Code generation is probably THE MOST VALUABLE application for Generative AI so far [1].

Seems everyone is building the same thing just from different initial conditions.

• Cursor: real-time synchronous copilot in the IDE
• Devin by Cognition: long-running async agents in the background

Cursor offers a differentiated copilot product experience. It’s built as a standalone IDE on top of VSCode, whereas Github Copilot is a VSCode extension. This allows Cursor to do “multi-line edits”, while Github copilot can only “insert”.

From my initial testing of Replit agents, it was able to solve one of our earlier front-end interview questions in just ~6mins [video]. The catch is that this problem is quite standard and likely GPT-4/Claude has seen it as part of its training dataset. But Replit agents do seem to fall somewhere between completely synchronous copilot and mostly asynchronous coding agents.

The future will be both: copilot helps boost your “IC” productivity, whereas coding agents allow you to delegate simple tests. Overall, we’ll be able to see 10-100x boost in terms of productivity output.