This makes a ton of sense to me with the huge corpus of web standards that exist. In theory we should be able to generate a browser from those specs, it just took a massive amount of effort before. Maybe if things like this get some traction, some of the human time spent implementing the spec could be spent on creating more/better specs, allowing for even better generation too.
Edit: replies are making good points about AI capabilities and where the effort really goes. Let’s just say that I meant this in an aspirational sense, rather than where the rubber actually meets the road today.
As someone who has spent the last 3 years implementing a browser engine from scratch full time (that currently passes ~half the 200k "css" tests that test style/layout/rendering), I find that (without a lot of close hand holding) the AIs are very far from being able to do this. They'll give you something that passes the tests, but it will do it a ridiculous way be far too slow to be useful (and is wrong architecturally such that it's not going to converge on a better solution).
I guess my hope is that while this may be the case now, perhaps it won’t be eventually. If you can generate from a spec, different implementations may be slower or worse in a variety of ways but still meet the spec.
At that point it’s about optimizing and deciding trade offs, both of which might mean heavily directing the generation in some way. Over time, maybe LLMs or whatever succeeds them won’t go down so many bad architectural paths. We’ve certainly seen that already in the last year or two.
I also don’t think folks like you ever lose your value in these efforts, even if all of it came true tomorrow. At worst it would be you plus the AI, which would always be a more potent combination than say, me with an AI trying to make a browser engine.
Not sure why you are being downvoted, but this is definitely a limitation of current models and difficult to optimize for. Anything you are not specifically optimizing for is essentially unconstrained: the model could learn it by pure chance, but the likelihood is exceedingly low.
Techniques, like the newly announced RL-XAR from Meta [1] are being developed that will likely improve reward models and guide RL training to optimize for metrics like software architecture that are hard to verify otherwise.
There’s quite a few edge cases where browsers allow non-compliant code to render as if it were to specs. Because if a site doesn’t render “correctly” from a user perspective, they blame the browser rather than blaming the web developer.
I think browser performance (and security) improvements are really difficult to tackle at the scale of browser complexity, and are both art and engineering. In other words, if this were possible, we would see the results first in existing browsers. This project is very interesting though.
Maybe some day this will be the basis but not the complete solution. Agreed that there’s a lot of semi intangible art that goes into browser engine decisions. I’m guessing a lot of which have to do as much with people and current landscape dynamics as they do with the tech.
> In theory we should be able to generate a browser from those specs, it just took a massive amount of effort before.
No, not really. The absolute vast majority of those specs are human technical and technical-adjacent language, not machine-readable specs.
On top of that many web specs often invent new terminology because a lot of things are specced years or decades after something popular has taken over the term in userland.
Maybe over time the specs will become more machine readable (bounded and concrete) if we take these paths. Although in some sense everything is machine readable these days, just not deterministically so (less helpful for specs, but not nothing either).
Your second point makes a lot of sense to me too. I’m not sure how this could handle that facet of human nature, except possibly to indirectly contribute to speeding up the cycle of spec creation.
Very cool idea, though I think we're a long way away from this. Hopefully they file bugs on the specs when they find ambiguities.
Also the specs are written to define observable behavior, and there's a fair bit of ambiguity that's UA defined, but to be actually compatible with the Web requires doing what Chrome does.
The good news is that you can look at the source of the 3 major engines (and Ladybird!) and your AI agent can do comparisons to figure out optimizations.
I'm sure this is not the origin, but Bez is a celebrity in England, a member of a 90s band called happy Mondays who didn't sing or play anything but just danced. Famous on the panel show circuit
So far it's only HTML/CSS (and we have our Rust-based framework to write the apps in). But it has been designed for fast incremental rendering, so it could be extended with JavaScript support quite easily.
This makes a ton of sense to me with the huge corpus of web standards that exist. In theory we should be able to generate a browser from those specs, it just took a massive amount of effort before. Maybe if things like this get some traction, some of the human time spent implementing the spec could be spent on creating more/better specs, allowing for even better generation too.
Edit: replies are making good points about AI capabilities and where the effort really goes. Let’s just say that I meant this in an aspirational sense, rather than where the rubber actually meets the road today.
As someone who has spent the last 3 years implementing a browser engine from scratch full time (that currently passes ~half the 200k "css" tests that test style/layout/rendering), I find that (without a lot of close hand holding) the AIs are very far from being able to do this. They'll give you something that passes the tests, but it will do it a ridiculous way be far too slow to be useful (and is wrong architecturally such that it's not going to converge on a better solution).
