Cursor 2.0 might have me coming back to Cursor

Detailed Summary

Introduction and Industry Shift (0:00 - 2:07)

The AI coding sector has seen rapid changes, with companies like Augment Code and warp.dev adjusting their pricing due to high inference costs. Warp.dev, for instance, now offers a $20/month basic plan with a 'bring your own key' model. This shift is driven by the realization that companies not owning their inference and infrastructure face unsustainable costs. Cursor's release of Composer 1 and Windsurf's SWE 1.5 model are significant as they represent a move towards proprietary models. SWE 1.5 is a 950 token per second model, currently free, while Composer 1 is a 250 token per second model, believed to be an upgraded version of Cursor's 'Cheetah' model.

The Strategic Importance of Owning Models (2:08 - 4:43)

The speaker emphasizes that the future of AI coding lies with companies that own their models and infrastructure, allowing for fixed monthly pricing. Cursor's development of Composer 1 is a strategic move to reduce its dependency on Anthropic and OpenAI, especially after past incidents where access to third-party models was cut off. The training data for Composer 1 is a point of curiosity, with the assumption that Cursor is building it themselves. While third-party models like Anthropic and OpenAI might still be superior, the gap is narrowing, and proprietary models offer sufficient quality for a large percentage of users. The speaker expresses frustration with 'rug pulls' by companies changing pricing but understands the underlying economic reasons.

Speed and Practical Demonstrations (4:44 - 10:00)

The speed of Composer 1 and SWE 1.5 is a major highlight. A text-based adventure game was generated in 55 seconds using Composer 1, demonstrating its rapid output. The speaker notes that while AI-generated code might not be production-ready immediately, its speed allows for real-time ideation, prototyping, and collaborative development. Composer 1 is preferred over SWE 1.5 for its slightly better performance. Examples include:

• Text-based adventure game: Generated quickly and fully functional.
• Calendly clone: Composer 1 successfully set up a server and created a functional clone with event types and scheduling, while SWE 1.5 produced a simpler version without sharing or booking capabilities.
• Digit recognition: Composer 1 created a Python backend that recognized digits using pre-trained data, an impressive feat. SWE 1.5's version required training and had issues with recognition.
• Personal finance app: Composer 1 generated a visually appealing and functional app with budget, savings, reports, and investments. SWE 1.5's version was aesthetically pleasing but lacked functionality like editing or deleting entries.

Further Demonstrations and Model Comparison (10:01 - 14:30)

The text-based adventure game generated by SWE 1.5 was also functional, though the speaker felt it was slightly weaker than Composer 1's output. Both models are considered solid and not 'slouches.' The speaker used SWE 1.5 in his codebase and found it acceptable, though not as impressive as Composer 1. Composer 1 has a 200k context window, which can be quickly consumed due to its high speed, leading to frequent auto-compaction by Cursor. Composer 1 also successfully worked in 'plan mode,' a feature that Cheetah previously lacked. Cursor 2.0 introduces a new UI with an agents and editor toggle, allowing users to run multiple models (up to seven) simultaneously with work trees, which is powerful but can be expensive.

Conclusion and Future Outlook (14:31 - 18:07)

The speaker expresses strong approval for Cursor's direction, particularly its move towards owning its models. This independence from Anthropic and OpenAI is seen as crucial for Cursor's long-term relevance, countering predictions of its obsolescence. While Anthropic and OpenAI models are still considered superior in some aspects, the margin is thin and depends on the coding task. Composer 1 is highly recommended for front-end, Express NodeJS, and Python development. The speaker acknowledges that Composer 1's cost is comparable to GPT-5, but its speed justifies the price for competent tasks. The video concludes by inviting audience feedback on the new models, their performance, and the implications of using potentially Chinese-based models, which some companies might be hesitant to adopt.

Notable Quotes

"Honestly, I feel like the AI coding game has just been flipped on its head the last probably month." "My general sentiment is the game has changed and it's changed to where fixed monthly pricing only is going to work if you own the model and the infrastructure." "The companies that have their own models and their own infrastructure to actually host these models, that can do their own inference, are ultimately going to be the ones to win. But only if the models are good." "Speed matters a lot." "I've been coding with the composer one model more than the 31.5 model because I do think it's slightly better." "I think the likelihood of that happening is near zero now especially now that they're even if they are fine-tuning a Chinese model it doesn't matter. They own the model now." "The anthropic models and OpenAI models still are superior, at least in my early testing. But the margin is thin and it's probably going to highly depend on the type of code you're doing." "To me, I think this just it it really does change my perspective on Cursor and Windsurf to some degree now that they're they seem to be hammering in on having their own models."