Code models and AI for programming

In short: Language models write, explain and fix program code because they were trained on huge amounts of code. From a suggestion in the line to an agent that completes whole tasks in the project.

In more detail: Code is cheap for models: it has fixed rules and the result can be checked automatically (tests, compiler). That makes reinforcement learning with verifiable results possible.

In detail

• Levels: autocomplete → chat in the IDE → agent that reads files, runs commands and runs tests.
• Training: public code (with licence questions), tasks with tests, reinforcement learning on passed tests, see Reasoning models and chain of thought.
• Measuring: HumanEval, SWE-bench (fix real bugs in open-source projects).
• Typical errors: invented functions or libraries, outdated APIs, security holes, seemingly right but wrong code.
• Good practice: read and understand code, write tests and run them, small steps, version control (git), never secrets in prompts.
• Learning: beginners should not copy code blindly but have it explained and rebuild it.
• Legal: licences of training code and copyright are disputed, see Training data, copyright and privacy in AI.

See also: Tool use, agents and MCP, Agent design patterns