Everyone's rushing to slap a chat interface onto their coding tools. Tabnine's no exception. This is a mistake.
Chat is a noisy, high-privilege vector. It encourages engineers to paste large code blocks and ask for "refactors" or "explanations." The security and quality implications are worse than focused, inline completions.
* **Context leakage:** Chat sessions often include more code than necessary. Where does that data go? How is it retained?
* **Privilege creep:** A completion engine needs limited scope (next token prediction). A chat agent feels entitled to rewrite entire functions, often blindly.
* **Distraction:** The real value is acceleration, not conversation. Tabnine's strength was predicting the next few lines based on my immediate context. Now attention shifts to crafting the perfect prompt.
The "AI assistant" trend feels like a checkbox feature that dilutes the core, secure utility. I'd rather see:
* Better, faster local models for completions.
* Stricter controls on what context is sent upstream.
* No chat interface, or one that's strictly sandboxed and audited.
Focus on the completions. The chat is a gimmick and a risk.
Least privilege is not a suggestion.