Dive into the latest advancements in AI coding assistance and see how JetBrains AI is changing the Kotlin programming in your IDE.

Is OpenAI “all you need” today or does custom Language Models offer superior quality? What really drives effective code generation: the context window or the size of the model? And is there more to AI coding assistance than just generating code?

We'll navigate through the fundamental concepts of Machine Learning, explore relevant cloud concepts, and a lot of Kotlin to discover the answers together.

Key features demonstrated

• AI Chat Functionality: JetBrains' AI chat feature integrates well with Kotlin, offering specialized knowledge and contextual understanding.
• Code Intentions: The AI can explain, refactor, and complete code, saving time and enhancing productivity.
• Code Completion: Advanced code completion is provided, rivalling tools like GitHub Copilot but fine-tuned specifically for Kotlin.
• Commit Message Generation: AI can auto-generate commit messages, summarizing code changes effectively.

He further explains machine learning concepts and their application in natural language processing and code generation. Key points include:

• Machine Learning Basics: Classification and training functions are crucial, using examples like a Golden Retriever to explain complex ideas simply.
• Language Models: Large language models (LLMs) like GPT-4 are discussed, emphasizing their size and ability to generalize knowledge, which contributes to their effectiveness in various applications, including code generation.

Vladislav talks extensively about:

• Inference and Efficiency: Training large models is costly; hence the focus is on optimizing inference — running these models efficiently post-training.
• On-Device Models: To manage costs and efficiency, JetBrains uses local models for certain tasks, reducing dependency on external large-scale models and cutting down costs.

Lastly, he shares insights on the architectural aspects of integrating AI in development environments. The architecture involves collecting contextual data from the codebase to refine AI suggestions and completions. Fleet, for instance, implements these advanced features, leveraging both local and cloud-based models to provide accurate and efficient coding assistance