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).
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.
Alacart · · focus · HN ↗
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.
nicoburns · · focus · HN ↗
warkdarrior · · focus · HN ↗
Speed is probably not part of a web spec.
> wrong architecturally
Nor is software architecture (though web security specs may have some influence here).
nsagent · · focus · HN ↗
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]: <a href="https://facebookresearch.github.io/RAM/blogs/unslop/" rel="nofollow">https://facebookresearch.github.io/RAM/blogs/unslop/