Meeting minutes
Accessibility of machine learning and generative AI.
jgw: Introduces Josh ...
Joshue108: Can intro broadly
Joshue108: Finding scoping issues; much seems to be repair; believe we need goals beyond
Joshue108: Even blue sky, what the future could look like
Joshue108: Is that all there is??
jasonjgw: Notes scope recently updated? Would your reaction change based on the update?
Janina: scoping - had some creative ideas for future and being aware/helpful. Person with a profile, tracking, and directly helpful as a result of that. Notice certain features weren't invoked becuase use wasn't aware, offering mini courses. One possible way AI can do things that aren't easily achieved otherwise. Hoped to address some thoughts like
that. Open up TOC to leave room for things like that. maybe we're too subtle in our scoping.
Joshue108: other bits i'm intrigued by. That's what we should be looking at those things
Janina Discussed scoping, and we want to cover those topics. nice it can caption and describe (yes problems with that, less capture that), but it's the less interesting stuff. Interesting in what's possible, if we throw AI at it, what does it open up?
janina: Notes wider scope is intent; AI augmenting user skills being mindful of individual user's specific a11y needs
Joshua AI as a tool, not as it's there as fait accompli, but as an entity that's complete, it's really looking at what it can do, comes around to brute forrce processing, algorithmic potential
jasonjgw: Was looking whether GPT could provide info on how to accomplish tasks both cli or web
jasonjgw: Can answer linux cli questions; sometimes not accurate; sometimes helpful
jasonjgw: Sometimes provides answers more complex than could be
jasonjgw: tendency to generate inaccurate info is a concern; would be an elevated concern for COGA
Joshue108: will need to update my comments; phps obe?
Joshue108: Agree to expanding TOC to encompass wider opportunities
jasonjgw: Also q on where to fit ml models; obviously author side to help correct what's being said; all for the better
jasonjgw: What we worry about is auto generated content without human qos
Joshue108: Q: should address, distinctions between ml and generative
Janina: we have to be clear on what that means. I'm not clear right now, but need to be clear for the document
Joshue108: to be taken seriously we need to be clear about the distinction between AI and ML and what the strengths and weakness of each are in terms of their potential for accessibility
jasonjgw: One technique for developing AI is to use ML applying large lang learning modules
Jason: the way i think this is explained in the intro text, one technique for developing AI is to use machine learning especially neural networks (most widely used today) and one application is large language models, going beyond language, basic strategy is to predict what the next phrase/word is going to be in a sequence (prompts)
jasonjgw: Based on large artificial large models
Janina: one side is machine learning (the benefit of GPUs, how it is achieved in the machine). It's an inference approach vs heuristic rule-based model?
<Joshue108> +1 to Jason
jasonjgw: replacing how a neural system works with a probability based calculations that can be optimized to solve specific problems. Inferences run on them once created.
NLS Book: "Why Machines Learn: The Elegant Math Behind AI"
Joshue108: skirting around same thing. Generating things from stuff. Whatever info is there, AI, ML things will brute force process. CAn they anticipate user need? Change assitive tech? Hcnage support depending on context of use? How AI can be used to generate user interfaces and support users who don't have access to their assistive tech.
Joshue108: appears to be intelligent but really its just brute force calculations with some probability Markovian generated options sprinkled on top
markovian chain
rajakushalnagar: Probabilistic calculations are part of the reason why AI can hallucinate. There hasn’t been a good way to reduce this other than to merge results from different models
rajakushalnagar: we have the probability, sometimes it guesses wrong and it hallucinates. And ongoing problem especially for a11y, bc you get the wrong results.
rajakushalnagar: humans can also misunderstand and may assume incorrectly. But humans have more context typically. It's easier for humans not to guess incorrectly. With Ai, we don't know what the solution will be yet
rajakushalnagar: what's the best guess between two of them, less risk if fewer assumptions. Wrong guess for a particular model. Might be multiple models, risk of reducing...different models have to come to some sort of agreement. It's the guess between them
<Joshue108> right, interesting Raja
raja: Reducing risk by recourse to multiple models
janina: Reminds of another book I need to understand: "The Alignment Problem"
jasonjgw: ways of using them to correct each other to some degree (i think). techniques worked on recently engaging in a reasoning process. I don't know the mathmatics of neural networks. Could be interesting in improving accuracy. Area at forefront of reseach investment
janina: Has to do with optimization
raja: When models know everything about the world, results will be better; but not possible now
rajakushalnagar: fundamentally, models don't get good enough world data from print. Still missing the world contect because the model includes everything as it happens, but not possible right now. People think about what's important in the picture and speech. Interaction betweent eh two in the training is missing
… what's the most important?
<Zakim> Joshue, you wanted to say that especially in context of what Raja says it is important to keep some control so the users need is met, not the need of the AI
Joshue108: We need to talk about what serves people's needs, not the needs of the tech
Joshue108: It's novel tech, and some would want to focus on that
Joshue108: Our view is what benefits folks, not what's cool tech
jasonjgw: Persistent worry then task of evaluating system output lands on pwd, but they're least able to do so
<Joshue108> +1 to Jason
Janina: how can AI help build smarter AT? We can generalize what constitutes better AT so we don't have to rely on the person with the disablity to say what's good. Every screen seems to be loaded...here expand this, look at this, but I want to squeeze in more chrome (scaffolding around, not the browser). All the stuff the provider is happy to do
for me. This is the bump in change your attention.
<Joshue108> Stacey: I hate this purposeful friction!
<Joshue108> Stacey: I get annoyed at designers for this
[discussion of using AI to declutter an overly busy, and tangentially related content]
Janina: items of scoping we want to discuss