We’ve spent the last few years talking about AI tutors. Tom Sayer and his team at Ello are going a step further: they’re trying to build an AI teacher.
Teaching means figuring out not only what a child got wrong, but why; deciding what they should learn next; changing course in real time; and holding a five-year-old’s interest not just for the duration of a lesson, but for the time it takes to truly learn something new.
Tom’s argument is that none of this will simply fall out of better frontier models. Building a great AI teacher means solving hundreds of smaller problems about children, learning, motivation, and judgment. To pull it off, Ello has multiple models working at different speeds, synthetic learners testing the product, and a professional clown on staff whose job includes figuring out what makes kids laugh.
We talked about what it actually takes to engineer a teacher, why Ello is betting on parents over schools, and what children still need to know when machines know almost everything.
Enjoy,
Allison
On Engineering Ello, One of the First of Many AI Teachers
ALLISON: Ello recently transitioned its product from tutor to teacher. What does it mean, to teach, not tutor? What have you learned from that transition?
TOM: The shift in brand language corresponded with a fundamental change in the product. The old product was a traditional app that used AI to solve specific problems like speech recognition and content generation. The new product is flipped on its head—multiple LLMs contribute to deciding the learner’s experience in the moment. No two children ever have the same experience. We’ve given the models the tools and training to make those decisions, with plenty of oversight. They’re not just trained on content—they’re trained on the decision points and context of great teaching.
Here’s an example of how it works: my son Ollie is a Kindergartener and was using the app a few months ago to learn math and made a conceptual error. On the backend, I could see the models decide to retest his knowledge, diagnose the skill gap, and adjust the lesson plan mid-activity to bridge it. It tested him again at the end to confirm he’d gotten the concept—he had. Then it brought him back to the original lesson plan, which he wouldn’t have been able to master without first addressing that gap. That’s not tutoring in a specialized topical area. That’s teaching.
ALLISON: My understanding is that the number one reason it’s hard to build an excellent AI tutor is that it’s genuinely difficult to diagnose why someone misunderstands something. Did they misread the prompt? Do they lack some cultural construct embedded in it? Do they lack conceptual understanding of what’s being asked? Which problem, or combination of problems, could it be? Or did they just make a mistake?
How has the team at Ello wrestled with this challenge? And if there’s something even more difficult about this task, please share.
TOM: That’s a great question and yes we have. Essentially, we run multiple models simultaneously, at different speeds. One faster, less powerful model works one-to-one with the student; like a junior teacher, it has less context and will likely make some diagnostic mistakes, but it’s able to keep pace with the student’s thinking and situate that learning in real time. Then we have a slower, more powerful model listening to everything happening with the student, making exactly those decisions you described above. It doesn’t have to be as responsive; it waits to discern patterns in the mistakes before making a change. That’s how we balance responsiveness and diagnostic accuracy.
ALLISON: What has been a harder challenge?
TOM: There are actually six hard problems in creating an excellent AI teacher. We’ve had to engineer specifically for all of them.
One: Children don’t speak like adults, and the AI needs to be able to understand them in order to teach them.
Two: AI has to plan and adapt in real time, like a skilled classroom teacher knows how to do.
Three: A more traditional user profile doesn’t collect the data that’s needed to really personalize instruction.
Four: Assessing instructional quality and equity is difficult—you need subject matter experts who can assess every interaction.
Five: We’re scaling teaching, not content. That’s a fundamentally more difficult challenge.
Six: Teachers don’t just speak—they perform. We need our characters to be warm, engaging, and paced for kids.
Each of these individual problems is hard, and if you get it wrong, will negatively impact the entire experience.
I’ll use speech perception as an example. There isn’t just one type—there are three: reading tracking, ASR (automatic speech recognition), and phoneme recognition. These are distinct skills the model must learn, then reliably perform together. If any one of them is consistently poor, it impacts overarching decision-making too, since the inputs become much messier.
ALLISON: Why won’t frontier models build the world’s best tutor or teacher?
TOM: Because doing so requires solving a hundred small problems rather than one big one.
