Week in Edtech 4/8/26: Anthropic’s Mythos Sparks AI Security Concerns, EdTech Efficacy Debate Intensifies, Screen Time Backlash Grows, OpenAI Faces Pressure, AI Reshapes Entry-Level Jobs, Higher Ed Adapts, and More! Feat. Yoon Yang of Pensive

Join hosts Ben Kornell and guest co-host Matt Tower, as they break down the biggest stories shaping AI, edtech efficacy, cybersecurity, and the future of work in education.

✨ Episode Highlights:
[00:00:33] Rapid rise of AI-generated “vibe coding” raises concerns about software vulnerabilities and cybersecurity risks
[00:02:44] ASU+GSV preview highlights the importance of relationships over deal-making at major edtech gatherings
[00:08:26] Anthropic’s Mythos model withheld due to its ability to uncover critical security flaws across systems
[00:11:37] Growing need for AI systems to defend against AI-driven cybersecurity threats
[00:14:18] Schools question edtech effectiveness amid too many tools and limited evidence of impact
[00:15:23] Debate over screen time intensifies as some classrooms move toward eliminating devices
[00:20:47] Discussion on whether smaller, more personalized school models better serve students
[00:26:08] OpenAI faces leadership changes and increasing competition from Anthropic and Google
[00:30:30] Big Tech’s varying levels of investment in education reshape the competitive landscape
[00:33:25] AI disrupts entry-level job markets, raising concerns about college graduate employment
[00:36:12] Future workforce will demand adaptability, entrepreneurship, and continuous learning

Plus, special guest:
[00:37:34] Yoon Yang, CEO and Co-founder of Pensive, on AI-powered grading and personalized tutoring in higher education

😎 Stay updated with Edtech Insiders! 

Follow us on our podcast, newsletter & LinkedIn here.

🎉 Presenting Sponsor/s:

Every year, K-12 districts and higher ed institutions spend over half a trillion dollars—but most sales teams miss the signals. Starbridge tracks early signs like board minutes, budget drafts, and strategic plans, then helps you turn them into personalized outreach—fast. Win the deal before it hits the RFP stage. That’s how top edtech teams stay ahead.

Tuck Advisors was founded by entrepreneurs who built and sold their own companies, frustrated by other m and a firms, they created the one they wished they could have hired but couldn't find. One who understands what matters to founders and whose North Star KPI is the percentage of deals closed. If you're thinking of selling your ed tech company or buying one contact Tuck advisors now.

Cooley LLP is the go-to law firm for education and edtech innovators, offering industry-informed counsel across the 'pre-K to gray' spectrum. With a multidisciplinary approach and a powerful edtech ecosystem, Cooley helps shape the future of education.

Innovation in preK to gray learning is powered by exceptional people. For over 15 years, EdTech companies of all sizes and stages have trusted HireEducation to find the talent that drives impact. When specific skills and experiences are mission-critical, HireEducation is a partner that delivers. Offering permanent, fractional, and executive recruitment, HireEducation knows the go-to-market talent you need. Learn more at HireEdu.com.

Transcript

[00:00:00] Alex Sarlin: Innovation in pre-K to grade learning is powered by exceptional people for over 15 years. EdTech companies of all sizes and stages have trusted higher education to find the talent that drives impact when specific skills and experiences are mission critical. Higher education is a partner that delivers offering permanent, fractional, and executive recruitment.

Higher education knows the go-to-market talent. You need learn more at higheredu.com. That's HIRE edu. 

[00:00:33] Ben Kornell: And I think what is happening at the same time, which is even more concerning, is that more and more of the world's products and internet are gonna be built with Vibe code, which is probably more permeable to bugs and errors and assault than in the past.

Maybe not. Maybe the total inverse is true. It's actually way easier to check. But the rate and speed at which people have been pushing new products and new code is unparalleled in our history. And meanwhile, we have essentially a cybersecurity weapon that could be unleashed. 

[00:01:12] Matt Tower: I do believe our access to technology provides a fundamentally better experience for students today.

