New Intelligence Report
Editor: Aeneas, Very Sleepy
【New Intelligence Summary】 The claim that GPT-5.2 worked non-stop for 7 days to create a browser has just been debunked! A developer confirmed in a post that the Cursor project is just "AI Slop," and the code cannot even be compiled. Cursor was too impatient this time.
A few days ago, the entire AI community was stunned by a major announcement from Cursor.
Here's what happened:
This sounds incredibly enticing—
As tokens become as cheap as electricity and water, AI can self-iterate indefinitely until it achieves its goal.
Whether it's an operating system, office software, or a game engine, as long as there's enough computing power, it seems AI can "grind out" anything for you.
However, before everyone could even recover from the shock, the "Leeuwenhoeks" of the technical community stepped in.
They carefully examined the open-source project code released by Cursor and discovered a shocking truth—
This so-called "AI browser" cannot even pass the most basic compilation!
In a technical blog post, the author pointed out sharply:
The so-called "breakthrough progress" touted by Cursor is essentially a pile of "AI Slop" lacking engineering logic.
They essentially played a clever "smokescreen" in their promotion, making everyone believe the project was actually running.
But in reality, this is just a pile of non-executable garbage code.
Blog Address: https://embedding-shapes.github.io/cursor-implied-success-without-evidence/
Is it fake that GPT-5.2 built a browser in 7 days?
Next, let's take a close look at the "debunking" article from the developer community, which meticulously exposed the false claims in Cursor's promotion.
First, the author analyzed what Cursor actually did.
On January 14th, they published a blog post titled "Scaling long-running autonomous coding."
Official Blog: https://cursor.com/blog/scaling-agents
In this article, they discussed the experiment of letting "coding agents run autonomously for weeks." Their explicit goal was:
To understand how far we can push the boundaries of agent coding, thereby completing projects that would typically require human teams months to finish.
Then, Cursor's researchers discussed some methods they tried, analyzed the reasons for failure, and how to solve the problems.
Finally, they found a solution that "resolved most of our coordination problems and allowed us to scale to very large projects without relying on a single agent."
Ultimately, this solution achieved a stunning result—
To test this system, we set an ambitious goal for it: to build a web browser from scratch. These agents ran for nearly a week, writing over 1 million lines of code across 1,000 files.
At the same time, they released the source code on GitHub.
GitHub Project: https://github.com/wilsonzlin/fastrender
This is strange, so did the agents successfully complete this task?
If you weren't swayed by that statement, you'd find this confusing point—
They claimed that "despite the large codebase, new agents can still understand it and make meaningful progress," and that "hundreds of workers run concurrently, pushing to the same branch with minimal conflicts," but they never explicitly stated whether this attempt was successful.
Can it actually run? Can you run this browser yourself? We don't know, and they never made it clear.
The so-called demonstration is just a short 8-second "video":
Below, they wrote:
Although this looks like a simple screenshot, building a browser from scratch is very difficult.
In short, from start to finish, they never definitively admitted that this browser is runnable and functional!
Open it up: Full of errors, cannot even run
In short, if you only look at the README, demo screenshots, or even a few promotional descriptions, this project seems really impressive.
However, as soon as you actually clone the repository and run a cargo build or cargo check, the problems will immediately be exposed.
error: could not compile 'fastrender' (lib) due to 34 previous errors; 94 warnings emitted
It can be said that this codebase is far from being a "workable browser," and it can even be said that it has never been successfully built!
The article author discovered the following multiple pieces of evidence.
First, multiple recent runs of GitHub Actions on the main branch have all failed, including errors in the workflow files themselves.
Additionally, if you try to build independently, you'll find dozens of compiler errors, and the most recent PRs were merged even with the CI failing.
More exaggeratedly, if you look through the Git history, tracing back 100 commits from the most recent one, you can hardly find a single commit that compiles cleanly.
In other words, this repository has never been in a "runnable" state since its inception.
Swipe up and down to view
https://gist.github.com/embedding-shapes/f5d096dd10be44ff82b6e5ccdaf00b29
Now we have no way of knowing what the "agents" released by Cursor's researchers on this codebase actually did, but they apparently never ran cargo build, let alone cargo check.
Because both commands would produce dozens of errors and about 100 warnings. If you actually tried to fix these errors, the number of errors would definitely explode.
Currently, in their repository, there is an unresolved GitHub issue about this.
Issue Address: https://github.com/wilsonzlin/fastrender/issues/98
The conclusion is already very clear:
This is not real engineering code at all, but typical "AI Slop."
This low-quality code pile-up may mimic a certain function in form, but behind it lacks coherent engineering intent, and it cannot even pass the most basic compilation.
In Cursor's demonstration, they talked about grand plans for the next step, but said nothing about "how to run," "expected effects," or "how it works."
Moreover, besides dropping a code repository link, Cursor did not provide any reproducible demonstration, nor did it give any known usable version tags (tag/release/commit) to verify those glamorous screenshots.
Regardless of the original intention, Cursor's blog post attempted to create the illusion of a "functionally normal prototype," but omitted the basic honesty most valued in the engineering world—reproducibility.
They indeed did not explicitly claim "it can run normally," which allowed them to technically avoid accusations of "lying," but this is highly misleading.
So far, the only thing they have proven is:
AI agents can output millions of tokens wildly, but the final generated code is still a pile of non-executable waste.
A "browser experiment" doesn't need to match Chrome, but it should at least have a reasonable minimum standard:
Compile successfully on a supported toolchain and render a simple HTML file.
