Anysphere

Problems, 2023. An unordered list of short-term, concrete problems. Better context: There are a lot of sources of information in a code editor: open files, semantically similar code chunks, symbolically connected classes, lint outputs, execution traces, git history, typing history, external documentation, and more. We want the model to instantly understand what is most relevant to the user’s question, and are currently training a custom and fast reranker model to solve this problem. For each request, we will gather 500k tokens from all different sources, and use our reranker to filter them down to the most relevant 8k tokens. This is both a model problem and, increasingly so, an infrastructure problem. A “copilot for edits”: While Github Copilot is tremendously helpful for eliminating low-entropy keystrokes while writing new code, it does not help you save low-entropy keystrokes when you need to make small, simple changes to existing code blocks. Think the navigation, deletion and in

Problems, 2023. An unordered list of short-term, concrete problems. Better context: There are a lot of sources of information in a code editor: open files, semantically similar code chunks, symbolically connected classes, lint outputs, execution traces, git history, typing history, external documentation, and more. We want the model to instantly understand what is most relevant to the user’s question, and are currently training a custom and fast reranker model to solve this problem. For each request, we will gather 500k tokens from all different sources, and use our reranker to filter them dow