The AI Record
And that's why we have Cursor Bench. We have this private eval set that is mostly made up of things that happen in our code base, which is held out from the … So we ensure that the models aren't trained on it. And it's based on those real world engineering tasks.

Lee RobinsonMachine Learning Engineer, Model Behavior, Cursor · Practitioner, Cursor (Anysphere)

Context. Robinson has just said public benchmarks are an imperfect test of how models are used in practice, and is describing Cursor's private evaluation set for its own models.

Talk: Recursive Model Improvement, AI Engineer (YouTube), July 15, 2026, at 8:12.

Checked against the source [ Grade B ] Said in the last year Cursor (Anysphere): $4 billion revenue, annualised, early June 2026 Sells AI tools

Filed under

Who is speaking

Lee Robinson, Machine Learning Engineer, Model Behavior, Cursor. Works on the team that trains the Cursor models he describes, so he is close to the training process; Cursor sells AI coding tools and its own models, and the talk is promotional of its approach.

Cursor (Anysphere) sells AI or agent tools or services, so it has a reason to publish flattering numbers about AI. The passage stays in the archive; read it with that in mind.

How this was checked

Checked against the source by a first reader, then confirmed independently by a second reader who fetched the source without seeing the first reader's notes (Independent review agent (Claude Sonnet), blind second pass from passage text and citation only, no repository access, 2026-11-03).

Whisper at about 08:12-08:25 reads 'we have cursor bench… held out from the evals. So we ensure…'. Captions have 'Cursor Bench'. The passage's '…' covers 'evals.' and the capital letters follow the captions.

How to cite this passage

Lee Robinson (Cursor (Anysphere)), “Recursive Model Improvement”, AI Engineer (YouTube), 2026-07-15, at 8:12. https://www.youtube.com/watch?v=q4Tr-DknG2M Archived at https://the-ai-record.com/passages/cursor-bench-private-eval-own-codebase-2026

More from the same source

So we put out Composer 2.5 in May. And it's now the most popular model in Cursor, which is exciting. And we scaled up training here quite a bit by generating more RL environments, trying out some new methods for learning. And also just making more ambitious problems for the models to solve. And the results have been pretty promising so far.

Lee RobinsonMachine Learning Engineer, Model Behavior, Cursor · Practitioner, Cursor (Anysphere)

Context. Robinson is describing Cursor's own model, Composer, which Cursor trains in-house.

Talk: Recursive Model Improvement, AI Engineer (YouTube), July 15, 2026, at 2:41.

Checked against the source [ Grade B ] Said in the last year Cursor (Anysphere): $4 billion revenue, annualised, early June 2026 Sells AI tools Passage page

The last bit here is that the model is learning to train the next model. And it's a little hard to wrap your brain around. The way I like to think about it is every time you release a new version of this intelligence, then you can create or distill these derivative versions that you use to speed up other parts of the training process, …

Lee RobinsonMachine Learning Engineer, Model Behavior, Cursor · Practitioner, Cursor (Anysphere)

Context. Robinson is closing a talk on how Cursor trains its models and uses earlier models to speed up later training.

Talk: Recursive Model Improvement, AI Engineer (YouTube), July 15, 2026, at 18:30.

Checked against the source [ Grade B ] Said in the last year Cursor (Anysphere): $4 billion revenue, annualised, early June 2026 Sells AI tools Passage page

Just to kind of put an example to this, we've created these tools and these systems where researchers can run experiments directly from Slack. We … avoid this state of being bottlenecked on humans launching, and reviewing, and babysitting runs. And we actually have an entire team just working on automating every part of the research work that isn't freeing up the researchers' time to work on their most ambitious ideas. So every person on the ML team gets access to this fleet of agents they can basically train models directly from Slack.

Lee RobinsonMachine Learning Engineer, Model Behavior, Cursor · Practitioner, Cursor (Anysphere)

Context. Robinson is describing how Cursor's machine learning team uses agents in its own model-training work.

Talk: Recursive Model Improvement, AI Engineer (YouTube), July 15, 2026, at 17:14.

Checked against the source [ Grade B ] Said in the last year Cursor (Anysphere): $4 billion revenue, annualised, early June 2026 Sells AI tools Passage page

And in reality, the vast majority of our revenue today comes from agent usage. And that means that all of the data inside of Cursor is also coming from agent usage, and we can use that to train better models.

Lee RobinsonMachine Learning Engineer, Model Behavior, Cursor · Practitioner, Cursor (Anysphere)

Context. Robinson is explaining where the data for training Cursor's models comes from. The talk gives no revenue figure.

Talk: Recursive Model Improvement, AI Engineer (YouTube), July 15, 2026, at 4:41.

Checked against the source [ Grade B ] Said in the last year Cursor (Anysphere): $4 billion revenue, annualised, early June 2026 Sells AI tools Passage page

Something wrong? Report a correction. JSON: /api/v1/passages/cursor-bench-private-eval-own-codebase-2026.json