Large Language Models, or LLMs, are deep-learning algorithms based on a transformer architecture that was first published by Google Brain in 2017. Through training on publicly available code, LLMs are capable of generating relevant code suggestions, or autocomplete for even advanced programmers. Given the rate of improvement of models and associated products, many experts predict that the vast majority of developers’ work will eventually be automated with LLMs - already, some expert coders like Andrej Karpathy claim 80% of their work has been automated. Any of the world’s ~30M developers who are not giving themselves a coding assistant superpower are likely at a huge productivity disadvantage.
However, many developers do not use an AI-powered coding productivity tool today. Some companies don’t want to use SaaS generally, for maintaining compliance certifications or general security risk. Other companies are afraid of IP leaking; for example, Samsung banned the use of Generative AI after it was found that three employees leaked internal source code to ChatGPT. In addition, some developers can’t use most assistants due to their bespoke legacy codebases (COBOL, anyone?), nonstandard tools, and unique workflows - the IDE, programming language, and source code management market are relatively fragmented given the varied preferences of developers and nature of enterprise codebases.
Codeium solves all of these challenges, enabling every developer to have a coding productivity multiplier that is secure & personalized, through their abilities to deploy on-premise and continuously customize to particular codebases & workflows. Codeium does this by being free for individual users, fully vertically-integrated, and working with all other developer tools agnostically. As a result, Codeium has the opportunity to become a comprehensive developer productivity platform, starting with their IDE autocomplete, code search chat, and terminal offerings, and soon expanding out to other essential parts of a developer’s workflow including code review, unit testing, & automating increasingly complex end-to-end tasks. Codeium is now consistently writing upwards of 40% of users’ code, with reports of tickets being closed twice as quickly as they were before. As LLMs continue to advance, these numbers stand to increase dramatically. Moreover, somewhat technical or even non-technical people will also be able to more easily become developers themselves.
Few founders have the wherewithal and foresight to pivot a successful startup into one pursuing a larger opportunity. When founders Varun Mohan & Douglas Chen started Codeium in 2021, it was a GPU virtualization software company called Exafunction. The founder-market fit was obvious given their backgrounds as ML infrastructure experts at Nuro & Oculus, and early customer traction was impressive. However, when the Exafunction team realized the potential of generative models directly following their Series A, they decided to use their own infrastructure to build a full-stack AI-powered coding platform, Codeium. With over 300,000 free users, 100 enterprises, and companies like Anduril using the product, we think it was the right move. By being deeply customer obsessed and iterating quickly, Codeium has learned non-obvious lessons to build a differentiated product suite. We are thrilled to partner with them and lead their Series B. If you’re interested in building the software productivity multiplier of the future, Codeium is hiring!
ー Leigh Marie