Jacob Jackson relied heavily on AI early in his profession.
Jackson co-founded Tabnine, the AI coding assistant that raised nearly $60 million in enterprise capital while still a pc science student on the University of Waterloo. After selling Tabnine to Codata in 2019 (while taking his final exams), Jackson joined OpenAI as an intern, where he worked until 2022.
At this point, Jackson felt the urge to begin an organization again that focused on supporting common developer workflows.
“In the years since I built Tabnine, tools like ChatGPT and GitHub Copilot have modified the way in which developers work,” Jackson told TechCrunch. “It's a extremely exciting time to be working on developer tools since the underlying technology has improved a lot since Tabnine was founded – which has led to many more developers being focused on using AI tools to hurry up their workflow.”
So Jackson began Supermavenan AI coding platform much like Tabnine, but with some quality of life and technical improvements.
Supermaven's own generative AI model, called Babble, can understand a number of code directly because of a context window of 1,000,000 tokens, Jackson says. (In data science, tokens are chunked raw data – just like the syllables “fan,” “tas,” and “tic” within the word “incredible.”)
The context or context window of a model refers to input data (e.g. code) that the model considers before generating output (e.g. additional code). A protracted context can prevent models from “forgetting” the contents of current documents and data and becoming off topic and drawing incorrect conclusions.
“Our large context window helps reduce the frequency of hallucinations since it allows the model to attract answers from context in situations where it will otherwise must guess,” Jackson said.
One million tokens is indeed a big context window. But it isn’t any larger than that of AI coding startup Magic, which has 100 million tokens. Meanwhile, Google's recently launched tool Code Assist matches Supermaven's context with 1 million tokens.
So what are Supermaven's benefits over the competition? Jackson claims that Babble has lower latency because of a “latest neural architecture.” He wouldn't elaborate further, aside from to say that the architecture was built “from scratch.”
“Supermaven takes 10 to twenty seconds to process a developer's code repository and turn into acquainted with their APIs and the unique conventions of their codebase,” said Jackson. “Thanks to our in-house model serving infrastructure, latency is lower, and our tool is responsive even when working with the long prompts that include large codebases.”
The marketplace for AI coding tools is large and growing. Polaris Research project that it can be price $27.17 billion by 2032. overwhelming majority of respondents to GitHub's recent developer survey say they’ve adopted AI tools in some form, and over 1.8 million people—and about 50,000 corporations—pay for GitHub Copilot.
But Supermaven – together with startup competitors like Cognition, Anysphere, Poolside, Codeium and Augment – must overcome ethical and legal challenges.
Companies are sometimes cautious about revealing proprietary code to 3rd parties. For example, Apple According to reports banned its employees from using Copilot last yr because they were concerned about confidential data leaks. Some code-generating tools trained with restrictively licensed or copyrighted code were shown render that code in a certain way, which poses a liability risk (i.e. developers who incorporate the code may very well be sued). And because AI makes mistakes, assistive coding tools can result in more faulty and unsafe code pushed into code bases.
Jackson said Supermaven doesn’t use customer data to coach its models. He did, nevertheless, admit that the corporate stores the info for per week to “make the system fast and responsive,” he said. On the topic of copyright, Jackson didn’t explicitly deny that Babble was trained using IP-protected code — only that it was “trained almost entirely using publicly available code slightly than a snippet from the general public web” to “reduce exposure to toxic content during training.”
Customers don't appear to be deterred. More than 35,000 developers use Supermaven, Jackson says, and a good portion of those pay for the premium Pro ($10 monthly) and Team ($10 monthly per use) plans. Supermaven's annual revenue reached $1 million this yr as its user base has grown threefold because the platform launched in February.
This dynamic caught the eye of VCs.
Supermaven announced its first external funding this week: a $12 million round led by Bessemer Venture Partners and notable angel investors including OpenAI co-founder John Schulman and Perplexity co-founder Denis Yarats. Jackson says the plan is to spend the cash on hiring developers (Supermaven currently has a team of 5) and on developing Supermaven's text editor, which is currently in beta.
“We plan to grow significantly by year-end,” he added. “Despite the headwinds for technology overall, the coding copilot market is growing quickly. Our growth since our launch in February – in addition to our recent funding round – positions us well for next yr.”