Jacob Jackson, Tabnine co-founder, launches Supermaven, an AI coding platform, raising $12M to improve developer workflows with a million-token context window.
While still a University of Waterloo computer science student, Jacob Jackson co-founded Tabnine, the AI coding helper that raised nearly $60 million in venture capital. In 2019, when he was doing his final exams, Jackson sold Tabnine to Codata. He then started working as an intern at OpenAI and stayed there until 2022.
Jacob Jackson felt the need to launch a new business to facilitate standard developer workflows.
“Since I founded Tabnine, tools like Github Copilot and ChatGPT have revolutionized the way developers work,” Jackson said in an interview with TechCrunch. “Working on developer tools is exciting because a lot has changed in the underlying technology since I started Tabnine. As a result, many more developers are interested in utilizing AI tools to speed up their workflow.”
Jackson then founded Supermaven, an AI coding platform similar to Tabnine but with several technological and quality-of-life improvements.
According to Jackson, Supermaven’s proprietary generative AI model, Babble, has a context window of one million tokens, allowing it to comprehend a large amount of code simultaneously. Tokens, which are split chunks of raw data in data science, are similar to the syllables “fan,” “tas,” and “tic” in the word “fantastic.”
The context, also known as the context window, is the set of input data (such as code) the model considers before producing output (such as more code). Long context can stop models from “forgetting” the information in recent documents and data, straying from the subject, and drawing incorrect conclusions.
“Our extensive context window allows the model to infer answers from the context in scenarios where it would otherwise have to guess, which helps reduce the frequency of hallucinations,” Jackson explained.
Indeed, a million tokens is a significant context frame. However, it is not greater than the 100 million tokens AI coding firm Magic held. At one million tokens, Google’s recently released Code Assist feature corresponds with Supermaven’s context.
Benefits of AI Coding Platform Supermaven
What benefits does Supermaven offer over competitors, then? According to Jackson, Babble has reduced latency because of a “new neural architecture.” He would only remark that the architecture was created “from scratch” without going into further detail.
According to Jackson, “Supermaven takes ten to twenty seconds to process a developer’s code repository to get acquainted with its APIs and the specific conventions of its codebase.” “Due to our internal model serving infrastructure, our tool has reduced latency and maintains responsiveness even when interacting with lengthy prompts associated with large codebases.”
According to Polaris Research, the market for AI coding tools is sizable and expanding, and by 2032, it is expected to be valued at $27.17 billion. In GitHub’s most recent developer survey, the overwhelming majority of participants said they had used AI technologies in some capacity. Additionally, over 1.8 million users—including almost 50,000 companies—are paying for GitHub Copilot.
However, Supermaven faces moral and legal obstacles that must be addressed, as do upstart rivals like Cognition, Anysphere, Poolside, Codeium, and Augment.
Companies frequently avoid disclosing proprietary code to outside parties. According to reports, Apple prohibited employees from using Copilot last year due to worries about leaking sensitive information. It has been demonstrated that specific code-generating systems trained on copyrighted code or subject to restrictive licenses may reproduce that code when given specific instructions, which presents a liability concern (i.e., developers that incorporate the code could be sued). Additionally, assisted coding tools may lead to more incorrect and unsafe code being added to codebases because AI is fallible.
According to Jackson, Supermaven trains its models without using client data. To “make the system quick and responsive,” he acknowledged that the business keeps data for a week. Regarding copyright, Jackson stated that Babble was “trained almost exclusively on publicly available code rather than a scrape of the public internet” to “reduce exposure to toxic content during training.” Still, he did not specifically dispute that the software was taught on IP-protected code.
Consumers don’t seem to be deterred. According to Jackson, over 35,000 developers are using Supermaven, and a significant portion is paying for the Pro ($10 per month) and Team ($10 per month per usage) premium plans. Supermaven’s recurring revenue increased to $1 million this year thanks to a tripled user base since the platform’s debut.
VCs took notice of that movement.
This week, Supermaven said it had raised $12 million in its first round of outside funding, led by Bessemer Venture Partners and featuring prominent angel investors such as Denis Yarats, co-founder of Perplexity, and John Schulman, co-founder of OpenAI. According to Jackson, the money will be used to create Supermaven’s text editor, which is presently in beta, and hire developers (Supermaven currently employs five people).
He continued, “We intend to grow significantly through the end of the year.” “Despite challenges facing the tech industry, demand for coding copilots has been rising swiftly. Our most recent investment round and our progress since our February launch put us in a strong position going into the new year.