I think it is way more fun and interesting to learn today than it has ever been, but I also think that most ed tech tools probably don't work, which is like a weird and sort of scary thing to say.

[00:01:33] Alex Sarlin: Welcome to EdTech Insiders, the top podcast covering the education technology industry from funding rounds to impact to AI developments across early childhood K 12 higher ed and work. You'll find it all here at EdTech Insiders. 

[00:01:49] Ben Kornell: Remember to subscribe to the pod, check out our newsletter, and also our event calendar, and to go deeper, check out EdTech Insiders Plus where you can get premium content access to our WhatsApp channel, early access to events and back channel insights from Alex and Ben.

Hope you enjoyed today's pod.

Hi everybody. It's Ben and Matt, and we are here with EdTech Insiders, Week in EdTech. Matt, it's always great to have you on the pod, and we have a lot to cover on A-S-U-G-S-V Eve. I think generally the official start is like Monday, but we count it as a Friday start. Everybody gets in in town on the weekend.

What are you most looking forward to that's not just attending panels or doing one-on-one chats when you head out to A-S-U-G-S-V? 

[00:02:44] Matt Tower: That's a great point, Ben, particularly that. It sort of bleeds into either side of the week, and there are some people who are like Saturday to Tuesday or Saturday to Monday people and some who are Sunday to Thursday and some who are Monday, Tuesday.

It's all over the map. I think what I've appreciated now having gone for a number of years in a row, is that creativity that folks put into making their little corner of the space interesting. Whether it's a Padres game or walking the waterfront, or Cooley renting the aircraft carrier, there's all these fun things that happen in and around the conference that I think gives some sort of levity to it.

So I will be surfing Sunday morning and any listeners are welcome to come meet me at Ocean Beach. We'll grab a wetsuit from Ocean Beach Surf and Skate, including yourself. And I think that's the best part about it to me is I see folks like you every year and we get to have some fun with it just as much as we do the serious professional stuff.

[00:03:42] Ben Kornell: Yeah, I mean, a lot of people talk to me about A-S-U-G-S-V and maximizing the game plan and what deals are gonna get done, and at least current state, I find very little deal action actually happens at A-S-U-G-S-V. It's so condensed. It's so overwhelming that actually the value is these touchpoint, the human touchpoints and relational touchpoints of just continuing to stay connected with people.

And if you had some m and a conversation happening before, it's a great one to touch base on it or an investment. Conversation before it's a great time to touch base on it or to spark a new conversation. But ultimately it's so crammed and with the back to backs it makes it so chaotic and the happy hours are so A DHD for me.

So, but I am looking forward to seeing everyone. I will note that we have a little pre pool party over at my place from one to four on Sunday. And it's interesting to like be now living in San Diego and having people in my backyard. And then Monday night is our big A-S-U-G-S-V event where we will have 500 of your best friends in EdTech, all on the rooftop of the Marriott, Marques fourth floor, hanging out, having drinks.

Who knows, maybe there'll be a little karaoke put in there. I can't really say. All I know is it sounds dangerous. Micro, there will be a microphone and we know what happens with EdTech people drinks and a microphone. But anyways, that's the fun part and that's what really creates the memory. 

[00:05:12] Matt Tower: I totally agree, and my advice to most newcomers now is set up the time to shake somebody's hand.

Don't expect to cover a lot of ground, but then write 'em an email after saying, Hey, it was really great to like actually meet you in person. Let's set up a zoom to talk about something substantive. And that's, I think, worthwhile strategy to follow. Yeah. I'm wondering, are you expecting any deals? I feel like GSV used to be like the pinnacle of the EdTech deal announcements.

I, it was to you, edX was the highlight of the conference four years ago. Three years ago, gosh. Matt, what do you think? 

[00:05:48] Ben Kornell: I mean, this is a great, like instead of our end of year predictions, let's do our A-S-U-G-S-V predictions. If there's a deal that's gonna go down, what is the deal that's gonna go down?

