Sualeh Asif says he worked on a search-engine-type company with a friend in college, which fed his interest in language models.
“back in college, sort of, I had, I had gone on and, uh, worked on a sort of search engine type company with a friend.”
Person
Sualeh Asif
1 episodes · 21 quoted claims
Sualeh Asif says Cursor co-founder Michael had spent considerable time building high-performance games.
“I think Michael had spent quite a bit of time building these, uh, high performance games”
Sualeh Asif says the Cursor founders did competitive math in college, not only competitive coding.
“I think we're not only competitive coders, like we did competitive math in college.”
Sualeh Asif says Cursor co-founders Amon and Arvind were the last to move from Vim to VS Code, triggered by GitHub Copilot.
“Amon and Arvind probably were the last to switch over from from Wim to VS Code. And and the trigger there was was GitHub copilot,”
Cursor built about three agent prototypes before shipping its agent, because the models would lose track of the task.
“and we had probably done three agent prototypes before that. We didn't ship because some version of the the model would just lose track”
Sualeh Asif says Cursor's custom tab model handled about 100 million requests per day and was growing quickly.
“For example, for our custom tab model, we do something like 100 million requests a day, uh, and and and quickly growing it”
Cursor trains a large first version of its apply model, deploys it to users, and uses the resulting data to distill smaller, faster models.
“you train your 1st version of an apply that is quite a bit bigger. And, uh, you then deploy it for all users. You get lots and lots of data and then you can distill a slightly smaller model”
Sualeh Asif says Cursor's apply feature is a model rewriting the entire file rather than a deterministic algorithm, and that improvements came largely from making the model smaller.
“It's it's a model that is actually rewriting the entire file and a lot of, you know, a lot of the improvements have been making the model smaller.”
Sualeh Asif says early language models used context windows of about 4,000 to 8,000 tokens, which limited what a coding agent could do.
“when the models started out, you were sort of doing these 4k, 8k context windows.”
Inside Cursor's code base, the Cursor Tab feature is named CPP, short for 'copilot plus plus'.
“cursor tab in our code base is named CPP for being copilot plus plus”
Sualeh Asif says Cursor tries never to store user code beyond the life of a request, though it hopes to store some code privately in future.
“we try to make sure that the that we never store any code past the life of your request.”
Cursor's code-indexing infrastructure runs at a scale of billions of files per day, using Cursor's own inference for embedding models.
“the one that we've been sort of very public about is our indexing infrastructure that, uh, we spent a lot of time optimizing and running at quite enormous scales, like billions of files, uh, per day kind of infrastructure”
Cursor is working on an index architecture with a shared 'trunk' of a code base plus per-user branches, to avoid thousands of duplicate copies.
“can we have a shared trunk? And then, uh, every single person can have their sort of branch off that trunk?”
Sualeh Asif says Cursor has a fixed number of GPUs at any moment, which limits its token capacity and forces trade-offs in indexing.
“we at any point in time have a fixed number of GPUs, which means we have a fixed num amount of token capacity.”
Indexing uses a small share of Cursor's GPUs, with the tab model and larger models using far more compute.
“we run the tab model, like indexing is a very small percentage of, uh, of our GPUs.”
Sualeh Asif says Cursor's tab model serves hundreds of millions of calls per day.
“so tab model is like hundreds of millions of calls per day”
Sualeh Asif says Cursor chose DeepSeek models as a pre-trained base for its own post-training and reinforcement learning stack.
“for just picking a really well pre trained base, uh, DeepSeek does a phenomenal, phenomenal job.”
Sualeh Asif says reinforcement learning and post-training have made models much more coherent over tens of thousands of output tokens, which helps agents.
“one of the things that is talked about less is how coherent the models have gotten over producing tens of thousands of tokens of output”
Cursor uses an internal prompt-building library, called Preempt in the transcript, built to suit its own needs.
“we use this internal library called Preempt Um, and the way we built Preempt was it was well suited to our own needs at the time.”
Sualeh Asif says Cursor's evaluation of its agents was largely vibes-based at the time, though he expects more daily metrics later.
“it's, it's pretty, pretty vibes based. I suspect it'll be vibes based in the short term”
Sualeh Asif recommends reading the SQLite code base as an example of well-written software by a very small number of people.
“SQLite, for sure, also if you haven't read SQLite, um, again, very well written”