If GitHub Integrates o1, Cursor Might Be Doomed
Cursor, the viral AI-powered code editor, enhances programming with powerful features. Lex Fridman discusses its journey and future with the Cursor team.
Recently, the AI programming tool Cursor has gone viral worldwide and gained significant attention.
Cursor is a code editor based on VS Code, packed with powerful AI-assisted programming features that have excited both the programming and AI communities.
In a recent episode, renowned podcaster Lex Fridman had a tech talk with four Cursor team members, unveiling their current work and future plans.
Below is a summary of the conversation between Lex Fridman and Cursor's founding members Michael Truell, Sualeh Asif, Arvid Lunnemark, and Aman Sanger:
Origin of Cursor
What’s the origin story of Cursor?
Around 2020, OpenAI published a paper on scaling laws. It seemed like the field was making clear, predictable progress. Even without many new ideas, it appeared that more compute and data could improve these models.
By the way, we could talk about scaling laws for hours. But to summarize, one of those papers suggested that bigger models and more data lead to better results in machine learning.
It’s bigger and better, but the improvement is predictable. That’s another topic.
Right, that’s a different topic.
Yes.
During that time, we had many discussions about what this could look like. How would this technology improve knowledge workers across different fields?
There were moments when the theoretical gains predicted by that paper became very real. We started to feel that meaningful AI work was now possible without a PhD. There was a sense that a whole new set of useful systems could be built.
Another key moment was gaining early access to GPT-IV.
So, by late 2022, we began fine-tuning the model. The upgrade felt incredibly powerful. Before that, we were working on a few different projects.
Because of Copilot, scaling laws, and our previous interest in the tech, we had been refining programmer tools, though these were very specific.
We were building tools for financial professionals working in Jupyter Notebooks or using these models for static analysis.
Then, GPT-IV’s upgrade made the theoretical gains more concrete.
We realized that if we stayed consistent, we could build something beyond just a point solution.
Programming was going to flow through these models, requiring new types of environments and workflows. So, we started working toward a bigger vision.