Okay, so I'll give you my formulation for the deal. Is somebody with great distribution acquiring somebody with great technology, because that's the combination that I think makes the most sense. So HMH acquired NWEA about two and a half years ago around assessment. I would love to see them acquire one of the upstart AI companies.

I could go with a diffit, a snorkel, somebody who has created expertise in more of a narrow band. Looking to just blow it out across the network. And I think you've got HMH who's probably thinking, okay, we need to IPO in like a year or two, or we're gonna be back in another transaction cycle with private equity.

So somebody of that regard I think could be a really interesting time and maybe we will finally see some of this industry consolidation that follows the Let Every Flower Bloom moment of Early ai. What about you? What do you think? 

[00:06:59] Matt Tower: Yeah, first, I think that's a really great prediction, number one. Because HMH has a history of AI related transactions, both venture investment and acquisitions.

So I think they're definitely a company to watch. And I also think that crop of early LLM AI companies like Diffit and Snorkel, you could insert a handful of other names in there too, and wondering what their next steps are is top of mind to me. So I love that prediction. I think 

[00:07:26] Ben Kornell: it's spot on. Part part, part of why I picked those two is also this connection to assessment and like where is assessment heading?

And if you had the NWEA library of assessments and an AI infused enablement, what could that do for the space? 

[00:07:41] Matt Tower: Yeah, I think mine, thinking about the venture side of things. I wouldn't be shocked if one of the more outsider AI education companies raised a big funding round. So I think it was not quite around GSV that Outsmart did their big round last year, but something of that ilk of folks who have credibility in the space raise a pretty big round for our size and try to do something interesting.

So that's something I'm on the lookout for. Around bigger consolidations. It is a little bit harder to see. You already had Coursera and Udemy, so I wouldn't expect something of the two U edX ilk, but crazier things have happened and I will be excited regardless to be reading the Newswire on Monday morning.

[00:08:26] Ben Kornell: Yeah, for sure. Well, speaking of Newswire, we've had a lot of news in the last week on the AI front. The big one that's coming through is around Mythos, the new AI model from Anthropic. First help our readers understand what's the big deal here and what's your take on it from an EdTech perspective. 

[00:08:46] Matt Tower: Yeah, so at base level, mythos is a general model developed by Anthropic, so it's of the same genre as Sonnet, which a lot of people will probably be more familiar with.

In that it covers all shapes and sizes of the tech industry. And Anthropic is not releasing it for public use very specifically because it was so good at finding security vulnerabilities across what's quoted as every Major OS and browser that the philanthropic team was not comfortable releasing it to the public.

So instead, they are putting together a consortium of, I think tech players, for lack of a better way to put it, too long, didn't read to help them fix these vulnerabilities. So we've now reached the point where large language models are better at finding security vulnerabilities in the software we use every day, every major.

Operating system and browser than humans, and it's now there is a really strong, moral, ethical, whatever judgment call for these model providers to make in how you release them so that it's done in a way that hopefully doesn't break our economic system and whatnot. Is there something you would add to that?

[00:10:04] Ben Kornell: Yeah, I mean, this is where we are today. Now imagine where we might be two years from now. If you've got the ability for models to self-replicate and self-improve over time. Like this isn't a single immovable object, this is a vector. And I think what is happening at the same time, which is even more concerning, is that more and more of the world's products and internet are gonna be built with Vibe code, which is probably more permeable.

To bugs and errors and assault than in the past. Maybe not. Maybe the total inverse is true. It's actually way easier to check. But the rate and speed at which people have been pushing new products and new code is unparalleled in our history. And meanwhile, we have essentially a cybersecurity weapon that could be unleashed.

Yeah. And so like in that arms race, there's always this need for equilibrium that every time you get an advancement, the industry has time and opportunity to react. But if we're on this self-replicating trajectory. It's only a matter of time before those with mal intent get a leg up and they're not going to slow down.

They're not gonna stop. And so almost makes me think that ultimately in the modern tech stack, you need like an AI layer that's constantly checking your code for vulnerabilities and you're just gonna need to have that as part of your natural tech stack anyways. 

