Meeting minutes
Accessibility of Machine Learning and Generative AI draft.
jasonjgw: there are updates in the repository from Janina and Scott. Still some more edits?
Janina: not yet, let's see what Scott added. Then after CSUN will create a branch to review.
… there is one vastly revamped 4.2 section. Alt text, image description and a new 6.2 section to talk about agentic AI and point to innosearch. Can say more about agentic, but will work on it
Scott: 4.2 originally gen ai would give broad info about images and alt text, but be vague. We added - it's getting more detailed, but can use the same LLM and ex on Gemini to swipe on photo you can do that three times and get three different responses with the same platform and same image. A11y paradox - if you get three different responses, which
one is right? All? None? no way of telling. And the other section looking at agentic and capturing what was discssed previously
Janina: want people to read up on those two sections before we do more edits. What we have now we should look at now. Jason and I will get references into the docuement and support for Bibtec and easy order correctly add those references in the form the tools expect them in when we generate docs
jasonjgw: we need to know what the new edit or change references are
janina: we have the bibtec from Scott need to reference those and get the markers
jasonjgw: matter of finding the places and fixing them. If you can have a first look and let me know and I can help
… we thank you Janina for that work! Once that's done we can see where the first working draft is.
… any final questions on changes to the draft?
W3C Web and AI Interest Group.
jasonjgw: follow up actions from last week on this. One, I had a look at ethical principles for web machine learning doc. Last change date on editor draft is Jan 2024. Based on that, I'd say it hasn't changed since last reviewed it as a group. I don't know or recall from their presentation whether they plan to make substantial changes. No under
active development currently.
Janina: i think they're open to suggestions. I'm inclined to read it again. I have more experience now than when we first went through the doc. Not sure if that will change anything or not.
jasonjgw: Second matter, in the presentation on behalf of the web machine learning working group, it was noted that there would be origin trials of new API, and I found out it's happening in the current BETA release of chromium based browsers. When it's an actual release, there will be web machine learning APIs available in your web browser. Trial
- i think you have to enable it. It's still experimental. But will be possible to test it for real. Saw that in the release notes.
… happening in the next month and spot on with the timing.
… and the third matter from facilitator perspective. Janina- you plan on attending more interest group meetings. what are the right steps on the relationship? Want to make sure we have a clear sense of that and when coming back to this after we next interact?
Janina: Roy is staff contact for them as well. So we'll know what's going on. 2nd monday of every month? They will rotate the time. So not sure what time next month because of that rotation decision.
jasonjgw: Scott - it might end up being in your time zone
Janina: every three months at least
jasonjgw: APA is coordinating that?
Janina: Yes. Assume we'll hear about it and we're on behalf of APA.
Scott: if it's in my time zone and I can contribute, let me know
jasonjgw: it might be late or early, but fairly within reason
Janina: should be an Asia, Australia friendly soon...but maybe more toward Japan.
jasonjgw: anything else on the interest group before we move on? Janina - thank you for your contrubitions on that
Machine learning and generative AI: observations and conclusions.
jasonjgw: this is our ongoing feedback on our own use and other information coming through about it
janina: done relatively little myself. But using gemini recently because it has a copy to clipboard. That's persuasive.
Scott: similar - using Gemini live a bit more. More interaction. Found if you want to do some deep dives in computer hardware...trying to figure out waht OS might work on something and got into a good interesting conversations with Gemini live on the options. I don't know, as with all things in doing comparisons, but it gave me some good ideas. I
think that interactivity - as long as you go in with caveats on correct or justify the reasons. Especailly as it can pull out that technical aspects...It's good
jasonjgw: there have been some articles by lenux weekly news. Can't do links - it's subscription only. One article described the problem as low quality securioty bug reports all generated by large language models. Using the systems to find false security bugs and submitting reports. and devs overwhelmed by the reports. A negative consequence. And
second, apparently agents are starting to produce github pull requests and complaining if the requests are rejected.
janina: like DOS attacks?
jasonjgw: yes. no human guidance and point made in the article. there's enough human to human communication without enlarging it with substantial agents autonomously trying to communicate with people. Another line of trouble
janina: we got the bots on their own social media platform. they generated their own religion.
jasonjgw: automation potential is very real. Can imagine - Janina alluded to this - there will be official channels for these to interact so everyone knows what it is and security controls on what it can do. but don't have that infrastructure onthe web yet
jasonjgw: some kind of channel with security controls, and question for people with disabilities who may need to use these things to interact sufficiently, the agents have the appropriate permissions that the humans need them to do (the things). if you're disability is improved by the agent, then the ability to rely/do that with the agent becomes
more imporant
jasonjgw: a w3c community group looking at authoring tool accessibility guidelines from standpoint of the use of machine learning technologies including LLMs in auth tools. User involved in this effort and includes w3c, don't know much more about it. But the group has been established.
janina: Driving facilitator was Jutta
https://
jasonjgw: the group is open to participants. If you know anyone who wants to get involved?
Scott: article said so much AI slop on youtube, AI is using AI slop to create videos. using AI slop as training data
jasonjgw: if they learn off material that another system made, it degrades performance over time. the worlds AI and model developers will need to exclude that
… article on what are the leading LLM right now. Some are available to download and run locally, some proprietary, and article notes, they can be developed in a matter of months there's a proliferation of them
Janina: one of them is restricted by the US dept of defense
jasonjgw: let's understand that the AI systems are developing fairly rapidly, but a few orgs producing the LLMs now, and some of them have the ability to process images and video. Some cases they're done with a separate model. We shoudl be aware that there's wuite a few and likely to continue
janina: yes, and far more than the brand names we know of
… what's advantage of running locally besides privacy?
jasonjgw: one is latency. Example gave was suppose trying to suppress unwanted background. Another advantage is transfer the infrastructure costs to application provider to the user. Provider doesn't have to gain access to server resources to run the model. The long tail of organizations that want to use or run models they can do it using the
user's hardware. And that's a fairly significant advantage as access to server resources can be expensive. And a podcast noted, people using agents to perform system admin tasks are starting to buy hardward to run machine learning models on instead. Cheaper. usage goes up and not cost effective to buy access from a cloud based provider
… cost reasons. and then privacy was another one.
Janina: cost one strikes me as more relevant to business than individual users, in most cases
jasonjgw: although with most users if it reaches the point where most are deploying agents in daily computing tasks, would have to consider buy the hardware or buy the cloud. And some hardware becoming available by default in new systems whether they want that or not
… phones have had the hardware in it for a while. Maybe 8 years at this point? Some of the laptop and desktop machines have had them too. Some are dedicated to neural network applications, if you need more advanced processing can buy that separately. I think what we'll see is more of that hardward will be on devices and likely to continue.
Increasingly sophisticated models run on users' devices.
Miscellaneous topics.
Janina: CSUN is coming
jasonjgw: I've always attended because I'm presenting