Hacker News Reader

Comment by Alacart | Hacker News Reader

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.

View on HN

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.

> I, too, hope that one of the two or three companies will be able to improve and displace everything. Who needs Linux anyway.

In the context of browsers the domination of one or two companies is already complete, and being able to generate new browsers with AI would hopefully allow people to escape the stranglehold of Chrome/Blink.

That said, I doubt a new,competitive browser could be created that way. The (possibly AI) developers of web apps won’t want to test on somebody’s bespoke AI-generated browser.

What does this have to do with Linux? I want more control for individuals, not less.

Do you really think LLMs will only belong to 2 or three companies in the future, when we already have fairly competitive open weights usable from home (or rented servers)? How would this be anything but the opposite of centralization amongst 2 or 3 companies, which is already the case now? Software has always been displacing things, especially itself.

I think you meant to respond to bergheim not to me.

Sorry, you’re right, I did.

But what are you writing about, LoL? Linux is just normal OS. Why did you mention Linux?

> 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"

What do you mean by grandmother with wheels?

Reference to meme:

https://youtube.com/watch?v=2G9ZGIOiPjY

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/

I recognize the feeling of handholding and have the tingles telling me that given the step change of Opus 5.5 and what passed by in the news lately, agent swarms are about to become mainstream. That's exactly what a project like this would use. I think everybody will be able to create their own engine and out of the ones who try, a few will actually make gains beyond what's available in parts of the engine.

It's a matter of "bring the best non-conflicting wins together" at that point. The existing engines are too big to flip over their codebase without having seen competitors prove it too.

My estimate is that with the release of Opus 5.7, companies need to have their shit together because the work ecosystem will flip over at 6 at the current rate of progress. This generation is the first one that delivers local applications with a better toolset than a small scale SaaS in hours to days.

Could you say which AI/LLMs you mean?

I've tried most of them. Certainly all the Open AI and Anthropic models. Fable is still the best I've tried. But it still struggles with architectural issues.

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.

That does happen, but these days they tend to add that kind of thing to the spec (and the test suite). There was a minor drama several years when WHATWG (representing browser vendors) effectively took over the spec process from the W3C because they were fed up with the W3C taking an idealistic viewpoint, and speccing things that they couldn't actually implement.

To be fair - my recollection was "representing browser vendors and pragmatists". It was far from rhe industry stitch-up that that might sound like to modern ears.

That’s good to know. The W3C probably did more to hamper web standards than anyone else (except maybe Microsoft). So I’m not surprised WHATWG took over.

Where in the web standards do you find specifications like how bookmark toolbars and menus should be organized, or that Shift-Ctrl-T should bring back a tab that was closed?

We're talking about a browser _engine_. The shell/chrome is irrelevant in this context

But some parts are far from being trivial engineering and are not part of any specs. Just consider the security infrastructure (multiprocess, sandboxing etc.) and how that impacts the overall design of the engine itself.

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.

Reply on news.ycombinator.com