[00:11:37] Matt Tower: Yeah, and I think as a consumer and or as a business.

The onus is really what I hope these types of mythos consortiums bring together is standards around how to share your data. Because again, it getting hacked is one thing. Understanding what is specifically vulnerable to you as a user is another right. And I think that is sort of the operative question for like you and I, right?

Of what am I willing to expose and how, and do I understand that even more so than do I understand the software that I'm using? So like, is it okay to share my name with FIFA's vibe coded ticketing website, which does not work particularly well for mm-hmm. As someone who wants to buy World Cup tickets.

But I think really, it's hard for me to imagine that. AI is gonna keep AI in check from like a vulnerability perspective across the whole open internet, and it will be more, can we protect the actually critical personal information of individual users and build standards around that so that when things get hacked, it's more like, oh, all this data is junk or inaccessible.

[00:12:49] Ben Kornell: Yeah, I mean, one thing that we've talked about many times before, both you and I, but also with Alex, is this concept of AI slop. And that's a content conversation generally, but what we now know is that so much of code is being produced by AI that we can't tolerate AI code swap. And this is where I feel like anyone who has personal information.

And especially with kids, you have to have a much, much higher bar around that. And sometimes I wear my school board hat, this is where as a school board member adopting any new products, just feels like, uh, well we need to pump the brakes here. This new AI product, what are the vulnerabilities? No one really, really knows.

So I think the rate of change is speeding up so much that the rational response for educators and education institutions might actually be to slow down. Even more than they have already. 

[00:13:50] Matt Tower: Yeah, I agree. It makes me a little bit sad because theoretically these tools give access and optionality to students and teachers on a level they've never seen before.

But like you're right. The threat of your data getting out there or somebody maliciously manipulating you via AI is huge. And I think to your point, the rational response is to pull back. And I think we're seeing that as evidenced by some of the articles we were talking about this week. 

[00:14:18] Ben Kornell: Yeah, so it does segue, unfortunately.

Really well to a lot of questions and concerns about screen time and EdTech efficacy. Ed Surg has a really great article talking about schools stepping back and seeing so many tools, so little evidence of impact. And second, there's a wave of articles about screen time. The one that I think has captured most attention was the article about the teacher who basically cut all screens from their classroom that's in the Atlantic.

It was also a blog post shared by Dan Carroll. So shout out to Dan for being early on that story. But we read all these headlines and sometimes it's the human interest side of things where people are saying. Look, a world with no screens has some real opportunities. Let's take these in two pieces. What's your first read on the EdTech efficacy and this abundance of tools and lack of evidence claim?

Do you feel like that's accurate? Where's the takeaway or the nuance here? 

[00:15:23] Matt Tower: Yeah, I think it's super squishy. I think it's like two things that I think are true at the same time is that one, I don't think school was fundamentally better in the 1970s and eighties. I think it left a lot of students behind and did a disservice to a lot of students.

I don't think our baseline was that great at the same time. Yeah, I think a lot of tools are drunk and if I was a teacher. I totally understand how I would feel overwhelmed by all of the options at my disposal and just throw out my hands of like, well prove anything. Right? And it gets back to the thing that I, I feel like I think about constantly, which is the best way to improve test scores is to give kids an apple before the test.

Right is make sure they're well fed and slept enough hours the night before, and then you can get into the ed tech. So, I don't know. I struggle with, I do believe our access to technology provides a fundamentally better experience for students today. I think it is way more fun and interesting to learn today than it has ever been, but I also think that most ed tech tools probably don't work, which is like a weird and sort of scary thing to say.

[00:16:35] Ben Kornell: Yeah. So three thoughts as you dive into my brain on this topic. Number one is what's ed tech? What's not ed tech? Basically, all ED education has technology in it now. So this differentiation of ed tech tool or not ed tech tool, like all curriculum providers, all use technology. Our lives have integrated with technology, so I think it's a false dichotomy.

