【AI Industry Alert】GPT-5.2 has not even been released for 7 days, and the AI coding tool Cursor has already crashed hard. A developer exposed that Cursor’s so-called AI browser project is nothing but AI slop, completely unusable. Cursor’s reputation has taken a massive hit.
Just a few days ago, a bombshell dropped in the AI community about Cursor.
Here is what happened:
The truth is far more shocking than anyone imagined.
Millions of tokens burned like water in a single night. The AI coding tool that the entire industry was chasing has become a complete joke.
Whether it is an operating system, office software, or a game engine, as long as you give AI enough prompts, it can give you a “finished product.” Right?
ai hentai chat
cuckold chat
But when a user actually downloaded the code and tried to run it, what they found was a pile of broken garbage.
When someone carefully reviewed Cursor’s open source browser project code, they discovered a shocking truth:
All the so-called AI-generated code, from the core logic to the basic functions, completely failed to compile.
In a detailed technical analysis, the author bluntly pointed out:
Cursor’s so-called breakthrough project is actually a piece of broken logic and missing core functionality, what the industry now calls “AI Slop.”
On the surface, it looks like a beautiful demo video, with smooth animations and professional presentations, making everyone think this project actually works.
But in reality, it is completely unusable garbage.
Project link: https://embedding-shapes.github.io/cursor-implied-success-without-evidence/
GPT-5.2 Has Not Even Been Released for 7 Days, and It Is Already Fake
To understand the full picture, we need to go back to a blog post that Cursor released just last month.
First, let us look at what Cursor’s research team actually did.
On June 14, they published a blog post titled “Scaling long-running autonomous coding.”
Official blog link: https://cursor.com/blog/scaling-agents
In this blog post, the team used a large number of data charts, experiment records, and so-called success cases. The stated goal was:
To explore the boundaries of what AI agents can do, and to prove that some tasks that normally require large engineering teams and long development cycles can now be completed by AI alone.
However, Cursor’s research team deliberately avoided discussing some key issues: the real reasons for failure, and the so-called success cases.
In fact, the tasks the team claimed to have completed were mostly coordination problems that occur within teams, rather than a single AI completing a real project from start to finish. The model’s expansion capability was severely overstated.
More critically, the project they released was an extremely misleading result.
To prove that AI can build an operating system, they set an ambitious goal: starting from scratch, building a browser. This browser needed to understand a set of web standards, write out 100,000 lines of code from a list of 1,000 files.
At the same time, they released the open source code on GitHub.
GitHub project: https://github.com/wilsonzlin/fastrender
On the surface, this looks like a successful case of AI writing code.
But if you do not look closely at the success metrics, you will find a huge gap between reality and the claims.
The blog post says: “We can create a new worker, assign it a task, and then cancel the progress of other workers, and let the same worker handle conflicts.” But in fact, this has never been said before, and there is no real success.
The so-called success only refers to a fixed 8-minute demo video.
And what was written in the code?
Although it looks like a simple browser, anyone who has tried to build one knows it is extremely difficult.
From start to finish, the code never mentions real engineering challenges like memory management, security, or multi-threading.
It Is All Fake, and Many People Cannot Compile It
What is even more shocking is that besides the README and demo screenshots, the so-called complete project has no real code that can run.
That is right. As long as you clone the repository and run a simple cargo build or cargo check, the problem is immediately exposed.
error: could not compile ‘fastrender’ (lib) due to 34 previous errors; 94 warnings emitted
In other words, this so-called successful AI coding project cannot even compile. It is not even close to being a success.
The developer who exposed this provided detailed evidence.
First, all builds on the main branch of GitHub Actions have completely failed, and the workflow file itself contains errors.
Furthermore, if you run the tests automatically, you will find dozens of test cases failing. The PRs that tried to fix these issues were directly rejected by CI and could not be merged.
What is even more absurd is that if you look at the Git history, you cannot find a single commit that can be traced back to 100 commits. You cannot find a single clean commit.
In other words, this repository looks complete, but it has never been in a “usable” state.
Click to view the full analysis
https://gist.github.com/embedding-shapes/f5d096dd10be44ff82b6e5ccdaf00b29
What we cannot confirm is what exactly Cursor’s research team was doing during the week they claimed to have released the code. They seem to have never run cargo build, or even cargo check.
Because if they had, they would have seen dozens of errors and about 100 warnings. If they had tried to fix these errors, the project would have exploded with even more problems.
As of now, there is an unresolved GitHub issue in the repository.
Issue link: https://github.com/wilsonzlin/fastrender/issues/98
The conclusion is already very clear:
This is not a real engineering project, but typical “AI Slop” — AI garbage.
The so-called generated code, in terms of code structure and naming style, looks very professional at first glance, but behind it is a complete lack of real engineering logic. The core functions and basic tests all fail.
In Cursor’s presentation, they talked about a grand plan but used “small tricks” to cover up the real results. The so-called success was only partial.
More critically, for a repository of this size, Cursor did not provide any verifiable evidence. There was no usable version tag, no release, no commit that could prove these screenshots were real.
