Jacob Jackson was committed to AI early in his career.
While studying computer science at the University of Waterloo, Jackson co-founded Tabnine, an AI coding assistant that raised nearly $60 million in venture capital. After selling Tabnine to Codata in 2019 (during his final exams), Jackson joined OpenAI as an intern, where he worked until 2022.
At that point, Jackson felt the urge to start a company again, one focused on supporting common developer workflows.
“In the years since we created Tabnine, tools like ChatGPT and Github Copilot have changed the way developers work,” Jackson told TechCrunch. “It's a really exciting time to be working in developer tools, because the underlying technology has advanced significantly since we started Tabnine. As a result, more developers are interested in using AI tools to accelerate their workflows.”
So Jackson launched Supermaven, an AI coding platform similar to Tabnine but with some quality of life and tech improvements.
Jackson says that SuperMaven's in-house generated AI model, Babble, can understand large amounts of code at once thanks to a context window of one million tokens. (In data science, a token is a small bit of raw data, like the syllables “fan,” “tas,” and “tic” in the word “fantastic.”)
A model's context, or context window, refers to the input data (e.g. code) that the model considers before producing an output (e.g. more code). Longer context can prevent a model from “forgetting” the contents of recent documents or data, or from going off topic and making incorrect inferences.
“Our large context window helps reduce the frequency of hallucinations because it allows the model to derive answers from the context even in situations where it has to guess,” Jackson said.
One million tokens is certainly a large context window, but it's not as big as AI coding startup Magic's 100 million tokens, while Google's recently introduced Code Assist tool matches Supermaven's context with 1 million tokens.
So what makes Supermaven better than its rivals? Jackson claims that Babble's low latency is due to its “novel neural architecture.” He didn't offer any details other than to say that the architecture was developed “from the ground up.”
“Supermaven takes 10-20 seconds to process a developer's code repository and become familiar with the API and the unique conventions of the codebase,” says Jackson. “Our in-house model serving infrastructure ensures low latency, so our tool remains responsive even while processing the lengthy prompts that come with large codebases.”
The market for AI coding tools is growing significantly, with Polaris Research predicting it will be worth $27.17 billion by 2032. The majority of respondents to GitHub's latest developer survey said they are deploying AI tools in some form, with over 1.8 million people and nearly 50,000 companies paying for GitHub Copilot.
But Supermaven, like startup competitors such as Cognition, Anysphere, Poolside, Codeium and Augment, has ethical and legal challenges to overcome.
Companies are often wary of exposing proprietary code to third parties. For example, Apple reportedly banned employees from using Copilot last year over concerns about leaking sensitive data. Some code-generation tools trained with restrictively licensed or copyrighted code have been shown to spit out that code when prompted in certain ways, posing liability risks (i.e. developers who incorporate that code could be sued). And because AI makes mistakes, assistive coding tools can push more erroneous and unsafe code into the codebase.
Jackson said Supermaven doesn't use customer data to train its models, but acknowledged that the company keeps the data for a week to “make our systems fast and responsive.” On the copyright front, Jackson didn't specifically deny that Babble was trained with IP-protected code, but said it was “trained almost exclusively with publicly available code, rather than scraping it from the public internet,” to “reduce exposure to harmful content during training.”
Customers don't seem to be giving up: Jackson says more than 35,000 developers use Supermaven, with a significant number paying for the premium Pro ($10/month) and Team ($10/month per use) plans. Supermaven's annual recurring revenue hit $1 million this year, on the back of a user base that's grown threefold since the platform launched in February.
That momentum has attracted the attention of VCs.
Supermaven announced its first outside funding this week: $12 million led by Bessemer Venture Partners and prominent angel investors including OpenAI co-founder John Schulman and Perplexity co-founder Denis Yarats. Jackson said the money will be used to hire developers (Supermaven currently has a team of five) and further develop Supermaven's text editor, which is currently in beta.
“We plan for significant growth through the end of the year,” he added. “Despite general technology headwinds, Coding Copilot's market is growing rapidly. Our growth since launching in February, along with our recent funding round, positions us well for next year.”