Devin by Cognition: When AI Becomes Your Best Developer | LangChain Interrupt
Watch our recorded sessions from Interrupt here: https://interrupt.langchain.com/video/?utm_medium=social&utm_source=youtube&utm_campaign=q2-2025_interrupt-2025_co
Russell Kaplan, President of Cognition, reveals the technical architecture behind Devin and introduces Devin Search -- their tool for deep research across your codebase -- and how they built it. Their specialized 32B model outperforms frontier models on coding tasks through custom post-training and reinforcement learning.
What You'll Discover:
- Why Devin is fundamentally different from Copilot
- The exact architecture behind Devin's codebase understanding system
- How their 32B model achieves 91% accuracy on CUDA kernels
- The "narrow domain specialization" approach that beats frontier models
Watch our recorded sessions from Interrupt here: https://interrupt.langchain.com/video/?utm_medium=social&utm_source=youtube&utm_campaign=q2-2025_interrupt-2025_co
Russell Kaplan, President of Cognition, reveals the technical architecture behind Devin and introduces Devin Search — their tool for deep research across your codebase — and how they built it. Their specialized 32B model outperforms frontier models on coding tasks through custom post-training and reinforcement learning.
What You’ll Discover:
– Why Devin is fundamentally different from Copilot
– The exact architecture behind Devin’s codebase understanding system
– How their 32B model achieves 91% accuracy on CUDA kernels
– The “narrow domain specialization” approach that beats frontier models
