← ASU+GSV Summit
ASU+GSV Summit
Building Better AI for Education with Sandra Liu Huang & Thomas Thompson | GSVtv Podcast
00:14:22
Loading summary
Transcript29 lines
- [00:00]
A
Foreign.
- [00:07]
B
We're here at the ASU GSV Summit 2026. We're here with Sandra Lu Huang. She's the president of Learning Commons, Mark Zuckerberg and Priscilla Chan's education initiative, and Thomas Thompson, the co founder and CEO of Eduaid AI. Welcome to the podcast.
- [00:22]
C
Thanks for having us.
- [00:23]
B
So, Sandra, you're making some big announcements at ASU GSB this week. Tell us what Learning Commons is releasing this week. And for people who haven't heard of evaluators or the knowledge graph or some of your work, how does it work to elevate the entire beard?
- [00:37]
A
Yeah. So at Learning Commons, you know, we are really thinking about how can great learning and teaching practices get to more scale? Because there's really, it's not for sure that what we learn, what we understand from the learning sciences actually reaches classrooms. And the work of Learning Commons is really to think about what are the open public goods sort of data sets and tooling that are needed that can be assets that help every ed tech developer, every district that's thinking about how to bring more rigor and quality into classrooms, do that. On the evaluators this week we have a couple announcements really excited about. But first, evaluators are basically tools that help make sure that AI generated content is evaluated against educational rubrics that are really grounded and time tested. And so, you know, this week we've got a set of literacy evaluators that we're announcing. But basically these are building blocks that help ed tech companies and organizations really understand what is the AI generating? You know, is it grade level appropriate? You know, this week's announcements are quite nuanced with. We can talk more about it, but thinking about what's the background knowledge you might need to have to access some text and knowing that I think will help us really match what students need to, you know, what will help them really grow and develop.
- [01:54]
B
Makes sense. The nuances is important. And education, I think all these evaluators evaluate different aspects of the education stack. You mentioned raising the field for every edtech developer. And this is exactly where you come in, Thomas. So let's get concrete about how evaluators work in practice. Walk us through a real teacher moment with Eduaid. Someone generates a passage, they run it through a grade level appropriate evaluator. So what do they see? What does it mean and how does that help you and the end users?
- [02:23]
C
Well, first, prior to them even seeing that, we've already tested all of our prompts and resources against the evaluators, but the user, the teacher Also has access to the evaluator as well. So they get four signals. They get the grade level, that is the band at which students could read this independently without support. They have a scaffold at grade level, a band at which students could read this with certain supports in place. They get a list of what those supports are, and then we show transparently the judgment of how the evaluator arrived at this conclusion. Then, quite simply, a teacher can one click create the differentiated version of that text for that secondary grade band. So you have an 8th grade independent band, a 6th grade scaffolding band, and then you get the resource with those scaffoldings in place.
- [03:03]
B
Wow. And so the evaluators allow you to do all of this amazing functionality without having to build it from scratch yourself.
- [03:08]
C
100%. We could do it behind the scenes.
- [03:11]
B
And for the teacher, that's amazing. And you know, so Sandra, back to you. What about eduaid? Eduaid is a really amazing tool, but what was it about Eduaid that made them sort of a right first partner in evaluating these types of evaluators? Because you're out to make something for the entire field. Eduard's way out in front already using it. What's. How is it bake a good match?
- [03:29]
A
Yeah. I think Eduaid is a great example of. Well, first of all, just like it's former educator building for other educators and just having that grounding of what is really needed in the classroom I think is really powerful. And with AI, you know, the effort to build is a lot lower. But that doesn't mean we don't need to think about quality. In fact, quality is almost a harder challenge for us in this moment. But we can also lean on AI to help us assess quality. And so, you know, I think Eduaid is really committed to thinking about what teachers need to support different types of learners and different levels of learners within a classroom even, and creating those correct sort of the right appropriate texts for them. And it really lines up well to where we're starting to experiment. And I think just the feedback that we've been getting and you know, as early partners kind of figuring this out, it's been really awesome.
