Jacob Jackson was completely into AI early in his career.
Jackson co-founded Tabnine, an AI coding assistant that raised $60 million in venture backing while he was a computer science student at the University of Waterloo. After selling Tapenine to Kodata in 2019 (during his final exams), Jackson joined OpenAI as an intern, where he worked until 2022.
It was at that point that Jackson had the urge to start a company again, one focused on supporting common developer workflows.
“In the years I’ve been building Tabnine, tools like ChatGPT and Github Copilot have changed the way developers work,” Jackson told TechCrunch. “It’s a very exciting time to be working on developer tools because the underlying technology has improved so much since I started Tabnine – which has led to many developers becoming interested in using AI tools to speed up their workflows.”
So Jackson launched Supermaven, an AI coding platform along the lines of Tabnine, but with some quality of life and technical upgrades.
Supermaven’s in-house generative AI model, Babble, can understand multiple codes simultaneously, thanks to a 1-million-token context window, Jackson says. (In data science, tokens are subsets of raw data—like the letters “fan,” “toss,” and “tick” in the word “awesome”)
A model’s context, or context window, represents the input data (e.g. code) that the model considers before generating output (e.g. additional code). Long contextual models can prevent “forgetting” the content of recent documents and data, and prevent subject matter from being ignored and misinterpreted.
“Our large context window helps reduce the frequency of hallucinations because it allows us to draw responses from the context in situations where the model has to guess,” Jackson said.
A million tokens is a huge context window, indeed. But it’s no bigger than Magic, an AI index startup with 100 million tokens. Meanwhile, Google’s recently introduced Code Assist tool matches Supermaven’s environment with 1 million tokens.
What are Supermaven’s advantages over competitors? Babble has low latency thanks to a “new neural architecture,” says Jackson. He would not elaborate beyond saying that the architecture was created “anew.”
“Supermaven spends 10 to 20 seconds processing a developer’s code repository to familiarize themselves with its APIs and the unique conventions of its codebase,” Jackson said. “Our tool is responsive when working with long instructions that come with large codebases, with low latency due to our built-in model infrastructure.”
The market for AI coding tools is large and growing, predicted by Polaris Research to reach $27.17 billion by 2032. The majority of respondents to GitHub’s latest dev survey say they’ve adopted AI tools in some form, and more than 1.8 million people — and ~50,000 businesses — pay for GitHub Copilot.
But Supermaven — along with startup rivals like Cognition, Anysphere, Poolside, Codeium and Augment — has ethical and legal challenges to overcome.
Businesses are often wary of exposing proprietary code to third parties; For example, Apple reportedly banned employees from using GoPilot last year, citing concerns about leaking confidential data. Some code generation tools trained using restricted licensed or copyrighted code have been shown to reproduce that code when prompted in a certain way, which poses a liability risk (ie, developers who include the code may be sued). Also, as AI makes mistakes, assistive coding tools are forced to code sites that are highly incorrect and insecure.
Supermaven doesn’t use customer data to train its models, Jackson said. However, he admitted that the company keeps the data for a week to “make the system faster and more responsive”. On the copyright matter, Jackson apparently did not deny that he was trained in IP-protected code — only that he was “trained exclusively on publicly available code, rather than a scrap on the public Internet, to minimize exposure to toxic content.” During training.”
Customers are not discouraged. More than 35,000 developers use Supermaven, Jackson says, and a significant portion pay for the Premium Pro ($10 per month) and Team ($10 per app) plans. Supermaven’s annual recurring revenue hit $1 million this year, on the back of a user base that has tripled since the platform launched in February.
That speed caught the attention of VCs.
Supermaven announced its first outside funding this week: a $12 million round led by Bessemer Venture Partners and top angel investors including OpenAI co-founder John Shulman and Perplexity co-founder Denise Yards. Jackson says the plan is to spend the money on hiring developers (Supermaven currently has a team of five) and building Supermaven’s text editor, which is currently in beta.
“We plan to grow significantly by the end of this year,” he added. “Despite the upside to technology overall, the market for coding copilots is growing rapidly. Our growth since our launch in February – as well as our most recent funding round – positions us well as we head into next year.
#coding #assistant #Supermaven #raises #money #OpenAI #Perplexity #cofounders #TechCrunch