An experimental macOS browser that replaces traditional HTML and CSS rendering with AI-generated interpretations launched on GitHub on May 4, 2026, gaining 177 stars. The Cursed Browser asks vision language models to examine HTML and generate what they imagine the page should look like, rather than parsing markup through conventional browser mechanics.
VLMs Hallucinate Web Pages Token by Token
The browser operates on a fundamentally different principle than Safari, Chrome, or other conventional browsers. Instead of parsing HTML documents and applying CSS rules through established rendering engines, the system feeds HTML to a large language model token-by-token. The LLM interprets CSS through next-token prediction, then a vision language model generates the final pixel output based on the HTML structure.
The repository describes the result: "Every page load is a surprise. Every render is a work of art. It's better than correct, it's AI Native." The outputs often differ dramatically from standard browser renderings, sometimes producing surreal or amusing interpretations of website layouts.
Built in Swift as Native macOS Application
The project is implemented as a native macOS application written in Swift. While technically functional as a browser, the practical utility is questionable by design. The creator embraces the absurdity while exploring genuine questions about AI's potential role in web technology infrastructure.
The project represents experimental art at the intersection of AI and web technology. It asks: what if we replaced every layer of browser rendering with LLM inference? The answer demonstrates both the current capabilities and limitations of vision language models in structured interpretation tasks.
Intentionally Humorous Exploration of AI-Native Interfaces
The tone is deliberately playful, positioning the project as both technical experiment and artistic commentary. The name "Cursed Browser" acknowledges the impractical nature while highlighting how AI could theoretically replace traditional software components—even when the replacement produces unpredictable results.
The project serves as a thought experiment about AI-native applications, demonstrating that not all software components benefit from replacement with neural networks. The inconsistent, hallucinated renderings make clear that some problems are better solved through deterministic algorithms than probabilistic models.
Key Takeaways
- The Cursed Browser replaces traditional rendering engines with VLM hallucinations, released May 4, 2026 with 177 GitHub stars
- LLMs parse HTML token-by-token and interpret CSS through next-token prediction, then VLMs generate final pixel output
- Built as a native macOS application in Swift, producing dramatically different results from conventional browsers
- The project intentionally embraces absurdity while exploring questions about AI's role in software infrastructure
- Results demonstrate that deterministic algorithms remain superior to probabilistic models for structured rendering tasks