I guess my hope is that while this may be the case now, perhaps it won’t be eventually. If you can generate from a spec, different implementations may be slower or worse in a variety of ways but still meet the spec.
At that point it’s about optimizing and deciding trade offs, both of which might mean heavily directing the generation in some way. Over time, maybe LLMs or whatever succeeds them won’t go down so many bad architectural paths. We’ve certainly seen that already in the last year or two.
I also don’t think folks like you ever lose your value in these efforts, even if all of it came true tomorrow. At worst it would be you plus the AI, which would always be a more potent combination than say, me with an AI trying to make a browser engine.
> I guess my hope is that while this may be the case now, perhaps it won’t be eventually.
I hear you brother. I, too, hope that one of the two or three companies will be able to improve and displace everything. Who needs Linux anyway.
Hope you are on the right side of the fence.
> far too slow to be useful
Speed is probably not part of a web spec.
> wrong architecturally
Nor is software architecture (though web security specs may have some influence here).
> Speed is probably not part of a web spec.
"being a grandmother is not part of the bike frame spec so my grandmother with wheels is a bike"
Not sure why you are being downvoted, but this is definitely a limitation of current models and difficult to optimize for. Anything you are not specifically optimizing for is essentially unconstrained: the model could learn it by pure chance, but the likelihood is exceedingly low.
Techniques, like the newly announced RL-XAR from Meta [1] are being developed that will likely improve reward models and guide RL training to optimize for metrics like software architecture that are hard to verify otherwise.
[1]: https://facebookresearch.github.io/RAM/blogs/unslop/
There’s quite a few edge cases where browsers allow non-compliant code to render as if it were to specs. Because if a site doesn’t render “correctly” from a user perspective, they blame the browser rather than blaming the web developer.
I think browser performance (and security) improvements are really difficult to tackle at the scale of browser complexity, and are both art and engineering. In other words, if this were possible, we would see the results first in existing browsers. This project is very interesting though.
Maybe some day this will be the basis but not the complete solution. Agreed that there’s a lot of semi intangible art that goes into browser engine decisions. I’m guessing a lot of which have to do as much with people and current landscape dynamics as they do with the tech.
> In theory we should be able to generate a browser from those specs, it just took a massive amount of effort before.
No, not really. The absolute vast majority of those specs are human technical and technical-adjacent language, not machine-readable specs.
On top of that many web specs often invent new terminology because a lot of things are specced years or decades after something popular has taken over the term in userland.
Maybe over time the specs will become more machine readable (bounded and concrete) if we take these paths. Although in some sense everything is machine readable these days, just not deterministically so (less helpful for specs, but not nothing either).
Your second point makes a lot of sense to me too. I’m not sure how this could handle that facet of human nature, except possibly to indirectly contribute to speeding up the cycle of spec creation.
I'm looking forward to the day when we have fully functioning web browsers that we have full programmatic control over in all aspects.
With luck all chromium/blink-based browsers will go the way of the dodo bird.
I'm afraid internally they would still look like chromium/blink anyway.
Very cool idea, though I think we're a long way away from this. Hopefully they file bugs on the specs when they find ambiguities.
Also the specs are written to define observable behavior, and there's a fair bit of ambiguity that's UA defined, but to be actually compatible with the Web requires doing what Chrome does.
The good news is that you can look at the source of the 3 major engines (and Ladybird!) and your AI agent can do comparisons to figure out optimizations.
Is this named like Jev intentionally? Or is this just a new naming convention that I haven't been watching?
I'm sure this is not the origin, but Bez is a celebrity in England, a member of a 90s band called happy Mondays who didn't sing or play anything but just danced. Famous on the panel show circuit
Slander! What about the (occasional) maracas?
Those were bowling pins, it was all fake
Maybe it’s one of those weird LLM smells, like the OpenAI goblin problem.
https://openai.com/index/where-the-goblins-came-from/
Ok, but will you be able to understand the code?
This is a neat idea. It opens up a whole new way to make a web app into a native app, while also adding native features not available in web views
I have a working (and not vibe coded) implementation of a "browser engine for apps": https://github.com/DioxusLabs/blitz
So far it's only HTML/CSS (and we have our Rust-based framework to write the apps in). But it has been designed for fast incremental rendering, so it could be extended with JavaScript support quite easily.
But can it dance
Can we do the same for MacOS, iOS? I want to run them inside Linux.