Magic, a startup developing a code-generating platform similar to GitHub’s Copilot, today announced that it raised $23 million in a Series A funding round led by Alphabet’s CapitalG with participation from Elad Gil, Nat Friedman and Amplify Partners. So what’s its story?
Magic’s CEO and co-founder, Eric Steinberger, says that he was inspired by the potential of AI at a young age. In high school, he and his friends wired up the school’s computers for machine learning algorithm training, an experience that planted the seeds for Steinberger’s computer science degree and his job at Meta as an AI researcher.
“I spent years exploring potential paths to artificial general intelligence, and then large language models (LLMs) were invented,” Steinberger told TechCrunch in an email interview. “I realized that combining LLMs trained on code with my research on neural memory and reinforcement learning might allow us to build an AI software engineer that feels like a true colleague, not just a tool. This would be extraordinarily useful for companies and developers.”
Steinberger teamed up with Sebastian De Ro to found Magic, an AI-driven tool designed to help software engineers write, review, debug and plan code changes. The tool, not yet generally available, can “communicate” in natural language and collaborate with users on code changes, Steinberger claims — operating like a pair programmer that’s able to understand and continuously learn more about the context of both coding projects and developers.
“Magic aims to drastically reduce the time and financial cost of developing software,” Steinberger said. “Giving teams access to an AI colleague who can understand legacy code and help new developers navigate it will enable companies to scale the impact of their current employees and train new employees with less personal coaching. In turn, employees will grow their skills faster and will be able to move among high-impact projects with increased agility.”
Steinberger isn’t revealing much about Magic’s technical underpinnings yet — making it tough, frankly, to compare the tool with the competition. The elephant in the room is the aforementioned Copilot, which was trained on public code to suggest additional lines of code in response to a description of what a developer wants to accomplish — or even explain what a portion of code does.
Steinberger promises that Magic will be able to do the same — and more — thanks to a “new neural network architecture that can read 100x more lines of code than Transformers.” (The Transformer, pioneered by Google researchers, is perhaps the most popular architecture at present for natural language tasks, demonstrating an aptitude not only for generating code but also for summarizing documents, translating between languages and even analyzing biological sequences.) But absent a demo, we have only his word to go on.
“Early releases will need human supervision, but our ultimate aim is for AI to complete large tasks reliably for you, end-to-end, without babysitting,” Steinberger added.
Perhaps the bigger, more existential problem for Magic is that Copilot already has a large following — and substantial corporate backing. It’s been used by over 1.2 million people, and GitHub is aggressively positioning it as an enterprise-scale tool, recently launching a corporate-focused plan called Copilot for Business.
Copilot’s traction might’ve contributed to the demise of Kite, a startup that was developing an AI-powered coding assistant not unlike Magic’s. Despite securing millions in VC backing, Kite struggled to pay the bills, running into headwinds that made finding a product-market fit impossible. Training AI is notoriously expensive; Kite founder Adam Smith estimated that it could cost over $100 million to build a “production-quality” tool capable of synthesizing code reliably.
“Within AI more broadly, training state-of-the-art models remains expensive,” Steinberger admitted. “This raises the bar for new entrants like us.”
Legal issues might stand in the way of Magic’s success, too — although some have yet to be resolved in the courts. Like most AI-powered code-generating systems, Magic was trained on publicly available code, some of which is copyrighted. The company argues that fair use — the doctrine in U.S. law that permits the use of copyrighted material without first having to obtain permission from the rights holder — protects it in the event that Copilot was knowingly or unknowingly developed against copyrighted code. But not everyone agrees. Microsoft, GitHub and OpenAI are being sued in a class action lawsuit that accuses them of violating copyright law by allowing Copilot to regurgitate sections of licensed code without providing credit.
Some legal experts have also argued that AI-powered coding systems could put companies at risk if they were to unwittingly incorporate copyrighted suggestions from the tool into their production software.
To these questions, Steinberger answered that Magic is taking steps to prevent copyrighted code from showing up in the tool’s suggestions and citing the source of suggested code where possible. (GitHub has taken similar steps with Copilot, filtering its output in some cases and experimenting with code and project citation.) Steinberger says that customers’ data will not be swept up for Magic’s proprietary AI training — excepting “personalized systems” used by individual customers.
“We will launch with a feature that flags any potential license issues with generated code to help the user make an educated decision on what to do with it,” he said, clarifying the earlier point.
Steinberger argues that, in any case, tools like Magic — and rivals such as Tabnine, Mutable and Mintlify plus open source projects like BigCode — are a net good for both developers and their employers. He pointed to statistics showing that skilled software engineers — who are increasingly hard to come by — cost around $150,000 per year (and up) and that teams spend upward of 25% of their time integrating and maintaining their development toolchains.
Not all programmers are likely to agree, particularly those affected by the tech industry’s recent mass layoffs. But as Steinberger notes, there’s a “tremendous” level of excitement about — and investment in — generative AI. It’s clear that it’s here to stay, in other words, for better or for worse.
“The software industry has a never-ending hunger for more talent. Every organization and product would benefit from more and better software shipped faster and cheaper,” Steinberger said. “Even with all the dev tooling we have available today, output is limited by human thinking, typing, and communication speed. Giving teams access to an AI colleague who can understand legacy code and help new developers navigate will enable companies to scale the impact of their current employees and train new employees with less personal coaching. In turn, employees will grow their skills faster and will be able to move among high-impact projects with increased agility.”
Magic, which is pre-revenue with a distributed workforce of six people, plans to launch its product in the near future — Steinberger wouldn’t say exactly when. The short-term goal (i.e., within the next year) is to grow the team to 25 people with a focus on the engineering, product and go-to-market sides.
To date, Magic has raised $28 million.