## TL;DR
AI code generation has evolved from simple line-level autocomplete to models that can understand entire codebases, fix bugs in production systems, and implement multi-file features from natural language descriptions. LLMs like GitHub Copilot, Claude Code, and Cursor are transforming software engineering by serving as always-available pair programmers.

## Core Explanation
The journey of AI code generation: (1) Statistical language modeling of code (2015-2019) -- treating code as sequential text, models like code2vec learned semantic code representations for tasks like method name prediction; (2) Pre-trained code models (2020-2022) -- CodeBERT, PLBART, and CodeT5 applied the BART/T5 architecture to code, enabling code summarization, translation, and defect detection; (3) Decoder-only Code LLMs (2023-present) -- GPT-4, Claude, Code Llama, StarCoder, DeepSeek Coder, and Qwen Coder are autoregressive LLMs trained on trillions of code tokens, enabling instruction-following code generation and agentic software engineering.

## Detailed Analysis
Key research directions: (1) Fill-in-the-middle (FIM) training -- models learn to predict missing code segments given surrounding context, enabling inline code completion; (2) Repository-level code understanding -- repo-level retrieval-augmented generation (RAG) indexes the entire codebase to provide relevant context when generating code for specific files; (3) Agentic coding -- models act as autonomous agents that read files, run commands, examine outputs, and iteratively refine code. SWE-bench evaluates this capability: agents are given a GitHub issue description and a full codebase snapshot, and must generate a correct patch. Production tools: GitHub Copilot (1.8M+ paid subscribers in 2024), Cursor (AI-native IDE, $10B valuation), Claude Code, Aider and OpenHands (open-source coding agents). Benchmarks: HumanEval, MBPP, LiveCodeBench, SWE-bench, CodeContests. The field is converging on "agentic coding" where AI actively participates in the full software development lifecycle.

## Further Reading
- SWE-bench Leaderboard: swebench.com
- A Survey on Code Generation with LLM-based Agents (arxiv 2508.00083)
- GitHub Copilot / Cursor / Claude Code / Aider / OpenHands