> "So I agree this isn't just wiring up of dependencies, and neither is it copied from existing implementations: it's a uniquely bad design that could never support anything resembling a real-world web engine."
It hurts, that it wasn't framed as an "Experiment" or "Look, we wanted to see how far AI can go - kinda failed the bar." Like it is, it pours water on the mills of all CEOs out there, that have no clue about coding, but wonder why their people are so expensive when: "AI can do it! D'oh!"
They were making claims without the level of rigor to back them up. There was an opportunity to learn some difficult lessons, but—and I don’t think this was your intention—it came across to me as kind of access journalism; not wanting to step on toes while they get their marketing in.
The claims they made really weren't that extreme. In the blog post they said:
> To test this system, we pointed it at an ambitious goal: building a web browser from scratch. The agents ran for close to a week, writing over 1 million lines of code across 1,000 files. You can explore the source code on GitHub.
> Despite the codebase size, new agents can still understand it and make meaningful progress. Hundreds of workers run concurrently, pushing to the same branch with minimal conflicts.
That's all true.
On Twitter their CEO said:
> We built a browser with GPT-5.2 in Cursor. It ran uninterrupted for one week.
> It's 3M+ lines of code across thousands of files. The rendering engine is from-scratch in Rust with HTML parsing, CSS cascade, layout, text shaping, paint, and a custom JS VM.
> It kind of works! It still has issues and is of course very far from Webkit/Chromium parity, but we were astonished that simple websites render quickly and largely correctly.
That's mostly accurate too, especially the "it kind of works" bit. You can take exception to "from-scratch" claim if you like. It's a tweet, the lack of nuance isn't particularly surprising.
In the overall genre of CEO's over-hyping their company's achievements this is a pretty weak example.
I think the people making out that Cursor massively and dishonestly over-hyped this are arguing with a straw man version of what the company representatives actually said.
> In the overall genre of CEO's over-hyping their company's achievements this is a pretty weak example
I kind of agree, but kind of not. The tweet isn't too bad when read from an experienced engineer perspective, but if we're being real then the target audience was probably meant to be technically clueless investors who don't and can't understand the nuance.
It's like claiming "my dog filed my taxes for me!" when in reality everything was filled out in TurboTax and your dog clicked the final submit button. Technically true, but clearly disingenuous.
I'm not saying an LLM using existing libraries is a bad thing--in fact I'd consider an LLM which didn't pull in a bunch of existing libraries for the prompt "build a web browser" to be behaving incorrectly--but the CEO is misrepresenting what happened here.
> "So I agree this isn't just wiring up of dependencies, and neither is it copied from existing implementations: it's a uniquely bad design that could never support anything resembling a real-world web engine."
It didn't use Servo, and it wasn't just calling dependencies. It was terribly slow and stupid, but your comment is more of a mischaracterization than anything the Cursor people have said.
[0] https://github.com/search?q=repo%3Awilsonzlin%2Ffastrender%2...
[1] https://github.com/search?q=repo%3Awilsonzlin%2Ffastrender+h...
But it was accompanied by a link to the GitHub repo, so you can hardly claim that they were deliberately hiding the truth.
Well, yes and no; we live in an era where people consume headlines, not articles, and certainly not links to Github repositories in articles. If VCs and other CEOs read the headline "Cursor Agents Autonomously Create Web Browser From Scratch" on LinkedIn, the project has served its purpose and it really doesn't matter if the code compiles or not.
You have a reputation. You don’t need to carry water for people who are misleading people to raise VC money. What’s the point of you language lawyering about the precise meaning of what he said?
“No no, you don’t get it guys. I’m technically right if you look at the precise wording” is the kind of silly thing I do all the time. It’s not that important to be technically right. Let this one go.
The reason I won't let this one go is that I genuinely believe people are being unfair to the engineer who built this, because some people will jump on ANY opportunity to "debunk" stories about AI.
I won't stand for misleading rhetoric like "it's just a Servo wrapper" when that isn't true.
I was in a meeting recently where a director lauded Claude for writing "tens of thousands of lines of code in a day", as if that metric in and of itself was worth something. And don't even get me started on "What percentage of your code is written by AI?"
The rest is stuff like HarfBuzz for font rendering which is an entirely cromulent dependency for a project like this.
And the latter is what's driving the push for KPIs the most - "active" ETFs already were bad enough because their managers would ask the companies they invested in to provide easily-to-grok KPIs (so that they could keep more of the yearly fee instead of having to pay analysts to dig down into a company's finances), and passive ETFs make that even worse because there is now barely any margin left to pay for more than a cursory review.
America's desire for stock-based pensions is frying the world's economy with its second and third order effects. Unfortunately, that rotten system will most probably only collapse when I'm already dead, so there is zero chance for most people alive today to ever see a world free of this BS.
The reality was the AI made an uncompilable mess, adding 100+ dependencies including importing an entire renderer from another browser (servo) and it took a human software engineer to clean it all up.
Don't publish things like that. At the very least link to a transcript, but this is a very non-credible way of reporting those numbers.
I'd still be surprised if that added up to "trillions" of tokens. A trillion is a very big number.
Fully agree that the original authors made some unsubstantiated and unqualified claims about what was done - which is sad, because it was still a huge accomplishment as i see it.
Just like your manager.
> ...while far off from feature parity with the most popular production browsers today...
What a way to phrase it!
You know, I found a bicycle in the trash. It doesn't work great yet, but I can walk it down a hill. While far off from the level of the most popular supercars today, I think we have made impressive progress going down the hill.
I'm also getting really tired of claims like "we are X% more productive with AI now!" (that I'm hearing day in and out at work and LinkedIn of course). Didn't we, as an industry, agree that we _didn't_ know how to measure productivity? Why is everyone believing all of these sudden metrics that try to claim otherwise?
