January 2025 Summaries

Anysphere, an applied research lab focused on automating coding, has announced securing $105 million in Series B funding from prominent investors like Thrive Capital and Andreessen Horowitz to further its mission of creating a more effective human-AI programmer. The company aims to revolutionize programming by expanding its team and investing in frontier research, targeting real-world scalability. Anysphere's product, Cursor, is already in use by millions of programmers and boasts proprietary models that generate and edit significant volumes of code, contributing to over a billion characters edited daily. The company, which has surpassed $100 million in recurring revenue, seeks individuals interested in advancing coding automation through innovative research, design, and engineering.

Fusion, the next-generation Cursor Tab model, represents a significant leap forward in code editing by swiftly predicting edits near the cursor and suggesting subsequent navigation points, known as "jumps," thereby reducing the tedium often associated with coding. Since its inception in March 2024, the Tab model has evolved through numerous updates, leading to a substantial increase in both speed and intelligence, with the current Fusion model offering over a billion edited characters daily and a 100-fold growth in request rates. Fusion enhances upon its predecessor by accurately predicting over 25% more difficult edits per line and extending the length of suggested changes by more than tenfold, all while significantly reducing latency and offering instant, higher-quality cursor jumps with improved context awareness. These improvements are due to cleaner data, extended context windows, and careful training for larger edits, with the model now outperforming many language models in generating code. Fusion's rollout in client release 0.45.0 promises further enhancements in codebase context and integration with Supermaven technology, marking a new era in efficient and intuitive code editing.

Character prefix conditioning is an algorithm designed to improve code completion by enabling language models to sample token sequences based on a character prefix, rather than a token prefix, which addresses issues arising when a user's cursor is not on a token boundary. This approach is necessary because language models typically process sequences of tokens, and naive tokenization can produce incorrect results if the cursor is misaligned. The algorithm involves sampling sequences from a distribution defined by an autoregressive model, ensuring that the sequence starts with the specified character prefix. The challenge lies in constructing an efficient algorithm for sampling from this conditional distribution, minimizing calls to the language model, and the blog invites readers to contribute solutions.