There’s a strange irony in the tech world that software has succeeded in automating pretty much everything apart from itself. While we use SaaS products to help us do everything from organise our time to secure our most sensitive data, the complex process of developing an app remains one that needs humans firmly in the loop.
And while tools like Github Copilot have made it possible for software engineers to write code faster, using them still requires substantial technical expertise.
Startups around the world are trying to tackle this problem, by using AI to build tools that let anyone build an app by simply describing what they want it to do with natural language.
One of them is Stockholm-based Loveable, which is today unveiling a new product which it says can automate almost everything that a full-stack engineer can do.
“We claim 95% of the value of software can be created by using the hottest programming language, plain English,” founder and CEO Anton Osika tells Sifted.
The company, which raised a €6.8m pre-seed round in October, is one of a growing batch of startups popping up around Europe to tackle the problem. And it’s no wonder, given that the software development industry is projected to grow from $203bn 2022 to $1.45tn by 2031.
Others companies in the space have founders with experience at companies like Microsoft and Monzo, as some of Europe’s smartest technologists put their minds to building a future where software can be built by anyone.
The value of humans
Loveable says it’s already seeing big demand for its product, dubbed GPT Engineer, and has seen its user base double month on month, with “more than a thousand products currently being built on the app.”
Osika tells Sifted that, historically, people have struggled to get AI to build software because it struggles with complex tasks.
“The problem that we all faced when using AI is that it's super impressive at first, like, ‘Oh magic pixie dust! It does things that seem like a human.’ But then at some point you notice that it gets really confused and you hit the wall in what you want to achieve,” he says. “Now we moved it so that you can get to the point of value before you hit the wall.”
Osika says Loveable has done this by building an AI system that breaks up software development “into different pieces and makes sure that each piece individually works.” He adds that while other startups like San Francisco-based Cognition AI are using AI to check its own code, he believes that humans tend to be best at checking if an app is working well for users.
He says that by combining creative insight from a person with AI’s speed at writing code, GPT Engineer can accelerate product development 20x: “AI is extremely fast, but humans are better at many other things, and if you have these very fast iteration loops between human and AI, you get synergy.”
GPT Engineer users develop apps using simple, natural language prompts to develop each piece of their app, and then check them to make sure they’re working as they intended.
Osika does add that, currently, he wouldn’t advise people to build “mission critical” software with GPT Engineer just yet, and that its biggest user base currently consists of founders who use the tool to build prototypes quickly.
Different schools of AI
Another startup automating software is London-based Agemo, founded by Aymeric Zhuo, former data science lead at Microsoft-owned gaming studio Activision Blizzard, and Osman Ramadan, a former AI research scientist at Microsoft.
The company raised a $4m seed round led by Firstminute Capital in 2023 to build a product that combines generative AI capabilities with other technology to enhance the system’s ability to reason.
Specifically, Agemo is using techniques similar to “reinforcement learning,” which powers Google DeepMind’s AlphaGo programme — a tool that made headlines by beating a world champion player at the game of Go in 2017. This, Zhuo says, allows the product his team is building to constantly learn and improve its understanding of how software works.
It’s using that AI approach combined with third party large language models (LLMs), which it fine-tunes to specialise them for software development. Zhuo describes Agemo’s product as making the best use of the powerful capabilities of generative AI, combining them with the more reliable capabilities of traditional software, as a way of minimising errors in the systems output.
“We encode the rules of software in an environment, and we teach the LLM to operate in that environment,” he explains.
Like Loveable, Agemo isn’t yet able to just build any type of software from scratch, and is first focusing on letting users build backend systems, rather than the frontend that users interact with. Zhuo says the startup is already working with design partners — one of them is a YouTuber with 100k followers who’s used the tool to build a software programme that automatically transcribes their videos and turns them into Linkedin posts.
A hot market
Startups like these are operating in a competitive market, taking on companies that have raised far larger sums like Paris-based Poolside, which raised a $500m Series B in October to continue building its AI model for automatically writing software. There’s also San Francisco-based Magic, which closed a $300m round in August to build its own foundation model for code.
Despite the juggernaut competition, new startups are being founded by experienced operators like Monzo cofounder Jonas Templestein, who paired up with Oliver Beattie, former VP of architecture at the digital bank, to launch London-based AI software builder Nustom this year.
Leaner outfits like them, Loveable and Agemo can go up against the likes of Poolside and Magic with less funding, as they’re not training their own large models from scratch — a process that costs tens of millions of dollars — and are choosing to adapt existing models.
Whoever comes out top of the race to automate software is likely to be the company that can best combine different AI approaches, while building for a product that lets human creativity continue to shine during the process.