Replit integrates GhostWriter into the core platform to provide comprehensive open source AI development tools
Development tools provider Replit recently announced a series of new initiatives aimed at bringing artificial intelligence to all developers.
Replit's cloud software development platform is widely used and claims to have more than 20 million users. Over the past year, Replit has gradually built generative AI capabilities for users, including the GhostWriter AI code completion tool and a partnership with Google. Until now, GhostWriter's use has been limited to a subset of Replit's users, but that's about to change.
As a major move on October 9, 2023, Replit directly integrated GhostWriter into its core platform and made the generative AI code completion tool available to all users, naming it "AI for all" (AI for all). Along with the GhostWriter integration, Replit also announced a new version of its next-generation open source Large Language Model (LLM) for programming, called replit-code-v1.5-3b.
Replit's open source coding LLM is positioned as a competitive alternative to StarCoder LLM, jointly developed by ServiceNow and Hugging Face, and Meta's Llama CodeLlama7B.
In recent years, a large number of AI-driven coding tools have emerged to help developers write code. Microsoft has its GitHub Co-pilot suite of services, Amazon has Codewhisper, and even Stack Overflow has joined the OverFlow AI effort. Masad believes that tools like GitHub Co-pilot are just add-ons to existing development tools. "We don't think that's the way forward," he said. "We think AI needs to be truly integrated into every programming interaction you have and needs to be part of the default experience with Replit, and I'm sure the same will be true with other products in the future."
With this new integration, Replit is completely ditching the GhostWriter name and instead making AI a core feature that will be enabled for all users. Masad said Replit has users around the world who code on a variety of devices, including laptops and even mobile phones, and now all of these users can become AI-enhanced developers. "We think this is going to be the largest deployment of AI-enhanced coding in the world, and we're going to burn GPUs as much as we burn CPUs, so keep your fingers crossed for us," he said.
Replit's generative AI capabilities are not just wrappers for other vendors' LLMs, but are based on open source technology built by the company itself. "Our code completion capabilities are powered by our own specialized large language model," Replit AI VP Michele Catasta explained during a live broadcast of the AI Engineer Summit. "We build on open source code published on GitHub and code developed by the Replit user community. train."
In May of this year, Replit released its replit-code-v1-3b LLM and is now launching an updated version of replit-code-v1.5-3b, which significantly expands the performance and scope of the LLM.
Catasta noted that the model update was trained on 1 trillion code tokens and supports 30 different programming languages. According to Catasta, the “secret ingredient” of this new LLM is Replit’s work on data processing. At the same time, another key factor in enabling this model is the powerful hardware used by this LLM during training. "We trained on 128 Nvidia H100-80G GPUs, which are as scarce as gold these days," he said.