- [04:21]
B
So, yeah, as Sandra mentioned, you're a former educator, you're a former classroom teacher, and Eduaid is now in 700,000 educators hands. That's quite amazing. You care a lot about quality. I mean, AI can do anything, but can it do it? Well, that's the question we're all grappling with in edtech. When you first encountered what Learning Commons was building. What was your immediate realization about what this could do for your product and what did you want to use it for? To make sure that your outputs were going to be at the level you needed them to be for your students.
- [04:53]
C
It's exactly what you're getting at. The primary bottleneck for the productive use of AI in education is validation and verification. And what the AI produces is not only grade level appropriate, but pedagogically rigorous fit to that teacher's unique style. So the tools that Learning Commons were working on just seemed like a natural fit to help us get closer to that goal. There's too many AI ed techs rely on kind of vibes or a feeling of whether or not this looks like it would be appropriate. This gives us an empirical way to actually test. Is this appropriate for this classroom population?
- [05:23]
B
Yes. Vibes are good for AI, but not for qa, not for quality. Exactly. So, Sandra, the evaluator surfaces, concrete scaffolding suggestions. Right. As Thomas mentioned, it's the grade band, the scaffolded grain band. And what scaffolds are available? It could be vocabulary to pre teach. Some teachers have described this as a transparency mechanism. Meaning for the first time, rather than just accepting or rejecting AI output, saying this seems good, vibes seems good enough for now. They have something concrete to act on. They can actually work with what underlying building blocks. So let's get semi technical here. What underlying building blocks and relationships in the data sets that Learning Commons provides to edtech providers make this type of scaffolding suggestion possible? How do you figure out what students need at any given moment?
- [06:13]
A
Yeah, absolutely. And you know, I think as we've talked about, AI is it's not deterministic. Right. So what you need to guarantee quality and think about your QA process, it's a very difficult problem. And I just really think in the K12 sector, like we have, we have, we ought to have a very high bar of that. And so the way we think about that is really sort of at a high level is thinking about what is the right criteria. How do you then put together the right data set to really capture and benchmark against that criteria? And then how do we develop the evaluation tool itself and validate that? So it's kind of three pieces on the criteria. You know, this is not in house. We have experts in house, but really we're relying on our partners and working closely with experts like student achievement partners, Anet they have decades of knowledge around literacy, how to teach it, how to evaluate it. We are working with them closely as experts, but also like the tools they've built. So they have evaluation rubrics which a human can really like go look at and evaluate text, but we want to make that more scalable. So then we go to think about how do we put together the data set that can really be a high quality benchmark against those rubrics. And essentially are both working with these partners, working with existing data sets. Common lit has a sort of a standard text passages and really kind of putting together the data set to make sure humans and some AI are. We're really sort of grounding on a precise evaluation against the rubric. Then once we have that, then we're building the evaluator itself, the prompt sort of a model of this. And we are experimenting hundreds of different ones to make sure we're tuning the evaluator to actually work well against this data set, validating that. And then those steps together kind of lead to an evaluator and hopefully saves folks like Thomas some work and can really be a building block for anyone who's presenting text into K12 context.
- [08:11]
B
Yeah, it's a powerful model. And one of the things I love about the Learning Commons work is that in some other sectors there's really a shared infrastructure. A lot of medical technology. It's not like you have everybody inventing what they're doing completely from scratch. The shared standards, it's shared underlying piping and infrastructure. And we don't have as much of that in edtech as we'd like to. And I feel like Learning Commons is really devoted to it. Question for you, this open approach that they develop all these amazing tools and frameworks and they make them available to everybody. How do you see that as an edtech vendor? Is that an advantage, a disadvantage? How do you view that kind of philosophical approach to building first education as
- [08:50]
C
a public good and a fundamental human right? And your right to have a quality teacher, quality resources for that teacher, is undeniable. So the fact that Learning Commons puts these things out under open licenses that any developer can use is part of what really attracts us to it. There's no point in developing closed infrastructure for a competitive advantage when we're talking about student learning, student outcomes, that is the priority. And we're not going to put our business interests before that. That's the primary goal. I'm a teacher first and an accidental founder. Because of that, I care about education primarily.
