Augment Code

Hello, world! We’re Augment Code, the first developer AI purpose-built for teams. What makes Augment different? Context. Every Augment feature is context-aware. Read more: https://bit.ly/4fdJ1UU

Code Instructions: Make quick changes and additions to your code with natural language.

Kong's development team shares how Augment helps them get up to speed with the codebase and stay in the flow.

Experience Augment's sub-100ms code completions through the eyes of Faruk Sahin, VP of Engineering and Product at Pocketlaw. Discover how the combination of speed and accuracy creates a seamless coding experience that "feels instant to the human eye."

We've achieved a significant breakthrough in AI code completion by pioneering a new approach that learns directly from natural coding workflows. Here's why this matters for the future of developer tools. The Challenge: Traditional reinforcement learning approaches rely heavily on human annotation and explicit feedback, which becomes impractical when dealing with large codebases. Consider this: it might take a human annotator an entire day just to understand enough context to label a single example in a million-line codebase. Additionally, public repositories rarely contain the incomplete, work-in-progress code states that developers regularly work with in their IDEs. Our solution, RLDB, takes a fundamentally different approach: 🛠️ Data Infrastructure: We've built a system that captures full repository content and IDE states at regular intervals, enabling context-aware reward model (RM) and reinforcement learning (RL) training. 💻 Algorithm Design: Rather than relying on traditional ranking-based approaches, we developed a custom RL algorithm that optimizes towards the distribution of real-world coding tasks. This allows our reward model to train on 100x more data than traditional comparison pairs. The improvements we've seen are substantial: 1️⃣ Performance boost equivalent to doubling model size or using 10x more training data 2️⃣ 8%+ perplexity reduction over traditional RL methods 3️⃣ Better handling of incomplete code states 4️⃣ Reduced hallucinations and repetitions in suggestions Most importantly, these improvements came without requiring developers to change their workflow or provide explicit feedback. The system learns from natural coding patterns, making it scalable and non-intrusive. #AI #AIforCode #DeveloperTools #AIResearch

Imagine building software with an AI that deeply understands your complex codebase. Join a live demo of Augment Code to see how it helps with onboarding, testing, and refactoring that fit your stack. Register here: https://lnkd.in/gwC6iMG8

Thrilled to announce a major step forward in our mission to advance open source development: Augment Code is now free for open source contributors and maintainers. What excites me most about this initiative is how it addresses two fundamental challenges in open source:  1. The steep learning curve for new contributors trying to understand large codebases  2. The time-intensive back-and-forth between maintainers and contributors on PRs We've seen remarkable results already. Jonathan Ellis, Apache Cassandra committer and DataStax CTO, reports a 50% hit rate in understanding Cassandra's complex codebase using Augment - a breakthrough for developer productivity in large-scale open source projects. What makes this different? Full access to Augment's capabilities - no rate limits, no outdated models. Contributors get comprehensive codebase understanding and intelligent suggestions that align with project patterns and practices. This means faster onboarding, more meaningful contributions, and less review friction. Sign up at augmentcode.com/opensource. Happy Holidays!

Jan Baryła, Engineering Manager at Handoff - AI Estimating for Remodelers, demonstrates how Augment Chat transforms the way developers understand and work with unfamiliar code. Learn why Augment's owned architecture and intelligent code navigation are reducing cognitive load for development teams.

Working on open source is hard. We just made it a little easier. Augment is free for open source contributors and maintainers. 🎉 Augment’s deep codebase understanding tears down some of the biggest pains for working on open-source projects– the context you need to contribute and the help you need to make quality PRs. Sign up at augmentcode.com/opensource

The secret to efficient LLM inference: rethinking how we batch requests We discovered something fascinating while optimizing our inference stack: traditional request batching is leaving massive GPU potential untapped. Even when batching 10 parallel decoding requests together, you're typically using just 2% of your GPU's FLOPS. We knew there had to be a better way. Our solution? Let decode tokens "piggyback" on context processing. Instead of traditional request batching, we: - Mix tokens from multiple requests in the same batch - Stay FLOPS-bound whenever possible - Optimize for real-world developer workflows The results speak for themselves: ✨ Higher GPU utilization ⚡ Lower latency 📈 Better cost efficiency 🎯 Faster response times The academic world calls this approach "chunked prefill." We call it the key to achieving deep context with low latency. Here's how we did it: https://lnkd.in/gVT5kQRg #MachineLearning #GPUOptimization #AI #Engineering #Innovation