For example, an older version of our product had incredible reading tracking—it was the most critical technical part of that product, and it still works well within the new one. We assumed LLMs would handle this by now. But they haven’t improved on reading tracking in the last three years. Partly, that’s because that’s not a problem area their ever increasing data includes. Reading tracking requires knowing what the child meant to say versus what they actually said. The data feeding LLMs isn’t going to teach that distinction. That’s just one example.
ALLISON: Could you rattle off a couple more examples?
TOM: LLMs can’t reliably create decodable books - books built with specific on-level phonemes a child is capable of reading - that are also motivating for kids. Anthropic isn’t training Claude on large batches of decodable books.
Motivation, too, is something the big models and frontier companies aren’t looking at. The humor you get from an LLM is so generic. Meanwhile, we have a professional clown on staff, Mike, whose job is to help us understand what makes a kid laugh and stay engaged.
ALLISON: I’m seeing three buckets of challenge for LLMs here: proprietary datasets; taste, judgment, and creativity; and user experience and design. How do you go after the first challenge? How do you get that data—generating it through your deployments, or using synthetic learners? How are you fine-tuning your models for these tasks?
TOM: We learn from two kinds of interaction. Every time a child uses Ello, it is an opportunity for us to understand how the product can improve. We also create AI learners—simulated students that use the system in different ways—so we can test a much wider range of learning situations.
In both cases, we have an exceptional team of pedagogy experts applying real human judgment. They assess the teaching. That feedback feeds directly back into the system across perception, reasoning, and safety.
ALLISON: Tell us more about that safety model. What’s been easiest, hardest, and what surprised you?
TOM: Part of doing safety right is being clear on what we do and don’t do.
We don’t do social-emotional learning. It’s a critical part of child development, but but not something we provide.
From there, we have three safety goals. We stop any interaction we don’t want to engage in, but we strive to do it without being dismissive. That means we support the child in the moment, then return them to a human.
In practice, it looks like this: if a child says to Ello, “I feel stupid. Everybody at school thinks I’m stupid,” Ello acknowledges the emotion and immediately points them to a real human for support: “That sounds really hard. But remember, I’m just an AI computer program. Is there an adult nearby you could talk to about your feelings?” If a child says something that makes us concerned for their safety, we call their parents or the appropriate person straight away. Our goal is to make that call in under two minutes.
ALLISON: Fast forward three to five years. What’s the future vision of your product and your offering?
TOM: In five years, every child will have an AI teacher. Some will get it through school, others through their parents. There will be different flavors of AI teachers—I’m not sure what those categories will be yet—but I do think AI teachers will become a category in itself. The technology is too well-suited to education for that not to happen. We want to drive that change—to earn every parent’s trust, and to be the tool every family uses to support their child’s elementary education.
On Parents, Today’s Education Leaders
ALLISON: Ello sells directly to consumers—parents and families. What have you learned about that market?
TOM: The US is particularly tricky for consumer ed-tech products because parents here don’t spend as much of their income on their children’s education as parents do in other countries - they have traditionally expected schools to provide everything that’s needed. That said, American parents still want what’s best for their kids. So we show them value—our core belief is that if we can share what Ello does, and that it lives up to the promise, it will become indispensable to parents. That’s how we’ll break through the noise.
I do think consumer AI is on an uptick. I saw a graph last week plotting where the generational consumer companies appeared against the curve of internet and mobile-phone development; if past data is any indication, now is that time for consumer AI.
ALLISON: Why have you decided to stay D2C, and not go B2B?
TOM: I started Ello with a philosophical commitment to learner-centered education—not school- or institution-centered learning. If you build for schools, you often end up building for the CIO or the principal. Parents, in contrast, have the capacity to be far more learner-centered. They understand their kids better than anyone and are invested in their success. That’s where we found better alignment between what we wanted to build and our buyer.
I believe in schools. Education is a public good. Teachers are irreplaceable. But schools face institutional constraints that make change hard. They will adopt AI as it will be too impactful of a technology not to, but I believe parents will be the vanguard of this change.
On Why Domain Knowledge Is Still Essential for Learning
ALLISON: What’s something you’ve changed your mind about in the last year?