- [09:23]
B
Yeah, yeah, that's. Yeah, that's. That's a great attitude. That's also why they make a Fantastic partner. I feel like you both committed to quality and to actually making education better. So let's talk about these two new literacy evaluators you mentioned at the beginning of the conversation. So they're called the Subject Matter Knowledge Evaluator and the Conventionality Evaluator. Tell us about what those mean. I. I admit I don't know what a conventionality evaluator is at all, so I'd love to hear it. What do they mean and what do they mean for edtech builders and innovators?
- [09:51]
A
Yeah. So some of our earliest evaluators were around grade level appropriateness, and I think that's sort of easier to get your head around. There's like quantitative measures like Lexile and our evaluators of qualitative grade level appropriateness. The new evaluators we're launching and announced this week, they are, I think they're really nuanced and very exciting to. To be thinking about, not things that we could think about actually scalably before. Subject Matter Knowledge evaluator is actually about background knowledge. It gives developers insight into. For a given passage of text, what actual prior knowledge or background knowledge or context do you need to have to even interpret what the words are that's helpful. So even if it's grade level appropriate, but you have not played baseball or cricket, depending on your context, that's background knowledge that's helpful to expose so that teachers can just be aware of that Conventionality is about how direct the passage is communicating the text. You know, do we use figures of speech and idioms and actually, personally, as a, you know, parents who didn't speak English growing up, I did not. That was real trouble. Like, oh, that's a weird idiom. Like, what does that, like literally have no one to tell me what that means. I have no Internet. That was really challenging. Right. But now we can create an evaluator to say, hey, actually be aware there are these figures of speech. This is more direct or less direct. And just that gives educators insight into how appropriate a piece of text is for their.
- [11:21]
B
It's like how much language conventions are used in a given passage. That makes sense. Yeah. So when you hear those ideas, what sparks your mind about what you might do in Edgeraid or what you're planning to do in Nigeria?
- [11:31]
C
Well, I mean, we learn in relation to what we already know. So a prior knowledge measure is vastly
- [11:36]
B
important to a teacher.
- [11:37]
C
If a student lacks the necessary vocabulary or declarative knowledge to access a text or material in class, those are. It's a huge hurdle that a Teacher has to cross. So being able for us as developers to get that out in front of teachers so they can understand this thing I created, what do my students need to master in order to access it is huge. I mean, it's fundamental to student progress.
- [11:59]
B
Yeah, it makes sense. And it would work for students of different ages as well. Different passages could be very powerful to know the idiomatic usage and prior knowledge, what subject it's about. So as we just end up our conversation about evaluators, I'd like to hear both of you talk just sort of broadly about the role of evaluators. You know, we basically, we're talking about how AI can do so much in education, but we're now, as a field, starting to realize it's not just about what it can do, it's really how it does it, how well it can do it, how it can support teachers in practice. What do you see as the end goal of how evaluators can make all of edtech better?
- [12:34]
C
Fundamentally, when you're doing something with an AI, it has no theory of mind, it has no understanding of your classroom. So you need a mechanism to understand how appropriate any AI generated content is for the students in front of you. And as developers making tools that will go in front of students or teachers. So fundamentally it's. You have to have it in place if you're going to promise outcomes, which is what every ed tech developer should be promising as outcomes for student learning. You can't guarantee it if you can't validate it, if you can't verify it empirically beyond the standard, trust me, bro, tech benchmarks, then it's not, it's all for naught. So this is a fundamental layer that you have to have in order for a productive use of AI in education.
- [13:14]
B
Powerful.
- [13:16]
A
I agree with that. But you know, and I think really to the transparency point you made earlier, we need to think about how to make sure AI we can harness AI to be. I think to unlock all that it can be for learners, we really need to really push on that transparency. And I think having open public data sets that are public goods that we can build on really keeps that context of K12 and learning very transparent, understandable, and really can kind of keep us as a sector accountable for delivering results for our students. And we can't rely on vibes for that. I just think the stakes are too high and we have this opportunity right now to just work together across the sector to build, build and invest in these resources for everyone.
- [14:02]
B
It's a really powerful mana senju Liu Huang is the President of Learning Commons and Thomas Thompson is the co founder and CEO of Eduaid AI. Thank you so much for being here with us at the ASU JSU Summit.
- [14:16]
A
Thanks Alex.