Makes model research inspectable through annotated code and training reports, from a Transformer implementation to the training and evaluation of a coding model.
Why read Sasha Rush?
Rush offers a route from using a coding model to understanding how one is trained. His Composer 2 introduction explains continued pretraining followed by reinforcement learning in coding environments. The coauthored technical report adds experiments and explains why the team's evaluation includes ambiguous requests and changes across several files.
For the underlying architecture, read his Annotated Transformer. It connects the original paper's equations to a line-by-line implementation, then demonstrates a copy task and translation. Some code dates from 2018; the page links an updated implementation. Familiarity with Python, tensors and basic neural networks will help.
Start with the original
Selected work
Research explainer ·
A technical report on Composer 2
How is a coding model trained for an agent's work? Rush introduces continued pretraining, reinforcement learning in Cursor sessions and evaluation on tasks from the team's own engineering work. The article leads to the full team report.
Paper ·
Composer 2 Technical Report
What evidence supports a training change? This coauthored Cursor Research report connects the recipe to ablations, agent behavior and the design of CursorBench. Read the benchmark conditions alongside the team's reported results.
Tutorial ·
The Annotated Transformer
How do the Transformer's equations become code? Rush's annotated implementation of Vaswani and coauthors' architecture covers attention, masking, training and translation, with help credited to Vincent Nguyen and Guillaume Klein.
Projects & roles
The Annotated Transformer
An annotated implementation of the original Transformer architecture; Rush's original article links the project.
Composer 2 research
Cursor Research's coauthored account of training a model for agentic software engineering, including Rush among the authors.
Community rating
Sources & further reading
- His current site identifies his Cursor research, lists his earlier Harvard, Cornell and Hugging Face roles as past, and links @srush_nlp.
- Cursor's original article carries Sasha Rush's dated byline and links the technical report.
- The author record includes Alexander M. Rush, the name used in his academic publications.