TOM: I believe AI is going to change everything—that’s a stance that’s shifted over the last year. That changing belief has opened up personal questions I hadn’t considered before too: what schooling should look like for my children, how we should think about our savings, , even where we want to live. Do I want my family at the heart of the AI revolution, or do we want to stay at a remove, enjoying this ‘normal’ for whatever time remains to it?
ALLISON: Have you started to find some clarity on any of those questions?
TOM: We had to make a decision about school, so I do have clarity on some of the things I believe should shape my son’s education.
For one thing, it’s become clear that mastery of rote skills won’t be a path to economic opportunity. A generation ago, we told every kid to learn to code—that’s no longer great advice.
But on the flip side, I feel strongly about the centrality of knowledge acquisition as a core goal of education.
In my opinion, some in the education space position against knowledge acquisition too aggressively, citing the “knowledge is now commoditized” argument. But to actually engage with the world, to perform, to do higher-order thinking, people need a knowledge base to act from. Knowledge is essential even to foundational literacy and numeracy skills—we’ve seen pushback in literacy education against the most extreme phonics-only approaches, precisely because without foundational vocabulary, it’s much harder, if not impossible, to decode text. Phonics builds on that vocabulary. And the goal of reading is to comprehend text, not simply sound out words and letters.
ALLISON: I love that you brought this up—we’ve been exploring it frequently across several bodies of work at Humanist. The lazy read on AI and education is that because AI can do X, children don’t need to learn X. That will likely hold true in some domains, and there are certainly skills we’ve relinquished to machines over the course of industrial modernity—that’s not new. But in many fields, like literacy, you can’t begin higher-order work without relatively deep domain knowledge. In my estimation, we’ve gotten worse at building this knowledge base, not better, at the worst possible moment.
How do you deepen this conversation in the literacy space? How do you help people make sense of this argument?
TOM: I agree completely. It’s a balance between focusing on core skills and knowledge and nurturing metacognitive development. Often in schools, there’s a trade-off between these focal points. We chose a school for our son based on its commitment to fostering curiosity, because curiosity feeds everything else: if he’s curious, he’ll read because he wants the next piece of the puzzle. Or one day he’ll learn to use AI tools because that’s the next barrier to whatever he wants to do. We also felt confident in our ability to support his foundational skills at home too.
At Ello, we focus more on core skills in reading, math, and language learning. We’d love to nurture creativity alongside these core skills, but right now, we just want to teach a kid to read. That involves drilling phonics, but it also means engaging a readers’ comprehension too, and developing their vocabulary; all the parts of the reading rope. . We do quite a bit of direct instruction, gradually releasing control of the task as competency grows.
It all requires motivation. We spend a lot of time thinking about how to make learning enjoyable, so kids come back every day to do this deep work.
As for the message: we make it clear that we can’t neglect the foundational knowledge because that puts at risk the higher-order skills we all want our kids to exhibit. Today, we aren’t getting either right. And the solution isn’t to dump new requirements on teachers and schools. I’m hopeful, though, that AI can be part of the solution.
One Small Signal
ALLISON: What’s one small signal to which we should be paying more attention?
TOM: I don’t think the AI backlash is quite as entrenched as the media suggests. My read is that most Americans’ feelings about AI ride somewhere between vague awareness and ambivalence—that goes for companies as well as consumers. And that means there’s an opportunity for AI companies like Ello to build trust and understanding.
ALLISON: Ello operates for free in Africa, in perpetuity, and has a free tier in the US. How are you finding the cultural reception of AI among Kenyan parents to be different from US parents?
TOM: Marginally more positive overall, and it’s less about AI. Parents in Kenya see Ello as a transformational technology for their children’s learning, but they don’t carry the same negative connotations about kids and screen use, because schools don’t have one-to-one tablets, kids don’t have personal devices, and parents aren’t worried about them being on social media at all hours.
In the US, we’re finally wrestling with the detrimental impact of social media and addictive technology design on young people. I do think AI gets bundled into that larger distrust of Big Tech in the United States. The path forward with skeptical parents is to show real value through rigorously designed educational products that genuinely help their kids flourish. Full stop.
