It seems every day there are massive advancements on vibe coding. I myself have already created small, little single-use apps with Manus. It’s only a matter of time where creating apps will be a single prompt. My fear is that Bubble isn’t moving fast enough, and the fact that they have a UI builder, a database, API connectors, and more means that they could be amazing at vibe coding or vibe Bubble building.
But as of yet, I’ve been unimpressed with Bubble’s AI implementation.
I’m worried about Bubble because I feel like one of these days someone’s going to crack the code of creating a complex, usable app that has API connections, a database hosted on a server, works on mobile, and is generated with simple prompts.
One of these days is a likely outcome, but what day, is that day 90, day 180 or day 1,600? For me, in what I understand about how to build apps on Bubble, and the issues that can arise from doing something wrong in one area, having a knock on effect on another area, plus my experience using AI for attempting to write code, AI is nowhere need capable and is likely trillions of dollars away from being capable.
AI, even if they slap the ‘reasoning’ model onto it’s name, do not reason like an intelligent being can. AI is incapable of picking the best path when met with 3 paths to follow, because AI is not always incorporating a true and full sense of the complete needs. For me that is, when I am aiming to fix a problem, the problem has been isolated and understood, but when there are then 3 options for how to move forward, the AI doesn’t choose the method that is optimal for the situation.
I personally think AI is stupid, because it is not intelligent. It is good at guessing, and using it’s training to regurgitate information better than a Parrot would, and no matter how much training is put into the AI, it still does not always follow orders.
From what I have taken away from the Bubble AMA and other announcements by Bubble is that they are aiming for a better experience than vibe coding. I personally do not even know what vibe coding is or what people are thinking they are doing; are they ‘vibing’ with the AI like they are on a date and hitting it off with the person and so their vibes are in vibration with one another and it’s like they are one, or are they just playing music in the background thinking ‘hey, this AI thing is doing my brains mental work for it’…I don’t know. But in my experience with AI, I need to use my logic to slow it down, get it to focus and to stay on task. Often enough I will allow AI to lead, only to stray so far away, I have to go back to step 1 to build it back up step by step.
All of that is said a a non-technical person. I’m sure if I were proficient in writing code, vibe-coding might be my thing, as I’d have the technical knowledge of how to spot check quickly poor code and fix on my own, but I am non-technical so vibe-coding to create a fully functional app from a few prompts is so far out of reach for me, and I do not imagine a near term future where the investors backing these AI models will have poured enough money into them to improve them fast enough to make it so somebody like me can actually do that anytime soon.
Here is what AI told me as a cost to create a new model
Putting it all together, a single, incremental advance beyond GPT-4 typically runs $100 million–$200 million end-to-end. A truly breakthrough next-generation model (e.g. “GPT-5”)—with 2–10× more compute, larger datasets, deeper RLHF cycles, and expanded safety guardrails—could easily push the price tag into the $500 million–$1 billion zone before you flip the “launch” switch.
And very importantly this
Infrastructure, overhead & safety evaluation
Datacenter ops: Power, cooling, hardware depreciation, networking—keeping an exascale GPU cluster humming can cost tens of millions annually.
There is a ton of investment globally right now into data centers. Thailand alone has around 45 such projects announced with an estimated total cost of $380 Billion…Building such large infrastructure projects takes time, sometimes 2-3 years.
I do not see AI building full apps from a few prompts as near term, I see it as medium term, like 3-5 years, and by that time, Bubble will have, like all other companies, piggy backed off the actual LLMs that power their AI builder to have improved as well.
Bubble likely will come out in relatively the same position they are today, which is serving a larger audience of non-technical users who may benefit from the visual development platform. Everybody has different ways of learning, some auditorily, some visual, and due to that ingrained difference between people, there will still be millions of people who would prefer Bubble over an AI coding agent.
I feel somewhat qualified to respond to this as I built Buildprint using code as someone that’s a non-coder to test what Bubble is up against. I am, however, technical. The conclusion I have come to is that good Bubble developers will be able to build with AI-assisted code. Bad Bubble developers will not.
The reason for this is that when you use AI-assisted coding, you cannot build what you do not understand. The AI will handle the syntax for you, but it does not handle logic. You need to know the principles that you’re instructing the AI to use and what questions to ask.
If you’re a good Bubble developer, you understand the principles of modular and maintainable development design, the same of which apply to AI-assisted code. There are lots of 1:1 mappings from Bubble to code. However, the second you build something you don’t understand, you are not able to maintain it because when it goes wrong, you have no chance of fixing it.
The example I always have is you’ve never seen an AI-assisted code generator app like Replit, Bolt, or Lovable which runs a database migration. Have you ever seen someone change their database structure while the app is in production without wiping it entirely to, for example, drop an existing table or add a new table and back-populate it?
In Bubble, that’s trivial because we create it and then run a bulk backend workflow. Do you know what to do if that goes wrong in traditional code? What if your migration is recorded but doesn’t happen as expected? Do you know how to revert it? No, you don’t.
