systems, solid MVPs, and fairly complex applications with it. But today, the AI feels many years behind what already exists in the market.
It tries to help, but in practice it generates fragile structures, inconsistent workflows, broken logic, and architectures that almost always need to be rebuilt manually. For anyone building something serious, it often creates more work than it saves.
As a concrete example: I recently asked it to generate a simple login page. What I got was completely broken, styles all over the place, colors totally out of context, components misaligned, and the page wasn’t even functional.
And when you factor in the cost of WUs, it becomes even more concerning. More and more, Bubble starts to feel obsolete compared to other tools that offer better structural efficiency, deeper technical control, and much stronger integration with truly capable AI models.
So my question is straightforward: is there a real plan for this AI to improve in a meaningful way?
Because honestly… a much more effective approach might look closer to what n8n does: provide clear documentation of Bubble’s internal structure (schemas, entities, relationships, workflows), so we could use a real AI (Claude, GPT, etc.) to generate the application externally, import it into Bubble, and then simply refine UI, logic, and workflows.
Right now, it feels like we’re locked into an internal AI that limits the platform’s potential, while the rest of the ecosystem has already moved far ahead.
Is anything concrete on the roadmap for this?
I’d really appreciate hearing a clearer perspective from the Bubble team about the future of this part of the product.
They know it’s bad and it their #1 priority out of everything they’re working on. They understand it’s life and death if they don’t:
I’m just not sure why progress is so slow moving. Maybe because they’re trying to ingest the entire app’s context instead of doing what I proposed, which is to create a “sandbox” editor where individual workflows/actions/elements can be created and edited by the agent then copied over to the app manually by the user.
I think they’re trying to skip a step and make it context-aware without having the model or context windows necessary to do that.
Ye of little faith!
Even AI models didn’t do this well* until Codex 5.3. Difference is:
I have faith, I just think they will take a long time to eventually come to my conclusion (Monthly Community Update - January 2026 - #16 by randomanon). It’s going to be the “Beta Expression Composer” all over again. It’s funny because they wasted hundreds (thousands?) of man hours on something that wasn’t technically feasible and might be doing the same thing all over again.
*and arguably still don’t depending on what your standards of “well” are
