|
Listen to this story
|
Last week, developer community platform Replit made their suite of AI products called Replit AI accessible to everyone, with default code completion and assistance features available to over 23 million developers.
The San Francisco-based startup is taking a significant step by integrating GhostWriter, a generative AI-powered ‘coding partner’ launched in October 2022, into its core platform, extending access to all its users under the banner of “AI for all.”
Replit also unveiled an upgraded version of its proprietary generative AI LLM for coding, replit-code-v1.5-3b, a cutting-edge 3B LLM with a strong focus on code-related tasks, trained on a vast dataset of 1 trillion tokens consisting of 30 programming languages and a developer-oriented subset of StackExchange.
Making AI Accessible for All
Many people, including students in India and elsewhere, lack the financial means to pay for services like GitHub Copilot, which typically costs $10 a month or similar amounts. Access to AI-driven workflows is essential because these skills will be expected by future employers or for starting one’s own business.
Replit believes in contributing to the open-source community. The goal is to ensure that AI becomes an indispensable tool, much like calculators for mathematics, as they strive to democratise AI in coding.
“AI plays an integral role in coding, and we envision a future where AI assistance is accessible to all developers, even if they cannot afford $20 a month for ChatGPT Plus,” Anshul Bhide, BizOps, India Head, told AIM in an exclusive interview.
Unlike Github Copilot or Amazon’s CodeWhisperer, their updated LLM replit-code-v1.5-3b is also open-sourced, serving as a solid foundation for anyone working on code completion models. “In terms of performance, it’s on par, if not better, than both Star Coder and Code Llama once fine-tuned,” he added.
Earlier this year, Google Cloud teamed up with Replit granting access to the former’s infrastructure and Vertex AI, promising to enhance productivity and enable complex software coding in a fraction of the time.
Compared to GitHub, Replit distinguishes itself with real-time collaboration, browser-based coding, and voice commands. The partnership underscores Google‘s efforts to compete with Microsoft in the developer community by forming strategic alliances with existing players in the AI space.
Additionally, the team used publicly available code from the developer community. “What’s interesting is that research suggests that the size of the dataset matters less than its quality,” he added. Training a smaller model on a high-quality curated dataset often yields similar results in terms of model performance. This is important as he suggested that there is a lack of quality data to train algorithms.
As far as the programming languages are concerned, Bhide said that they focused on the top 30 languages used by smart developers on our platform including Python, Java, JavaScript, C++, and others and training the model on these languages would enhance code completion and productivity.
Their proprietary LLM, the first model trained on 128 H100-80G GPUs released as open source, faced a significant challenge in obtaining these GPUs due to the ongoing chip crisis and high demand. Talking about the same, Bhide discussed the infrastructural changes needed for large-scale AI deployment, emphasising the shift from CPU to GPU usage.
This shift is driven by the increased accessibility of AI tools to a broader developer audience, resulting in a greater demand for GPU capacity for code completion and model inference. However, the challenges of securing these GPUs were highlighted, including fierce competition from tech giants and data center providers. This surge in demand, coupled with NVIDIA’s limited manufacturing capacity, has made H100 GPUs a precious and scarce resource for the foreseeable future, significantly impacting AI development and deployment.
India Expansion Plans
India is a crucial market for Replit, being the second-largest and one of the fastest-growing markets. The company’s product-market fit in India is strong, thanks to factors like the preference for mobile over desktop, fast 5G internet, and a growing interest in coding. Anshul also touched on the rise of auto-coding platforms and Replit’s role in enhancing developers’ productivity.
Within the framework of Replit’s India strategy, there is a strong emphasis on addressing latency concerns by establishing local servers. This strategic move aligns with the company’s commitment to catering to the specific needs of the Indian audience. Additionally, Replit is actively engaged in localisation efforts, including the launch of “100 Days of Code” in Hindi.
Moreover, Replit has observed the significant popularity of its mobile app among Indian students for coding purposes, even within college classrooms. This trend underscores the platform’s relevance and effectiveness within the Indian educational landscape.
“The ultimate goal is to provide access to Replit products for learning and skill development, potentially opening doors to employment opportunities and entrepreneurship on a global scale,” concluded Bhide.