Unfortunately, neither Cursor's article nor the public build reached this passing line.
GitHub flooded, developers angry
This behavior of packaging a "semi-finished product" as a "milestone" completely angered the developer community.
In the GitHub Issue section, angry comments flooded the screen:
I tried it too, and it simply doesn't run.
The code logic is completely unrelated, and they merged with all CI red? Are we supposed to worship screenshots?
Since the functionality is fake, what's the point of open-sourcing this repository? To prove AI can make electronic waste?
Some people pointed out the essence of this "bubble engineering" with a single sentence:
After all, investors don't understand code, and they don't even know what GitHub is.
As long as it's code written automatically by a computer, the performance curve can skyrocket. When the machine runs, gold flows...
Moreover, on Hacker News, there are nearly 200 discussions that have completely exposed the bottom of this project.
Netizen pavlov pointed out that the so-called "from scratch" and "custom JS virtual machine" are pure nonsense.
Just take a look at the dependency list (html5ever, cssparser, rquickjs) and you can see that this thing is essentially a "shell" version of Mozilla's Servo engine.
Netizen brabel was even more amused:
These people actually think claiming "hand-crafted from scratch" is a good move?
Programmers' first step is to check dependencies, and they can tell at a glance that it's just calling packages.
The only explanation is that they bet no one would seriously verify it, after all, most people would just cheer after seeing the title.
Is Anthropic too dominant, forcing Cursor into a corner?
Although Cursor never explicitly said "this is ready for production," they used grand narratives like "built from scratch" and "meaningful progress," along with carefully selected screenshots, to successfully create the illusion of a "successful experiment."
Their closest statement to success was:
Hundreds of agents can collaborate on the same codebase for weeks and make real progress.
However, this astonishing claim has no evidence to support it.
No working commit, no build instructions, no demonstration.
People don't expect it to become the next Chrome, but if you claim you've built a browser, it should at least be able to demonstrate compiling + loading a basic HTML file; otherwise, it's just pure deception of the public.
Actually, when the news of AI programming automatically for a week came out, it felt a bit strange.
In the past month, the highlight moments of the global AI community have basically been on Claude Code.
The X posts by Claude Code's father, Boris Cherny, basically cause a stir in the community. For example, he said that in the past 30 days, he didn't write a single line of code, and all contributions to the Claude Code repository were completed by Claude Code itself.
Further Reading: He didn't write a line of code in 30 days, but he earned 1 billion US dollars!
Google's chief engineer Jaana Dogan said that Claude Code completed the entire team's one-year task in just one hour.
Former Tesla AI director Andrej Karpathy bluntly stated: Silicon Valley is experiencing a magnitude 9 earthquake, and he has never felt so behind...
Further Reading: Goodbye, programmers! Silicon Valley is all AI Coding, Karpathy declares a magnitude 9 earthquake is coming
In this context, a narrative of "coding agents running for a week and writing a browser themselves" is so in line with the trend and so eye-catching.
No wonder after trying this idea, the Cursor engineers didn't think much and rushed to post it, only to be attacked by the sharp-eyed developers.
AI Programmer Super Evolution: "Cracked Engineer"
Although this "browser farce" was tragic, it unexpectedly revealed the true evolutionary path of AI programming.
Yuchen Jin, co-founder and CTO of Hyperbolic, pointed out the key lesson implied in Cursor's demonstration: simply piling up numbers doesn't work.
Letting a bunch of agents coexist at the same level and self-coordinate will only bring chaos.
· Role division must be clear: need a Planner, Executor, and Reviewer.
· Model differentiation: GPT-5.2 is more suitable for long-term planning tasks; while Opus 4.5 is prone to "early leave" and laziness.
· Organizational structure: Adding too many "management" agents will instead drag down efficiency, which is exactly the same as the "big enterprise disease" in human companies.
Matt Shumer, CEO of HyperWriteAI, also found during replication that as long as the runtime framework and capabilities are supported, AI agent clusters with clear division of labor can indeed produce substantial progress.
However, deeper evolution is not just happening in AI, but also in human engineers.
In Silicon Valley, a new buzzword is replacing the old "10x engineer"—"Cracked Engineer".
This term refers to top developers who can do the work of an entire team alone.
Cursor's failure this time is precisely because of the "vibe programming" trap mentioned by Karpathy—
Only enjoying the thrill of AI generating code wildly, completely abandoning engineering rigor.
What is born from this can only be the kind of "Cursor mover" waste that cannot run without patching.
Real "cracked engineers" are the opposite of this phenomenon. They use AI crazily but never blindly trust AI.
They have deep technical foundations, can spot logical flaws in AI-generated code at a glance, and can clean up the "electronic slop" that appeared in this browser project.
As the founder of the startup Intology said:
A few focused and knowledgeable people plus AI can do more than 15 people without AI in the past.
Future software development will not be thousands of unsupervised AI agents bumping around like headless flies (and then producing a browser that cannot compile); instead, it will be led by a "cracked engineer," with dozens of AI Agents, precisely and efficiently building a real product.
And these "cracked" programmers will also eliminate those who only "pretend to program."
References:
https://embedding-shapes.github.io/cursor-implied-success-without-evidence/
https://www.theinformation.com/articles/forget-vibe-coders-cracked-engineers-rage-tech?rc=epv9gi
Follow ASI Instantly
⭐Like, Share, and View - One-Click Triple Connection⭐
Light up the star mark, lock in New Intelligence's lightning-fast push!