Look, I'm not against AI. I'm finding it quite valuable for certain scenarios -- but in a constrained environment and with very clear guidance. Letting it loose with coding is not one of them, and the hype is dangerous by how much it's being believed.
EDIT: I retract my claim. I didn't realize this had servo as a dependency.
They marketed as if we were really close to having agents that could build a browser on their own. They rightly deserve the blowback.
This is an issue that is very important because of how much money is being thrown at it, and that effects everyone, not just the "stakeholders". At some point if it does become true that you can ask an agent to build a browser and it actually does, that is very significant.
At this point in time I personally can't predict whether that will happen or not, but the consequences of it happening seem pretty drastic.
And I'm an optimist, not one of the AI skeptics heavily present on HN.
From the post it sounds like the author would also doubt this when he talks about "glorified autocomplete and refactoring assistants".
Still, getting "something" to compile after a week of work is very different from getting the thing you wanted.
What is being sold, and invested in, is the promise that LLMs can accomplish "large things" unaided.
But they can't, as of yet, they cannot, unless something is happening in one of the SOTA labs that we don't know about.
They can however accomplish small things unaided. However there is an upper bound, at least functionally.
I just wish everyone was on the same page about their abilities and their limitations.
To me they understand conext well (e.g. the task, build a browser doesn't need some huge specification because specifications already exist).
They can write code competently (this is my experience anyway)
They can accomplish small tasks (my experience again, "small" is a really loose definition I know)
They cannot understand context that doesn't exist (they can't magically know what you mean, but they can bring to bear considerable knowledge of pre-existing work and conventions that helps them make good assumptions and the agentic loop prompts them to ask for clarification when needed)
They cannot accomplish large tasks (again my experience)
It seems to me there is something akin to the context window into which a task can fit. They have this compact feature which I suspect is where this limitation lies. Ie a person can't hold an entire browser codebase in their head, but they can create a general top level mapping of the whole thing so they can know where to reach, where areas of improvement are necessary, how things fit together and what has been and what hasn't been implemented. I suspect this compaction doesn't work super well for agents because it is a best effort tacked on feature.
I say all this speculatively, and I am genuinely interested in whether this next level of capability is possible. To me it could go either way.
Take a look in the Cargo.toml: https://github.com/wilsonzlin/fastrender/blob/19bf1036105d4e...
Maybe there is a main servo crate as well out there, and fastrender doesn't depend on that crate, but at least in my mind fastrender depends on some servo browser functionality.
EDIT: fastrender also includes the servo HTML parser: html5ever (https://github.com/servo/html5ever).
Is entropy increasing or decreasing the longer agents work on a code base? If it's decreasing, no matter how slowly, theoretically you could just say "ok, start over and write version 2 using what you've learned on version 1." And eventually, $XX million dollars and YY months of churning later, you'd get something pretty slick. And then future models would just further reduce X and Y. Right?
Maybe they just need to keep iterating.
I am an avid user of LLMs but I have not seen them remove entropy, not even once. They only add. It’s all on the verge of tech debt and it takes substantial human effort to keep entropy increases in check. Anyone can add 100 lines, but it takes genuine skill to do it 10 (and I don’t mean code golf).
And to truly remove entropy (cut useless tests, cut useless features, DRY up, find genuine abstractions, talk to PM to avoid building more crap, …) you still need humans. LLM built systems eventually collapse under their own chaos.
I think your analogy is quite fitting!
I've yet to see anyone in this space be negatively impacted by their outlandish claims.
They release a new model or add extra sub agents and the slate is wiped clean.
Management already doesn't trust developers in any way. Why would they believe you, who are clearly just trying to save your job, over a big company who clearly is the future!
Or do you trust your management to make the right decision?
We talked about dependencies, among a whole bunch of other things.
You can watch the full video on YouTube or read my extracted highlights here: https://simonwillison.net/2026/Jan/23/fastrender/
6 months ago with previous models this was absolutely impossible. One of the biggest limitations of LLMs is their difficulty with long tasks. This has been steadily improving and this experiment was just another milestone. It will be interesting a year from now to test how much better new models fare at this task.
project 1: build a text based browser using ratatui and quickjs.
project 2: base it on project 1. convert to gui, pages should render pure html.
project 3: acid1 compliance. Use constraint based programming to output final render, no animation support.
There was a story going around about LLMs making minesweeper clones, and they were all terrible in extremely dumb ways. The headline wasn't obvious, so I thought the take that people were getting from it is that AI is making the same dumb mistakes that it was making a year ago. Nope. It was people ranting about how coders are going to be out of a job next week. Meanwhile, none of them can do a minesweeper clone with like 50 working examples online, maybe 8 things you have to do right to be perfect, and 9000 articles about minesweeper and even mathematical papers about minesweeper to make everything about the game and its purpose perfectly clear. And then AI generates buttons that don't do anything and timers that don't stop.
Claude Opus 4.5: "Build minesweeper as an artifact, don't use react"
(Then "Fix it to work on mobile where right click isn’t a thing")
Play it here: https://tools.simonwillison.net/minesweeper
Transcript here: https://claude.ai/share/2d351b62-a829-4d81-b65d-8f3b987fba23
> tools like Cursor can be genuinely helpful as glorified autocomplete and refactoring assistants
That suggests a fairly strong anti-AI bias by the author. Anyone who thinks that this is all AI coding tools are today is not actually using them seriously.
That's not to say that this exercise wasn't overhyped, but a more useful, less biased article that's not trying to push an agenda would look at what went right, as well as what went wrong.