Because it's an extremely large and complex project that is also very clearly specified, to the point that the three word prompt "build a browser" encapsulates a huge amount of detail.
Similar to "build space invaders", another useful test prompt for seeing how well an LLM can do at a medium complexity task without having to give it a great deal of instruction.
Can still be a good capability test. Building a car is a real world equivalent. It’s highly complex and has been done billions of times. Still hard to pull off if you ask me.
I think a lot of people here are kind of concerned that there's only three browsers now, and overwhelmingly really only two: Google Chrome and Safari for iOS. Internet Explorer is just Chrome now and while I use Firefox it is still pretty tiny in usage stats. I do not consider things like Epiphany/Gnome Web as serious contenders, and even if I did that would still be only one more niche browser. Things without JS support like NetSurf, in my opinion, don't count at all.
It's a little scary when a single megacorp has so much power over something ostensibly open like the internet, but it also has historically taken an incredible amount of resources in order to make a browser making it hard for new players to break in. A modern web browser is arguably more complicated than an operating system...hell, it arguably is an operating system. It touches a ton of aspects of computer science, and requires lots of dedicated workers to keep up with web standards.
Because it has been such an intractable problem for so long, it's an extremely tempting target when the circumstances have changed. Anyone here can basically have a metaphorical intern working as many hours as they want for ~$20-$100/month. A problem that would be impossible for a single person five years ago suddenly seems "almost possible" when you can work at a higher level and have the pesky "code" details taken care of for you.
How does your custom AI-built browser challenge the current browser triopoly?
The LLMs are trained on the code of existing browsers. This is essentially a massive process of turning code you don't understand into code you don't understand.
The code details are pretty much all of the details, other than the protocols and standards.
If you understood the codebase of existing browsers (or at least could be confident in making arbitrary changes to existing browsers, perhaps with AI assistance?) then the triopoly wouldn't be threatening because you could just patch out manifestv3 whenever you want.
There is also the problem of people not testing their websites to be compatible with your custom browser. But I would say this is a problem on the protocol level.
> A problem that would be impossible for a single person five years ago suddenly seems "almost possible" when you can work at a higher level and have the pesky "code" details taken care of for you.
It was already possible - The Ladybird browser started with one person and then a team of experts in maintaining browsers joined in. It makes a lot of sense for them to try to build one from scratch; with / without AI.
The problem which applies to all non mainstream browsers: Is anyone going to use your browser over the established ones because of some technical detail such as, who wrote it or what it is written in?
99.999% of people only care about whether if it just works and disappears away from the user whilst they're using it as a daily driver for browsing or doing work.
How is AI surely going to make this better the other than inexperienced folks throwing PRs they have not read? In fact, it helps the experts rather than the vibe coders.
“A common complaint today from AI coding skeptics is that LLMs are fine for toy projects but can’t be used for anything large and serious.
I think within 3 years that will be comprehensively proven incorrect, to the point that it won’t even be controversial anymore.
I picked a web browser here because so much of the work building a browser involves writing code that has to conform to an enormous and daunting selection of both formal tests and informal websites-in-the-wild.
Coding agents are really good at tasks where you can define a concrete goal and then set them to work iterating in that direction.
A web browser is the most ambitious project I can think of that leans into those capabilities.”
“The browser and this project were co-developed and very symbiotic, only because the browser was a very useful objective for us to measure and iterate the progress of the harness. The goal was to iterate on and research the multi-agent harness—the browser was just the research example or objective.”
IMO it’s a pretty verifiably correct and complex piece of software. There are all sorts of test suites but no greater one than comparing a page rendered to how it renders in chrome.
Also, and maybe more importantly, a browser is the kind of the perfect entry point for a lot of businesses.
If you’re for example hoping to build an agent that people rely on to manage their calendars, emails, book them vacations, etc.. a browser is probably the perfect form factor for that.
Or if you’re a AI lab and want to tap into training data behind logged in websites - a browser everyone uses is the ultimately what you need.
If you can build and maintain a browser from scratch with just agents then there’s no need to pay millions of dollars to hire away chromium engineers from Google.
Feedle looks useful, I've already found one blog I wasn't aware of. Thanks for recommending it!
I wonder if Feedle will provide an API (even a paid API)? Could be nice to plug this into a metasearch engine as an additional data source. And I guess that might be necessary to add Feedle search to an RSS reader or native app as well.
Because the traditional web is dying due to AI. Search is now turning into read the AI summary. So, what does the future look like? I mean, how do we bridge the gap between where we are now and that future. I don't like it...
