Lovable: Let AI help you make your Minimum Lovable Product

Anyone should be able to build

In the past few decades, the productive output of a single developer has increased by a few orders of magnitude, due to improved hardware combined with higher-level abstractions enabled by software. We can build and ship products with smaller teams than ever before, and developer productivity is continuing to increase with LLMs generating code with tools such as GitHub Copilot.

But still, if you want to build a high-quality application, you need expertise in coding and design, or access to experts’ time. The average person does not have this. The prevalence of low/no-code tools has made product development more accessible, but building useful products is still hard, even with these tools.

Lovable wants to enable any person on earth to build and ship products by themselves that would currently require full teams of developers, designers, QA engineers and product managers. We are extremely proud to be partnering with them on the road ahead.

‍

The path to automating product development

Current LLMs are already good at writing code. The first use case for this has been in developer tooling to make developers even more productive — we have GitHub Copilot, Cody, GitButler and a whole host of similar tools. They will increase developer productivity by another order of magnitude or so, and somewhat enable the broader public to be able to write code snippets to solve small problems, such as plotting or analysing data. 

Current LLMs are also quite good at understanding human intent and using it to make decisions. Most big tech companies have already shipped AI Chatbot features as an alternative interface to their products. Wix ADI, Framer AI, Unbounce and similar products allow you to describe the website you want to build in text format which then uses an LLM to interpret and translate it into a schema for use in their templating software. This is a very impressive demo, but ultimately will only allow users to create products that are already possible to make in drag and drop UIs, but more quickly and conveniently.

In order to make end-to-end AI product development possible, LLMs need an environment in which they can do anything that humans do when building a product: i.e. write, modify and deploy code to the internet, create services and debug errors in the live production/test environments. Current LLMs can do this at a rudimentary (although still impressive) level.

Product development requires much more complex code generation than simple one-shot prompting. To create a product, you need to: 

• Understand and define the technical requirements
• Plan how to structure the project visually and technically
• Write and deploy the code
• Check whether the requirements have been satisfied
• Iterate

This requires an agent-like entity similar to the many agent frameworks that have built on top of existing LLMs such as AutoGPT, GPTEngineer (the foundation for Lovable) and smol-developer. Commonly, some sort of thought-loop and recursive delegation process is used to turn an LLM into an agent. One common framework is ReACT:

‍

‍

However, as anyone that has played with AutoGPT or GPTEngineer knows, LLMs are still not great at long term planning, and frequently fail, hallucinate, get side-tracked or stuck in loops. The high level reasoning that is observed when chatting directly with e.g. GPT-4o or Claude 3.5 Sonnet is downgraded by several steps when putting it into the out-of-distribution context that is required to make it agentic.

In order to make LLMs better in an agentic context, they will need to train specifically on data from an agentic environment. But this data does not yet exist online, and it requires some sort of reward signal to tell good from bad, which means we need humans in the loop. 

Additionally, while LLMs are still improving, they will need a forgiving environment in which to operate — where hallucinations are identified and corrected for, in which faulty code doesn’t get shipped straight to production, and where they can experiment and learn from mistakes in their planning process

Lovable is building this environment, and optimising it so that feedback from humans is captured and used to finetune and train models to get better at critical bottlenecks in the product development process.

What does this look like?

For the user, Lovable is a chatbox where you type a description of what you want to build. Lovable AI generates the code and spins up an environment in which the product can be previewed immediately and launched on the internet with one click of a button. The user provides feedback or more instructions via chat to Lovable AI which incorporates this feedback, modifies the code and shows you the new preview of each change. Each step is saved with version control and can be rolled back instantly.

The project is hosted automatically via a one-click deploy. The code is stored in a GitLab repository which can be modified directly if the user wants to migrate to another service or take over the project and start coding manually instead. This will be useful for very quickly prototyping MVPs internally, or making simple frontend apps to incorporate into a larger product.

This means the product is already a useful tool both for the average person who doesn’t know how to code, but wants to make a website, and for developers who want to quickly prototype or get inspiration for an MVP.

‍

‍

The current product supports building frontend-only websites (it also supports using open APIs inside the frontend code). The next step is to allow the AI model to choose and access a set of standardized backend services that can be incorporated into a product, such as mail, database, authentication, functions etc.

As users can build more complex products within Lovable, the environment will need to become more forgiving, both for the AI models and for the user. And signals of whether the AI has delivered what the user intended will become crucial, and form the basis for making models that are uniquely good at doing this work.

Founders

It’s hard to imagine a better team to tackle this problem than Lovable’s founding team. Anton Osika has a background in physics at CERN, was the first employee at Sana, founded Depict and created GPT Engineer in the summer of 2023 which became one of the fastest growing github repositories ever. Fabian comes with deep engineering expertise from Depict and CurbFood, and founded TenFast — software for property managers. 

Together, they are leveraging the massive amount of interest that GPT-Engineer kick-started, with more than 100k having used the GitHub repo, which accumulated 45k stars, 80+ contributors, 7,400 forks and 698 commits. The GPT-Engineer repository lives on as open source, with a thriving community of builders. Lovable’s first project (gptengineer.app) is also open source and allows building of one-page frontend-only applications. The next step is to allow full stack applications to be developed and deployed.

‍

First steps

gptengineer.app is already launched and being used by thousands of builders. Meanwhile, Lovable is expanding the team and building the next iteration of end-to-end application creation.

Welcome to the byFounders community, Lovable. We are extremely excited to be working with you. 

‍