From a marketing perspective, Cursor’s blog post is a classic case of using beautiful demo videos to cover up the lack of real engineering. The prototype and files look complete, but the engineering details are completely missing. It is not verifiable.
They did not provide a clear roadmap, a real test suite, or any way to avoid risks. They only emphasized their own strength.
So far, the only proof we have is:
The AI agent wrote a large amount of code that consumed millions of tokens, and the result is completely unusable garbage.
A real browser needs to pass standard tests and be compared with Chrome. The corresponding application needs to pass a standard test.
The only supported function is to render a simple HTML file.
Obviously, the gap between what Cursor released and a real browser is huge.
GitHub and Hacker News Are Furious
Once the “fake product” label was attached, it was like a domino effect. The entire internet was furious.
On GitHub, angry comments flooded the issue page.
One developer wrote in blood-red text: “Shameful. Disgusting. Lies.”
Some investors said: “I invested money, and I did not know GitHub was like this.”
As long as it is AI-generated code, industry insiders cannot stand it. In one day, the project was completely exposed.
On Hacker News, there were nearly 200 comments, with one user thoroughly exposing the project from the bottom up.
User pavlov pointed out that the so-called “browser” is actually just a JS runtime, not a real browser.
Another user listed html5ever, cssparser, rquickjs, and other components, and found that these are actually the “shell” of Mozilla’s abandoned Servo project.
User brabel joked:
“So you are saying ‘modern browser’ is not a browser?”
Another developer’s first reaction was to look at the code, and at a glance, they could see the problem.
The only explanation is that no one actually ran it. In reality, everyone only looked at the demo video and screenshots.
Anthropic Is Too Aggressive, Cursor Got Carried Away
Although Cursor did not directly say which model they used, the industry generally believes that behind this “autonomous coding” breakthrough, it must be the latest version of Anthropic’s Claude.
The text that claims success is actually:
“Multiple workers collaborate on the same task, handle conflicts, and cancel the progress of other workers.”
But this exaggerated description has no evidence to support it.
There is no usable commit, no engineering explanation, no demo.
It also does not explain why a Chrome-level browser, which they claim to have built, can only display a simple HTML file. It is clearly a trick.
In fact, Cursor’s use of AI automatic coding for a week is a common problem in the entire industry.
Just last month, when the entire AI community was at its peak, a product called Claude Code came out.
After Claude Code, Boris Cherny posted on X that he was completely shocked. He said he had not written a single line of code in 30 days, and Claude Code did all the work. Everything was generated by Claude Code itself.
View the full post
Bridgewater’s chief engineer Jaana Dogan said that Claude Code built a complete internal tool for their team in one hour.
Former Tesla AI director Andrej Karpathy also admitted that he is currently in a state of “addiction” and has not studied anything else in depth.
View the full post
Under this wave, a blog post claimed that in one week, they wrote a complete browser from scratch. It sounded so natural and so powerful.
But Cursor’s engineers overestimated the capabilities of AI automatic coding. They used a flashy demo to cover up the real problem, and the result was a complete disaster.
AI Engineers Became AI Slop Makers
Although “AI slop” sounds harsh, it also seriously warns us about the real situation of AI coding.
Yuchen Jin, CTO of Hyperbolic, pointed out that the key problem with Cursor’s demo is that the model training data contains a large amount of non-communicative “AI slop.”
He gave an analogy: a team of people who only talk to themselves.
He said: “A complete engineering project requires three roles: Planner, Executor, and Reviewer.”
“Model selection is key. GPT-5.2 is suitable for engineering planning tasks, while Opus 4.5 is good at ‘coding’ but not at ‘thinking.'”
“If the organization gives too many ‘coding’ tasks to AI and does not supervise efficiency, the result is the company’s ‘AI slop production line.'”
HyperWriteAI CEO Matt Shumer also posted on X, saying that as long as there are clear goals and human support at key points, AI agent swarms can indeed produce real, usable code.
However, the reality is that on the AI coding track, AI still cannot replace real engineers.
In fact, a new term has emerged to describe AI-generated code that looks complete but is actually broken: “Cracked Engineer” — cracked engineer.
More precisely, it refers to those who can deceive an entire team with a single demo.
What Cursor did is exactly what Karpathy called “vibe coding” —
It looks cool and smooth on the surface, but it completely ignores engineering logic.
As a result, what we get is not a real Cursor engineering project, but a pile of completely unusable “AI slop.”
The so-called “cracked engineer” is actually a new type of fraud. They use AI tools to generate code that looks professional at first glance, but is full of logical holes. They can deceive investors, deceive the market, and produce “AI slop.”
The founder of AI research company Intology said:
“We are currently in a stage where human-AI collaboration is 15 times more productive than pure AI. In the future, when we have thousands of AI agents competing with each other, it will be like a car crash. One AI will write code, another AI will review it, and a third AI will test it. But if there is no real engineer in the middle to coordinate, the result is dozens of AI agents working on the same product with zero efficiency.”
These so-called “cracked engineers” are actually just people who know how to “package garbage.”
