Jacob Jackson devoted his early career to AI.
While still a computer science student at the University of Waterloo, Jackson co-founded Tabnine, an AI coding assistant that raised nearly $60 million in venture funding. After selling Tabnine to Codata in 2019 (during its final exams), Jackson joined OpenAI as an intern, where he worked until 2022.
It was at that point that Jackson felt the urge to start a company again, one focused on supporting common developer workflows.
“Since I started Tabnine, tools like ChatGPT and Github Copilot have changed the way developers work,” Jackson told TechCrunch. “Working on developer tools is really exciting because the underlying technology has improved so much since I started Tabnine, which has led to more developers being interested in using AI tools to accelerate their workflows.”
So Jackson started Supermaven, an AI coding platform similar to Tabnine, but with some quality-of-life improvements and technical upgrades.
Jackson says Babble, Supermaven’s in-house generative AI model, can understand a lot of code at once thanks to its 1 million-token context window. (In data science, tokens are granular bits of raw data, like the syllables “fan,” “tas,” and “tic” in the word “fantastic.”)
The context or context window of a model refers to the input data (e.g. code) that the model considers before generating output (e.g. additional code). A long context can prevent the model from “forgetting” recent documents and data content and from incorrectly extrapolating off-topic.
“Our large context window helps reduce the frequency of hallucinations because the model can derive answers from context in situations where it has to guess,” Jackson said.
A million tokens is actually a large context window, but it’s not as big as the 100 million tokens from AI coding startup Magic. Meanwhile, Google’s recently released Code Assist tool matches Supermaven’s context with 1 million tokens.
So what advantage does Supermaven have over its competitors? Well, Jackson claims that Babble has lower latency thanks to a “novel neural architecture.” He wouldn’t elaborate beyond saying that the architecture was developed “from scratch.”
“Supermaven takes 10 to 20 seconds to process a developer’s code repository, while it learns the API and the unique conventions of the codebase,” Jackson said. “The low latency, thanks to its own model serving infrastructure, keeps the tool responsive even when handling the long prompts that come with large codebases.”
The AI coding tools market is large and growing, with Polaris Research predicting it will be worth $27.17 billion by 2032. The majority of respondents to GitHub’s latest developer poll said they have adopted AI tools in some form, with more than 1.8 million people and about 50,000 businesses paying for GitHub Copilot.
But Supermaven, like its startup competitors Cognition, Anysphere, Poolside, Codeium, and Augment, has ethical and legal challenges to overcome.
Companies are often wary of exposing their proprietary code to third parties. Apple, for example, reportedly banned its employees from using Copilot last year, citing concerns about confidential data leaks. Some code-generating tools trained on restricted or copyrighted code have been shown to spit out that code when asked in certain ways, creating liability risks (i.e., developers who integrate the code could be sued). And because AI makes mistakes, assistive coding tools can result in more mistakes and unsafe code being pushed into the codebase.
Jackson said Supermaven doesn’t use customer data to train its models. However, he acknowledged that the company retains data for a week “to make the system fast and responsive.” On copyright issues, Jackson didn’t explicitly deny that Babble was trained with IP-protected code, only saying that it was “almost exclusively trained on publicly available code, not scraped from the open internet,” “to reduce exposure to toxic content during training.”
Customers don’t seem to be discouraged. Jackson says more than 35,000 developers are using Supermaven, many of whom pay for its premium Pro ($10/month) and Team ($10/month/use) plans. Supermaven’s annual recurring revenue hit $1 million this year, thanks to a user base that has tripled since the platform launched in February.
This trend has caught the attention of VCs.
Supermaven announced its first external funding this week, a $12 million round led by Bessemer Venture Partners and a group of angel investors including OpenAI co-founder John Schulman and Perplexity co-founder Denis Yarats. Jackson said the company plans to use the money to hire developers (Supermaven currently has a team of five) and to develop Supermaven’s text editor, which is currently in beta.
“We plan to grow significantly by the end of the year,” he added. “Despite the overall tech headwinds, the coding pilot market is growing rapidly. The growth since our February launch and our most recent funding round position us well for next year.”
