Introducing Mercury Edit 2

Inception's Mercury Edit 2 predicts a developer's next code change from recent edits and surrounding code. Its training account describes a problem with an initially capable model: it suggested changes too frequently and made them too long, distracting users.

The team then used accepted and rejected suggestions as preference-training data. Inception reports higher acceptance and more selective suggestions after this work. The useful design principle is that edit prediction must learn when a suggestion is welcome, alongside what code to propose. A model that understands the next change can still interrupt the developer if it offers too much.