Can AI Code Completion Transform Java Development? One Engineer’s Journey

Encountering Large Language Models

Wu Ming, a Java senior engineer at Ant Group, first noticed the power of generative AI when a colleague recommended ChatGPT 3.5 for answering questions and writing code. After submitting a difficult technical question to ChatGPT, the assistant quickly provided a solution that resolved the issue for a teammate, leaving Wu Ming amazed and prompting him to explore AIGC and large‑language‑model technology further.

He began following the rapid rise of AIGC since 2022, noting its impact beyond art generation and its growing relevance to software development.

When Large Models Meet Development

In early 2023, Ant’s internal “Bailing” large model project released the CodeFuse intelligent development assistant. The first internal test in June offered a 7‑billion‑parameter model, which Wu Ming found under‑powered compared to public models with hundreds of billions of parameters. The web‑only interface also clashed with developers’ usual IDE workflow.

Four months later, CodeFuse was updated with a much larger model and an IDE plugin that supports code completion, code explanation, comment generation, code optimization, and automated unit‑test creation. The plugin allows developers to trigger predictions with the Tab key, making the assistant feel like an extension of their own coding process.

Wu Ming’s favorite feature is code completion, which can create a “flow” state where the programmer and AI seem to operate as one. He describes this as a heightened sense of focus and efficiency, akin to the psychological concept of “flow”.

However, he also notes limitations: the unit‑test generation often requires extensive manual correction, and the assistant struggles with Ant’s massive, highly customized codebase, highlighting current gaps in AI‑assisted development.

Looking ahead, Wu Ming envisions AI assistants extending beyond code writing to cover the entire software development lifecycle—requirements discussion, system design, code review, debugging, deployment, and operations—mirroring the ambition behind CodeFuse’s name, “development assistant”.

Three Tips for Using AI Development Assistants

Lower Your Expectations – Treat current models as early‑stage tools (comparable to an 8‑ or 9‑year‑old child). Accept that they will make mistakes, and use those moments as learning opportunities.

Integrate the Assistant into Your Workflow – Choose tasks where AI can naturally fit, such as invoking code completion via Tab within the IDE, so the tool becomes a seamless part of daily coding.

Be Willing to Experiment – Regularly try new prompts, explore hidden features (e.g., right‑click menus, sidebars), and even apply the assistant to non‑coding tasks like design or cross‑tool orchestration to discover unexpected value.

The Future Has Arrived

Wu Ming debates whether to teach his 8‑year‑old nephew about large models, weighing concerns that early exposure might diminish independent thinking against the efficiency gains he personally experiences when using AI for learning.

He argues that AI does not replace thought; developers must still verify and understand generated code, keeping their minds actively engaged. In his view, those who adopt AI tools will gain a competitive edge, and the technology will inevitably become ubiquitous across organizations.

He concludes that the momentum of large‑model AI is unstoppable, and encourages developers to embrace the tools now to avoid missing the coming wave.