5 ms·

>"The attempt at an industry wide consensus building effort around what HTML should be failed (W3C), largely because it spent most of its political capital on i

>"The attempt at an industry wide consensus building effort around what HTML should be failed (W3C), largely because it spent most of its political capital on implementation-independent academic dreams rather than incremental improvements." Interesting. Can you elaborate on what some of those academic dreams are/were that derailed things? Did they end up in the spec or were they just bikeshedded to death?

mike_hearn 5y ago
What I meant here was the combination of X[HTML] and RDF. 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.
- bogomipz 5y ago
  Wow, thanks for the comprehensive response. This jogs my memory a bit although at the time I wasn't able to pay as much attention to this as I would have liked. What a sad situation. I was curious about this: >"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." I apologize if this is a naive question but what is the connection between GPT-3, browsers and Berners-Lee?
  - mike_hearn 5y ago
    Before the 2000s-era machine learning revolution, AI was assumed to be solvable via symbolic reasoning. Pure logic, in other words. See: Cyc. Cyc was/is an attempt to encode all of human knowledge using predicate logic. If you have knowledge in this form then you can use the standard rules of logic to infer new knowledge and do reasoning. 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.
    - bogomipz 5y ago
      This is fantastic historical information. It's even more fascinating(and horrifying) to think that these things collided the way they did and derailed the evolution of a maintaining a proper HTML spec and that ultimately one company - Google shifted both the direction of AI movement as well as the direction of the browser. Thanks for the wonderfully detailed responses, I really appreciate it.

mike_hearn 5y ago

What I meant here was the combination of X[HTML] and RDF. 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.

bogomipz 5y ago
Wow, thanks for the comprehensive response. This jogs my memory a bit although at the time I wasn't able to pay as much attention to this as I would have liked. What a sad situation. I was curious about this: >"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." I apologize if this is a naive question but what is the connection between GPT-3, browsers and Berners-Lee?
- mike_hearn 5y ago
  Before the 2000s-era machine learning revolution, AI was assumed to be solvable via symbolic reasoning. Pure logic, in other words. See: Cyc. Cyc was/is an attempt to encode all of human knowledge using predicate logic. If you have knowledge in this form then you can use the standard rules of logic to infer new knowledge and do reasoning. 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.
  - bogomipz 5y ago
    This is fantastic historical information. It's even more fascinating(and horrifying) to think that these things collided the way they did and derailed the evolution of a maintaining a proper HTML spec and that ultimately one company - Google shifted both the direction of AI movement as well as the direction of the browser. Thanks for the wonderfully detailed responses, I really appreciate it.