It would be very useful if AI could do it (though I kind of doubt something high-performance can be created this way). It's economically unfeasible for humans to develop a browser engine from scratch anymore. Microsoft and Opera gave up. Apple probably will too someday.
It's the most complex piece of software on your computer (it's basically an entire sandboxed operating system, at this point), the standards it has to adhere to are expanding every day, its performance optimization is critical yet adversarial (i.e. website owners have no incentive to make their sites efficient - the browser will be blamed for slowness, not them), and it costs the user nothing. No company can afford to maintain that, unless it's serving some broader strategy that's earning billions.
I have just barely noticed this but I think this shows a really interesting trend that is long overdue. Recently many things were hard and now they are 'easy' (ish) which means we are likely to see an explosion of choices. This is good from an ecosystem point of view. Diversity will, initially, cause a lot of issues but in the long term it will lead to many new things.
this doesn't really matter, but, normal english would be either "Why are people all (of) the sudden writing browsers with AI?" or (slightly less normal) "Why all (of) the sudden are people writing browsers with AI?", with the (of) being more 'correct' but often omitted. Regardless, it should be "are people" instead of "people are" in a question.
It's a difficult, large problem - but one with very extensive tests available and that LLMs have a good understanding of. You can also look up the earlier automated ports of the JustHTML parser. A rendering engine is the next (admittedly large) step.
I say this as someone who also took on this task, about a week before I saw Cursor's attempt.
Why did Alex Honnold climb Taipei 101 when he could just take the elevator?
I know the analogy is not perfect but it's the kind of project that wasn't feasible for a single dev before LLMs so now it just seems like a fun thing to try for some people.
Who? Where? I’m blissfully unplugged from this bubble (despite lurking on HN). Show me a usable browser that was “written with AI”. Such claims are likely PR.
It remains an open challenge. I've seen three credible initial attempts so far, though none of them have got further than being able to render static HTML and CSS:
Because the most used, high value target for exploitation application, is a perfect candidate for autocoded slop wrangled by incompetent developers who don't understand the code enough to write it from scratch.
And browsers are one of those products where if it was simple to alter everyone would like different setups and features. Like the Notes app of before.
Naturally they are all incomplete implementations because the AI agent is mostly reusing open source components and for the stuff it decides to write itself, doesn't have the training data to implement a full rendering engine in a one-shot.
I've noticed a lot of desperation to validate the usefulness of AI. I guess partly from those who are trying to make money from it and partly from others who need the cognitive shortcut of offloading work and creativity to it.
> I've noticed a lot of desperation to validate the usefulness of AI. I guess partly from those who are trying to make money from it and partly from others who need the cognitive shortcut of offloading work and creativity to it.
I've noticed a lot of desperation to deny the usefulness of AI. I guess partly from those who feel threatened by it and partly from others who aren't curious enough to figure out how to use it effectively.
Because it's an extremely large and complex project that is also very clearly specified, to the point that the three word prompt "build a browser" encapsulates a huge amount of detail.
Similar to "build space invaders", another useful test prompt for seeing how well an LLM can do at a medium complexity task without having to give it a great deal of instruction.
I called building a browser the "hello world" of complex parallel agent coding harnesses the other day: https://simonwillison.net/2026/Jan/23/fastrender/#a-single-e...
I am not convinced either of these are good test prompts for generic complexity tasks. Many solutions have already been included in the training data!
You can trivially produce a web browser by copying and compiling the code for firefox, no transformer needed.
Can still be a good capability test. Building a car is a real world equivalent. It’s highly complex and has been done billions of times. Still hard to pull off if you ask me.
But that would produce Firefox.
The goal with these tests is to see if the models can make something new, not just copy an existing solution.
That is the goal, at least.
But how do you define, or indeed assess, novelty?
It's not that difficult to take an existing mature codebase and morph it such that it looks quite different but is functionally unchanged.
This is a very different task than building something that's not been built before.
I think a lot of people here are kind of concerned that there's only three browsers now, and overwhelmingly really only two: Google Chrome and Safari for iOS. Internet Explorer is just Chrome now and while I use Firefox it is still pretty tiny in usage stats. I do not consider things like Epiphany/Gnome Web as serious contenders, and even if I did that would still be only one more niche browser. Things without JS support like NetSurf, in my opinion, don't count at all.
It's a little scary when a single megacorp has so much power over something ostensibly open like the internet, but it also has historically taken an incredible amount of resources in order to make a browser making it hard for new players to break in. A modern web browser is arguably more complicated than an operating system...hell, it arguably is an operating system. It touches a ton of aspects of computer science, and requires lots of dedicated workers to keep up with web standards.
