AI @ W3C
By Dominique Hazael-Massieux (W3C)
See also the slides.
Transcript
Thank you all.
Hi, everyone, I'm Dominique Hazaël-Massieux, and among other roles in W3C, I'm the staff contact for the Web Machine Learning Working Group, and I organized, a couple of years ago, the Web Machine Learning Workshop.
And so, unless you've been living under a rock, you've probably heard that there is this thing called AI, Artificial Intelligence, that is making a bit of a wave.
And more recently, in particular, generative AI has been the hot topic.
So we thought it might be useful to give you an update on what's already happening in W3C in that space, and also some of the broader questions that don't quite have an answer yet, but intersect with our mission.
So we'll be looking at capabilities, ethics, and more generally, the impact of AI on the web.
Next slide.
So, in terms of capabilities, the main work that the Web Machine Learning Working Group is pursuing at the moment is called Web Neural Network API, or WebNN.
It reached Candidate Recommendation in April, and it provides a set of low-level primitives that allows hardware acceleration to run machine learning models.
That is currently being actively implemented in Chrome and Edge, and work on integrating this in the frameworks that will make use of these primitives has also started.
More recently, work that is specific to this latest generation of machine learning model for generative AI has been looked at.
In particular, support for what's called transformers, which are a key technical piece of supporting this.
So that's for the work where we have a clear and direct intersection in our standardization program at the moment.
Now, of course, if you think about all the discussion about AI, a lot of it is around its ethical impact on the rest of the world.
Next slide.
And so, part of the charter of the Web Machine Learning Working Group since inception has been to develop ethical principles for web machine learning.
And a first version of that document was developed in the course of the previous year, based on the principles that UNESCO had developed in that space, completed with some of the inputs that can be derived from the TAG, the Technical Architecture Group, Ethical Web Principles.
And we did that through a series of online discussions that were moderated and animated.
And I think the current document, while still a draft, is already a useful collection of principles, questions, possible mitigations.
But we definitely need more contributors to this work.
So if that's of interest, please get in touch.
If the topic itself is of interest, there will be a lightning talk later today on this topic, a breakout tomorrow.
I also notice the Web Accessibility Initiative released a report from a Symposium they organized on the specific intersection with accessibility.
All of these that are directly relevant to the questions that surface to us.
Again, this is about existing, ongoing work and discussion in W3C.
Next slide.
But in a way, you could say this is all looking at it from a fairly narrow perspective, because when you look at the broader landscape of the impact of artificial intelligence on the web, it is potentially huge.
The fact that generative AI now can very easily generate content of at least credible, initially, quality at a huge scale.
The fact that it also changes the way many of us interact with internet web content.
In particular, I'm thinking of the increase of chat and query interfaces.
The fact that machine learning models themselves create a new centralization risk in the architecture of our programming systems, since it requires collection of huge amount of data, the training of hugely computing-intensive resources, all of that creates natural bottlenecks in who can create and deploy these systems.
And something that, of course, has been surfaced quite clearly to many of us, the fact that it has shifted quite significantly some of the expectations about how web content, content published on the web, gets reused and repurposed beyond what at least most content creators would have expected.
So there are really a lot of questions about not just how we run machine learning in a browser, but what's the overall picture of the impact of these technologies on the web platform.
Next slide.
So just surfacing some of the questions that we, as a community, may want to consider in terms of that broader impact.
And again, this is probably the start of a conversation, not by any sense a full list of questions.
But, as I mentioned, the fact that nowadays AI content can become pervasive, or at least very easily deployed and distributed on the web, should that be surfaced more explicitly, either through technical or other means?
It's important for content creators, but it's also probably very important for models that will be trained on that content, because if you train machine learning models repeatedly on their own content, you create further risk in the quality that these models can generate.
As I said, a lot of people are saying, I'm not ready for my content to be used for machine learning training, at least within the current economical or legal context.
What should we do to make that possible?
Should we standardize a solution to that problem?
In terms of enabling the use of AI and machine learning models in browsers, the distinction between custom models and models that would be automatically provided by browsers, which already happens through APIs like the Speech API, there are some really interesting architectural questions that need to be solved.
The fact that machine learning models are huge is re-raising some of the questions about managing of large assets in browsers, and the cost of running machine learning models also raises questions around how to defer some of this workload to other pieces of the network.
Again, my small selection of questions.
Tomorrow there is a breakout organized, and I should say, not by me, but a breakout organized on the topic of the impact of generative AI on the web, where at least some of these questions have been noted as likely to be discussed.
So if you are interested, I encourage you to get there.
And finally, maybe a smaller but not less interesting question, on the next slide, how these new generative AI systems might be included, used, as tools in our own operations.
Do we want them to be?
So there may be also policy questions around them.
You could imagine them being used in the context of spec writing, tests, documentation writing, even to help with some of our outreach programs.
But none of these decisions ought to be implicit, or assumed to be neutral.
It's probably a conversation we need, as a community, to surface.
So again, maybe more questions than answers to the topic of AI web, but I hope this presentation can help at least spur some renewed discussions in that space, and looking forward to hear from you all on these questions.
Thank you all.
