Summary
Highlights
Cursor, a new AI Frontier Lab, has announced a brand new Opus-class model, directly competing with the latest Opus and GPT models. This model is trained from scratch with 10 to 20 times more compute than previous models, thanks to Cursor's acquisition by SpaceX. This acquisition strategically positions Cursor alongside XAI and the Colossus Supercluster, providing immense computing power. This fusion resolves a significant bottleneck in AI model development, turning a weakness into a superpower for both Cursor and XAI/SpaceX.
Anthropic has demonstrated the compounding advantage of focusing on the best coding models, gathering user data, and continuously improving. This has led to a brain drain from Google DeepMind, with prominent figures like John Jumper and Nom Shazir leaving for Anthropic and OpenAI, respectively. Despite efforts from Sergey Brin to improve Google's coding AI models, Google and XAI have struggled to catch up to the coding abilities of OpenAI and Anthropic, highlighting the difficulty in closing this gap even with vast resources.
Cursor, despite its large user base, faced challenges due to its strong reliance on expensive Frontier AI labs like OpenAI and Anthropic for API usage. To overcome this, Cursor developed its own Composer models, which were cheaper and competitive. Their Composer 2.5 model, based on the open-source Chinese model Kimmy K2.5, used additional training and reinforcement learning techniques. This success paved the way for their new model, trained from scratch with significantly more compute, aiming for lower inference costs and better workflow tuning.
Cursor's approach to training models is unique and effective. They've found a way to address the credit assignment problem in reinforcement learning, where models struggle to identify specific errors in long rollouts. Their method involves 'targeted textual feedback,' where a teacher model provides precise hints at the exact point of mistake in a trajectory. This allows the student model to learn and adapt more efficiently, leveraging Cursor's extensive user data for highly effective training.
Cursor also introduced Origin, a new platform designed as 'Git for AI agents,' serving as an alternative to GitHub. Origin is built from the ground up to accommodate agentic AI workflows, addressing GitHub's recent reliability issues caused by an unexpected surge in AI agent usage. This development by Cursor's co-founder, Michael Tru, highlights a forward-thinking approach to AI development infrastructure, anticipating the increasing role of AI agents in coding.
The success of Cursor's new model will indicate whether it's possible for a company to rapidly catch up and even lead in the coding AI space despite not starting as a frontrunner. Cursor's evolution from building on open-source tools and relying on other AI models to developing its own foundational technology is a significant achievement. Elon Musk, through SpaceX and XAI, continues to play a pivotal role, not only by providing immense compute power to Cursor but also potentially becoming a 'compute landlord' for future AI data centers, including those in space, as envisioned by Google's project SunCatcher.