Because it has been such an intractable problem for so long, it's an extremely tempting target when the circumstances have changed. Anyone here can basically have a metaphorical intern working as many hours as they want for ~$20-$100/month. A problem that would be impossible for a single person five years ago suddenly seems "almost possible" when you can work at a higher level and have the pesky "code" details taken care of for you.
How does your custom AI-built browser challenge the current browser triopoly?
The LLMs are trained on the code of existing browsers. This is essentially a massive process of turning code you don't understand into code you don't understand.
The code details are pretty much all of the details, other than the protocols and standards.
If you understood the codebase of existing browsers (or at least could be confident in making arbitrary changes to existing browsers, perhaps with AI assistance?) then the triopoly wouldn't be threatening because you could just patch out manifestv3 whenever you want.
There is also the problem of people not testing their websites to be compatible with your custom browser. But I would say this is a problem on the protocol level.
> A problem that would be impossible for a single person five years ago suddenly seems "almost possible" when you can work at a higher level and have the pesky "code" details taken care of for you.
It was already possible - The Ladybird browser started with one person and then a team of experts in maintaining browsers joined in. It makes a lot of sense for them to try to build one from scratch; with / without AI.
The problem which applies to all non mainstream browsers: Is anyone going to use your browser over the established ones because of some technical detail such as, who wrote it or what it is written in?
99.999% of people only care about whether if it just works and disappears away from the user whilst they're using it as a daily driver for browsing or doing work.
How is AI surely going to make this better the other than inexperienced folks throwing PRs they have not read? In fact, it helps the experts rather than the vibe coders.
Blame Simon Willison ;)
“A common complaint today from AI coding skeptics is that LLMs are fine for toy projects but can’t be used for anything large and serious.
I think within 3 years that will be comprehensively proven incorrect, to the point that it won’t even be controversial anymore.
I picked a web browser here because so much of the work building a browser involves writing code that has to conform to an enormous and daunting selection of both formal tests and informal websites-in-the-wild.
Coding agents are really good at tasks where you can define a concrete goal and then set them to work iterating in that direction.
A web browser is the most ambitious project I can think of that leans into those capabilities.”
https://simonwillison.net/2026/Jan/8/llm-predictions-for-202...
“The browser and this project were co-developed and very symbiotic, only because the browser was a very useful objective for us to measure and iterate the progress of the harness. The goal was to iterate on and research the multi-agent harness—the browser was just the research example or objective.”
https://simonwillison.net/2026/Jan/23/fastrender/
The Cursor FastRender project started in December, so it wasn't influenced by my prediction in January.
> Coding agents are really good at tasks where you can define a concrete goal and then set them to work iterating in that direction.
Specifically ones that are in the training data.
> A web browser is the most ambitious project I can think of that leans into those capabilities.”
I assume Linux and gcc are in the training data, so additional options may be OSes and compilers..
> Coding agents are really good at tasks where you can define a concrete goal and then set them to work iterating in that direction.
Wholly based on other people's work. Which is OK.
IMO it’s a pretty verifiably correct and complex piece of software. There are all sorts of test suites but no greater one than comparing a page rendered to how it renders in chrome.
Also, and maybe more importantly, a browser is the kind of the perfect entry point for a lot of businesses.
If you’re for example hoping to build an agent that people rely on to manage their calendars, emails, book them vacations, etc.. a browser is probably the perfect form factor for that.
Or if you’re a AI lab and want to tap into training data behind logged in websites - a browser everyone uses is the ultimately what you need.
If you can build and maintain a browser from scratch with just agents then there’s no need to pay millions of dollars to hire away chromium engineers from Google.
The list goes on!
I'd really like a pure RSS browser. Just nice articles in a standard format, no ads, no clickbait.
RSS readers are cool but the discovery is still tedious - looking for feeds isn't much fun.
I did just discover feedle though, which looks like halfway toward what I want! https://feedle.world/
I want feedle in a nice native app.
Feedle looks useful, I've already found one blog I wasn't aware of. Thanks for recommending it!
I wonder if Feedle will provide an API (even a paid API)? Could be nice to plug this into a metasearch engine as an additional data source. And I guess that might be necessary to add Feedle search to an RSS reader or native app as well.
Because the traditional web is dying due to AI. Search is now turning into read the AI summary. So, what does the future look like? I mean, how do we bridge the gap between where we are now and that future. I don't like it...
It would be very useful if AI could do it (though I kind of doubt something high-performance can be created this way). It's economically unfeasible for humans to develop a browser engine from scratch anymore. Microsoft and Opera gave up. Apple probably will too someday.