With that said, even though I was able to build with code in roughly the same time as I would build with Bubble, I still wouldn’t recommend any client to do it, at least not with us yet. The reason for it is that it makes us responsible for DevOps.
If a server breaks, that’s on us and we have to fix it. With Bubble, the data is backed up reliably, we have extensive logging history, and we have compliance options which exist by default. If a server goes down, someone in the US gets woken up and fixes it - not me.
I’ve had a few clients come to us having built prototypes on Lovable. I think on the whole the AI race will be an advantage for Bubble.
The AI-powered apps, even if they’re not necessarily better, will outpace Bubble in terms of growth, but I still think that Bubble will grow in terms of user numbers. Whether they grow fast enough is a different question.
I think the AI revolution has drawn more attention to the fact that anyone can build an app, and Bubble is a very good platform to do it for serious projects, rather than hobbyist tinkerers.
If you don’t know how to code, you won’t get very far with vibe coding
Unintelligible errors will popup
Complex logic is very difficult to express with a single prompt
Maintainability and scalability of the output code will be a nightmare
If you don’t know how to code you’ll get better results developing with Bubble.
AI capabilities to transform plain words into complex coded logic might improve a lot in the upcoming years, but I believe the aforementioned facts will always stand.
I agree that a pro Bubble developer has acquired and shares skills with a (code?) developer. Software architectural concepts and strategies like client/server-side, latency, recursivity or synchronicity are well known and mastered by such a person.
However, when you jump to actual code, there are other fundamental architectural concepts and strategies that you have to master to develop a product that performs similar to that developed with Bubbe. And those concepts are not covered when developing with Bubble.
Concepts like microservices, multithreading or architectural patterns like MVC (model-view-controller) are not covered by Bubble and are needed for successful code development, and I think it’s something vibe coding will hardly do for you. I doubt that a pro Bubble developer have mastered them without previous code experience.
I really don’t see how anyone could create a real app using AI. I am pretty adept at creating prompts. (The secret is to corral the AI in so it has no logical avenue to go down the wrong path) But take an app like Uber for example. There is no way in hell you are writing a prompt to create all the functionality of the backend and UI of Uber.
Even if by some miracle, you spend 6 months perfecting the perfect prompt, there will be so many details wrong with the product to fix that you might as well have started using SWIFT and Kotlin/Java
There are enough abstractions where you can actually not understand coding/SWE/computer science on a deep level and get away with it. That’s the whole point of Vercel, Supabase, Xano, (arguably AWS invented this “abstraction as a service” concept) etc. It’s hilarious when people think these are somehow different from Bubble.
You only need to learn that stuff if you’re working a FANG job. Otherwise if you’re just trying to build a profitable app you can just pick your slop poison and go to town.
Traditionally, over the past 7 years, our prototypes have been done in Bubble. Last week, the team put together their own prototype using Firebase Studio.
Although it only had mock data, they were able to put together the prototype quicker than what I could have even done in Bubble – And even though the code isn’t the best.. well, with Bubble, it’s not either.
Sure, the Firebase prototype has some chunky loads.. but so does my Bubble prototypes.
I am coming to terms with the fact that things are changing pretty rapidly – Bubble still has it’s uses.. But the walls are closing in. People are caring less and less how things are built and more “I want these 5 features.. And if AI can do it in 2 1/2 minutes vs you charging me 2 hours, fine”.
I don’t like it anymore than anyone else.. but things are changing.
Indeed, all those mentioned platforms are meticulously crafted following the aforementioned dev strategies (and others) so that everything complicated is abstracted for the developer and they can serve millions of users concurrently.
I agree that going all-in with code it’s not the best option nowadays unless you are super knowledgeable and comfortable with it. In fact I am not defending code as the only way to get superb software. I believe the future of software development goes through abstraction layers like Bubble, Supabase or Xano (that’s why I’m here myself, actually), just because it cuts down costs and dev time.
On the other hand, if one decides to vibe-code an app entirely, following code development best practices might not be needed at the beginning. You can just vibe-code a monolithic app, deploy it on a single AWS instance and serve your first thousands of users perfectly fine. But at some point your setup will choke and you’ll need to go for a microservices architecture, dockerizing those microservices and orchestrate them with Kubernetes. That’s the exact moment one understands that the app must be completely rebuilt.
I think it’s also important to get out of the idea that ‘vibe coding’ is the only way to use AI to code. It’s a bad way.
A technical user that cannot code can use AI to code so long as they ensure they understand what the AI has written. That’s not vibe coding.
And wait until you discover what a Bubble dedicated instance looks like!
Jokes aside, I’m obviously pretty heavily invested in Bubble but I’m also not married to it. Bubble is still the best option for most client apps and AI code has increased our workload rather than taken it away. So, just talking from our experience
I totally agree with this point. I would only use AI as a copilot, to automate time-consuming tasks, e.g. translating DB objects into whatever my frontend consumes, reducing the context the AI is working on.
