The Problems We'll Solve (Continued)

Cursor pushes the existing Copilot++ one step further, envisioning a next action prediction model that predicts not just the next edit, but the next cursor movement, the next file switch, and even the next terminal command. The goal is to pick up on the user's intent across the entire editor with almost no delay, so the model can immediately execute the next action.

The core direction splits into two axes.

• Cursor Flow: continuously predicts low-entropy actions like tab and keystrokes, creating the next flow of where the user will move and edit.
• Using more inference-time compute to generate larger and better perfect edits in the background.

The important problem here isn't simple accuracy but a flow-preserving experience. The direction proposed is one where the user only needs to write pseudocode, and the model expands it into a real implementation in the background, or handles changes across multiple files at once.

Another major axis is optimal context. It aims to tie together scattered information—code, documentation, commit history, logs, UI pixels, even Slack messages—through Retrieval, recurrence, and long-context attention, treating it like a single infinite context engine.

In particular, multi-hop context that doesn't end with a single search is seen as important. For example, after finding computeDiff, you then need to find how ITextModel is created, so a single embedding search isn't enough—a model that predicts the next hop is needed.

Finally, for bug detection and debugging, it proposes AI linting with reduced false-positives, better AI review, and an extension of the cursor/debug package that collects runtime information. In summary, Cursor is laying out the next set of problems aimed at unifying the editor, search, review, and debugging into a single predictive system.