It's the most complex piece of software on your computer (it's basically an entire sandboxed operating system, at this point), the standards it has to adhere to are expanding every day, its performance optimization is critical yet adversarial (i.e. website owners have no incentive to make their sites efficient - the browser will be blamed for slowness, not them), and it costs the user nothing. No company can afford to maintain that, unless it's serving some broader strategy that's earning billions.
Several years ago, I did write that every programmer should attempt to write a browser: https://austinhenley.com/blog/morechallengingprojects.html
:)
It’s a demonstrably difficult problem that, aside from lady bird, has not been easy for independent devs to accomplish.
I have just barely noticed this but I think this shows a really interesting trend that is long overdue. Recently many things were hard and now they are 'easy' (ish) which means we are likely to see an explosion of choices. This is good from an ecosystem point of view. Diversity will, initially, cause a lot of issues but in the long term it will lead to many new things.
Because they want to sell AI and browsers are stuff users spend a lot of time in.
I don't really see the added value there though. I do see it for search and deep research but much less for regular browsing.
this doesn't really matter, but, normal english would be either "Why are people all (of) the sudden writing browsers with AI?" or (slightly less normal) "Why all (of) the sudden are people writing browsers with AI?", with the (of) being more 'correct' but often omitted. Regardless, it should be "are people" instead of "people are" in a question.
(In edited writing, “all of a sudden” is far more common than “all of the sudden”. https://www.merriam-webster.com/grammar/usage-of-all-of-a-su...)
oops you're right. some grammar policer i am
(I didn't know that, actually)
It's a difficult, large problem - but one with very extensive tests available and that LLMs have a good understanding of. You can also look up the earlier automated ports of the JustHTML parser. A rendering engine is the next (admittedly large) step.
I say this as someone who also took on this task, about a week before I saw Cursor's attempt.
Because they can. It's fun.
AI makes it possible to do things you'd never have been able to do before, perhaps due to skill level or perhaps due to the time investment required.
It is a lot of fun to make software that you'd only been able to dream about making before AI.
There's a lot of "if AI is so good then why can't it do X", and depending on who you ask "writing a browser" is somewhere on the frontier of X values.
Why did Alex Honnold climb Taipei 101 when he could just take the elevator?
I know the analogy is not perfect but it's the kind of project that wasn't feasible for a single dev before LLMs so now it just seems like a fun thing to try for some people.
Who? Where? I’m blissfully unplugged from this bubble (despite lurking on HN). Show me a usable browser that was “written with AI”. Such claims are likely PR.
It remains an open challenge. I've seen three credible initial attempts so far, though none of them have got further than being able to render static HTML and CSS:
- https://github.com/hiwavebrowser/hiwave
- https://github.com/wilsonzlin/fastrender
- https://github.com/embedding-shapes/one-agent-one-browser
I've started tracking them on this tag on my blog: https://simonwillison.net/tags/browser-challenge/
Something about a solution in search of a problem methinks.
Because the most used, high value target for exploitation application, is a perfect candidate for autocoded slop wrangled by incompetent developers who don't understand the code enough to write it from scratch.
Yup. I think LLMs and AI in general can be useful, but I don't know what to call this wielding of it like a hammer in search of nails. A fad?
Because someone tried and it vaguely worked.
And browsers are one of those products where if it was simple to alter everyone would like different setups and features. Like the Notes app of before.
Naturally they are all incomplete implementations because the AI agent is mostly reusing open source components and for the stuff it decides to write itself, doesn't have the training data to implement a full rendering engine in a one-shot.
Can someone please build a great cross-platform email client. Last I checked there weren't a lot of great options outside of Outlook and Thunderbird.
Since most email is HTML these days the first step would be to build a browser!
It is a stepping stone to writing an entire feature rich operating system thereafter, if only a virtual one.
The same reason why all of a sudden people are rolling their own sandboxes using AI (without even looking at the code!):
Getting people to use it on the other hand...its just a meme
I've noticed a lot of desperation to validate the usefulness of AI. I guess partly from those who are trying to make money from it and partly from others who need the cognitive shortcut of offloading work and creativity to it.
> I've noticed a lot of desperation to validate the usefulness of AI. I guess partly from those who are trying to make money from it and partly from others who need the cognitive shortcut of offloading work and creativity to it.
I've noticed a lot of desperation to deny the usefulness of AI. I guess partly from those who feel threatened by it and partly from others who aren't curious enough to figure out how to use it effectively.
(Sorry, couldn't resist.)
Neodymium Magnet Dealers in India