GPT-5.2 built a browser in 7 days. Why engineers are horrified by this "achievement."
A week ago, Cursor CEO Michael Truell announced a supposedly remarkable achievement. He claimed that, using GPT-5.2, Cursor had created a browser that ran continuously for an entire week. This browser consists of three million lines of code across thousands of files. The rendering engine was written from scratch in Rust and includes HTML parsing, CSS cascading, layout, text generation, rendering, and its own JavaScript virtual machine.
Truell noted that the browser works, albeit with caveats. It has some issues and is far from the level of WebKit or Chromium, but the team was impressed that simple websites render quickly and mostly correctly. Some developers managed to compile the code after fixing bugs, while others reported success after revising the build instructions.
Overall, however, developers aren't convinced Cursor has achieved a breakthrough. Jason Gorman, director of the UK-based consulting firm Codemanship, considers it proof that agent-based AI can scale to create non-functional software. Oliver Medhurst, a software engineer and former Mozilla employee, agrees. He noted that while working with a codebase of this size is impressive, objectively, it's not a good browser engine. Furthermore, the code is incredibly bloated—the Ladybird and Servo projects do much more, each clocking in at around a million lines.
Writing a web browser is one of the most challenging tasks for a programmer. Chromium, the open source foundation of Google Chrome, contains over 37 million lines of code. The Cursor browser, called FastRender , has around three million lines. Back in 2022, developer Joshua Marinacci wrote about how complex the web has become, to the point that only a few companies are capable of building a browser from scratch. The fact that Microsoft discontinued development of its own browser engine and migrated Edge to Chromium underscores the enormous engineering resources required to develop and support a browser.
Cursor engineer Wilson Lin, who worked on the browser's code, published a blog post explaining the project's goals: to explore how far the boundaries of agent-based coding can be pushed for projects that typically take teams months to complete. Critics have accused Cursor of heavily relying on Servo, Mozilla's open-source Rust rendering engine. However, Lin dismissed claims that FastRender is built from libraries and frameworks, stating that the JavaScript virtual machine, DOM, rendering systems, and text pipeline are all being developed as part of the project.
Gorman remains unconvinced. He points to performance metrics in the FastRender repository that reveal code instability. A build failure rate of 88 percent is extremely high and points to a broken codebase. When asked about reports of successful builds, he expressed skepticism, noting that the CI build still crashes.
Gorman is critical of claims about the success of AI coding tools in general. He cites data showing that developers greatly overestimate the impact of AI on their productivity, and most teams experience a negative impact on metrics like development time and release reliability. The minority that see modest improvements have already eliminated bottlenecks in development processes such as testing, code reviews, and integration.
Many sensational claims of AI coding success, according to Gorman, come from developers working on small tasks independently, without clients, users, or dependencies on other teams. They've driven a car to 200 mph on a straight road with no other cars and decided that faster cars mean faster traffic. Then they return to the office and demand the same rapid improvements from their teams, who are essentially driving during rush hour.
Gorman notes that when output is measured—lines of code, commits, pull requests—an increase is definitely observed. But this doesn't translate into a real increase in productivity. He points to the lack of evidence that AI tools are leading to the creation of more software, measured by the number of products in app stores, and the lack of revenue that can be attributed to these tools.
Gorman finds AI technology impressive, but often flawed. He uses it daily as a coach and mentor to understand how best to apply it. But does he consider it revolutionary? No. The principles and practices that made development teams effective before AI—small steps, short feedback cycles, continuous testing, code review and integration, modular design—remain the same. Same game, different dice.
He adds: If AI agents could truly create a working product of three million lines of code in a week, where in the design process does feedback from users and customers occur? That's where the real value is created.
A week ago, Cursor CEO Michael Truell announced a supposedly remarkable achievement. He claimed that, using GPT-5.2, Cursor had created a browser that ran continuously for an entire week. This browser consists of three million lines of code across thousands of files. The rendering engine was written from scratch in Rust and includes HTML parsing, CSS cascading, layout, text generation, rendering, and its own JavaScript virtual machine.
Truell noted that the browser works, albeit with caveats. It has some issues and is far from the level of WebKit or Chromium, but the team was impressed that simple websites render quickly and mostly correctly. Some developers managed to compile the code after fixing bugs, while others reported success after revising the build instructions.
Overall, however, developers aren't convinced Cursor has achieved a breakthrough. Jason Gorman, director of the UK-based consulting firm Codemanship, considers it proof that agent-based AI can scale to create non-functional software. Oliver Medhurst, a software engineer and former Mozilla employee, agrees. He noted that while working with a codebase of this size is impressive, objectively, it's not a good browser engine. Furthermore, the code is incredibly bloated—the Ladybird and Servo projects do much more, each clocking in at around a million lines.
Writing a web browser is one of the most challenging tasks for a programmer. Chromium, the open source foundation of Google Chrome, contains over 37 million lines of code. The Cursor browser, called FastRender , has around three million lines. Back in 2022, developer Joshua Marinacci wrote about how complex the web has become, to the point that only a few companies are capable of building a browser from scratch. The fact that Microsoft discontinued development of its own browser engine and migrated Edge to Chromium underscores the enormous engineering resources required to develop and support a browser.
Cursor engineer Wilson Lin, who worked on the browser's code, published a blog post explaining the project's goals: to explore how far the boundaries of agent-based coding can be pushed for projects that typically take teams months to complete. Critics have accused Cursor of heavily relying on Servo, Mozilla's open-source Rust rendering engine. However, Lin dismissed claims that FastRender is built from libraries and frameworks, stating that the JavaScript virtual machine, DOM, rendering systems, and text pipeline are all being developed as part of the project.
Gorman remains unconvinced. He points to performance metrics in the FastRender repository that reveal code instability. A build failure rate of 88 percent is extremely high and points to a broken codebase. When asked about reports of successful builds, he expressed skepticism, noting that the CI build still crashes.
Gorman is critical of claims about the success of AI coding tools in general. He cites data showing that developers greatly overestimate the impact of AI on their productivity, and most teams experience a negative impact on metrics like development time and release reliability. The minority that see modest improvements have already eliminated bottlenecks in development processes such as testing, code reviews, and integration.
Many sensational claims of AI coding success, according to Gorman, come from developers working on small tasks independently, without clients, users, or dependencies on other teams. They've driven a car to 200 mph on a straight road with no other cars and decided that faster cars mean faster traffic. Then they return to the office and demand the same rapid improvements from their teams, who are essentially driving during rush hour.
Gorman notes that when output is measured—lines of code, commits, pull requests—an increase is definitely observed. But this doesn't translate into a real increase in productivity. He points to the lack of evidence that AI tools are leading to the creation of more software, measured by the number of products in app stores, and the lack of revenue that can be attributed to these tools.
Gorman finds AI technology impressive, but often flawed. He uses it daily as a coach and mentor to understand how best to apply it. But does he consider it revolutionary? No. The principles and practices that made development teams effective before AI—small steps, short feedback cycles, continuous testing, code review and integration, modular design—remain the same. Same game, different dice.
He adds: If AI agents could truly create a working product of three million lines of code in a week, where in the design process does feedback from users and customers occur? That's where the real value is created.
