We quickly improved our training infrastructure and trained our own models, which allowed us to pull ahead in terms of quality and capabilities compared to GitHub Copilot.

We've learned a lot through the process of training this model, discovering what scales well and what doesn't.

Currently, we train bespoke models for each problem, but the goal is to develop models that can generate solutions for multiple problems, enhancing architecture search.

We realized that the existing AI models were not sufficient for our needs, prompting us to develop our own models to enhance product quality.

The AI tools we build need to continuously deliver additional value as the baseline for foundation models improves.

We didn't expect to develop our own models, but now every significant feature in Cursor involves a custom model in some way.

The current landscape shows that large model labs are focused on building powerful foundation models that require less fine-tuning, which may lead to a decline in the relevance of fine-tuning in the near future.

The introduction of OER models has significantly changed our product strategy, allowing us to build systems we previously thought were impossible.

Starting small and having a clear vision helped us validate our approach and gradually work towards our larger goals.