Promising Problems on AI Coding from Cursor

Problem of building cursor

• bettter context
  • Context: open files, semantically similar code chunks, symbolically connected classes, lint outputs, execution traces, git history, typing history, external documentation, and more.
  • training a custom and fast reranker model to make model to instantly understand what is most relevant to the user’s question
  • For each request, gather 500k tokens from all different sources, and use reranker to filter them down to the most relevant 8k tokens
• A “copilot for edits”
  • need innovation in both UX and on the model-side
• Constrained, in-flow agents
  • make an agent works on folders of a few hundred thousand tokens
  • scale it up to work for entire codebases.
• Bug-finding
  • always be passively scanning your files to find potential bugs for yo
  • actively look for the bug with your help. There’s a lot of interesting data collection to be done here.
• Larger edits
  • model needs to be smart enough to pick out the parts to modify without rewriting everything
  • the changes need to be shown in a parsable, real-time form.
• Scale
  • have already built a really fast Merkle-tree-based codebase syncing engine in Rus
  • custom indexing system

Other problems from cursor

• Next Action Prediction

  • it's called cursor flow, predict the user's next edit and allow user to update it seamlessly

  • Directions:

    • Fundamental research on action prediction across a codebase.
      • Continued pre-training and post-training on ~5-13B active parameter code-models (for prefill-bound low-latency predictions).
      • Additional inference tricks similar to Speculative Edits
      • Clever UX for surfacing "actions" in a non-obtrusive way.

• Multi-File Edits

• Optimal Context

  • underexplored research direction.

• Multi-hop Context

  • multi-hop embbeders and rankers
  • customized attention for codebases
  • Embbed codebase into weights so that we can use model as a search index

• Bug Detection and Debugging

  • Problems: are plagued by false-positives
  • Identifying the trickiest bugs require a deeper understanding of the codebase
  • Benefits of AI Review
    • user is more tolerant of false-positives
  • AI Linter
    • a cheaper, faster model than AI-Review
    • must be low false-positive rate.
  • Smarter Debugging
    • built a cursor/debug package, tracks runtime information.
    • Clever dataset curation (likely synthetic data) and RL on frontier code models to improve calibration.
    • Infinite context and near-perfect codebase understanding.

Reference

1. Our Problems from Cursor team https://www.cursor.com/blog/problems-2023
2. More Problems from Cursor team https://www.cursor.com/blog/problems-2024