Show HN: We launched a new web browser

During the era when the W3C was firmly in charge, their tech output was all oriented around making the web stack more rigorous and - as they saw it - better suited for training AI. HTML itself was more or less abandoned and everyone told to move to XHTML, the benefits of which would be the ability to embed new DSLs like SVG, app-specific markup, and the ability to express abstracted "knowledge" in the form of RDF triples. All these specs still exist of course and they were all implemented in browsers, but hardly anyone uses them.

There were several big problems:

1. XML has strict validation rules. Much, much easier to build a parser and tools for, but, not compatible with most web content which is full of markup errors. For a brief period some people heroically tried to fix all the errors in their markup and make it fully validating, but the effort involved was high especially for big sites and the reward was ... well, there was no reward really. The only tool most people care about is the browser and browsers had complicated hacky HTML parsers nobody understood. But, pixels got to the screen and they got their reliably.

Especially consider how important that is given the prevalence of HTML-by-string-concatenation. In XHTML if you made a tiny error in that process then the browser would stop and render an error page, meaning a site outage that doesn't show up in your server logs. HTML has the opposite philosophy: keep on trucking no matter what. Your page might render a bit garbled at worst, but users can tolerate that occasionally.

2. RDF/XML/Tim Berners-Lee turned out to have the wrong idea about AI. Or, well, maybe. I suspect the jury is still out on that one given that stuff like GPT-3 doesn't really meet people's prior expectations of what AI will be like, but still. The idea of expressing human knowledge in the form of an abstracted graph of nodes, in which all the nodes and edges are labelled with URIs, and then serializing that to XML and embedding it into XHTML web pages. Yeah, no. Big, big specs. No tools that actually used any of it to do anything useful.

Meanwhile in all of this HTML4 was stagnant, missing lots of small quality of life fixes that ordinary webmasters and browser makers really wanted. So you can't blame browser makers for killing off the W3C. It wasn't meeting the world's needs.

On the other hand, once unleashed from any kind of broad consensus or standards process, HTML more or less ceased to be a spec you could actually implement. You can't even say you're compliant with it because a week after your statement it might have had another 100 pages added to it.

Re: canvas+HTML5. Yeah, you can definitely go that way but there are some issues. My own analysis took me down a slightly different route. The issue with using WASM/Canvas is:

1. You need a UI toolkit that can draw into the canvas. Those are hard work. Making a simple one that gets abandoned after a year is easy, making one that's got a competitive feature set and which is maintained over the long term; much harder. Some do exist and it'd make sense to leverage them.

2. Of the toolkits that do exist, none are particularly well suited to wasm (maybe Qt could work?). In particular you normally want to write GUIs in high level GCd languages, but wasm doesn't support GC nor language-specific JITCs.

3. If you look at things developers are expressing a need for today when they go outside the limits of HTML5, it's often a combination of things like performance, OS integration, hardware integration, a wish or need to use other programming languages etc. HTML5 is a poor UI toolkit or really barely a UI toolkit at all, but that's not what's driving people currently. So the question is what would your offer be in the short term if you're restricted to wasm and canvas?

4. Finally, if you look at where HTML5 is weak and not innovating, or not even in the game at all, in my view there's lots of areas but they mostly require you to be outside the browser to fix them.

So I think the way to go is to make being outside the browser a much more hospitable place and in that way encourage people to build competing neo-browsers. Let 1000 flowers bloom, if you see what I mean.

TBL/W3C felt like the next big upgrade for humanity would be AI that could understand the web. To achieve that, you need web pages to be encoded in the form of logical triples. Cyc used a custom lisp-ish language called CycL to do that, RDF was the same concept but 'web-ified' i.e. using XML and URIs for everything.

Of course that approach never took off. Even by 2003 Google was starting to master machine learning techniques like scalable logistic regression, the webbiest of web companies just didn't care about this symbolic AI approach at all because in reality nobody was using it. Too complicated, too abstract, no tools or cool demos. Pure vapourware in other words. GPT-3 has now shown that you don't need everyone to learn new syntaxes or predicate logic to train AI on the web. You can rely purely on statistical techniques and neural networks.