As we even call these things out, we joked whether we should rename EdTech insiders, education insiders, or learning insiders, because EdTech almost as a frame, it works from an investor standpoint and a industry standpoint, but functionally in classrooms, it's a misnomer. Second, I would say. I disagree with your statement that many of these EdTech tools don't work.

I think the question is, in what context do the tools work and in what context do the tools not work? And this is not an implementation blamer thing, but what I would say is I think there's strong intention from everyone who's building these tools to make things that work for learning. And I think the vast majority of them, I find using learning science and using research backed processes.

But in the combination of how I rolled out the product sets and features, the device that it's on, the time of day, the implementation, the rollout. We have the 5% challenge, which Lawrence Hole eloquently wrote about, which is basically even with efficacy studies, only those that are implemented well with engaged students show real longitudinal results.

And then I think the third one is that I think that taking the tech out of it and just saying, when are schools most effective at teaching and learning? It's when there's coherence, when there's alignment. Yeah. When the programmatic elements all flow and connect to each other. And I do think that the point the article makes most importantly is that having all of these tools actually creates a misaligned and discombobulated learning experience rather than a coherent structured and thoughtful one.

And so I applaud the people stepping back and saying, okay, how are we going to thoughtfully, intentionally construct an experience with education? Resources, which happen to all have technology. So they're ed tech together with our teaching and learning and and so on. And that is the work that is right to be done.

That also leads into the screen time element. So not to dominate the mic here, but just on the screen time front, I think the evolution has been social media, bad cell phone screen, bad computer screen, bad. And I think that, that it's such an alluring narrative and all of these cases, the technology itself is not the bad thing.

It's how it was used or how it was rolled out. And social media for so many kids was the first time, and I grew up in southern Indiana. For many kids, it's the first time to connect with other people around the world that are like them. There's so many goods that come out of some of these things, but the harms have to be balanced or minimized, and especially with kids.

And so I think we've gotten into this like black and white ban or don't ban kind of world. And what we really should be doing is digital literacy training to empower students over time to use the tools effectively and minimize harms. And I just think we're not willing to do that work with our kids and with our school teachers and our systems like that.

It's hard work to take fire and figure out how to cook with it rather than burn your house down. But yet here we are like thousands of years later and fire's been one of the best things that's ever been invented, I think. That is a technology just like these other technologies are. Yeah. I'm a little bit concerned that we're just going into this banning space and ultimately what we know when things get banned, those who understand good use, who tend to be more affluent or have access tend to accelerate ahead and others are left behind.

[00:20:47] Matt Tower: Yeah, and I would add, it's even harder to figure out how to do this programmatically, right? So it's hard to do that, you know, at the state or the national level more than anything. And I think it gets to a question that I've been sort of spinning on that I'll, I'll explain in a, in a second, but the question is basically should schools just be way smaller?

And this question gets at a couple of themes that are relevant today, right? Like obviously school choice sort of defaults to smaller providers figuring out what the future of the public education like. Physical plant looks like and how, you know, the buildings we built for school districts in the 1950s may not be appropriate for today's schools.

I think it gets at what you're talking about with screen time where you know, you and I might not actually agree on how much screen time is the right amount of screen time or like what settings are appropriate for screen time in. You know, we might even disagree on the difference between four and six year olds in the amount of screen time.

And I think like all of those decisions sort of lend themselves to smaller. Cohorts and you know, I think it's maybe the pendulum swinging away from, you know, both the way the school system has operated, which is sort of factory industrial, you know, trying to make as big a school as possible and the like school choice trend of let a thousand flowers bloom.

I'm just like, is the answer. Just smaller schools and like that allows people to find a place for their children that is more closely aligned to their values rather than just trying to do everything scaled. 

[00:22:30] Ben Kornell: Yeah. I mean, this comes back to a fundamental underlying question of this whole debate, which is what is quality?

What does quality look like? You talked about test scores. From my perspective, test scores represent a floor or not a ceiling, and maybe the number one way to gain test scores is an apple. Another way to gain test scores is to dumb down the test, 

[00:22:51] Matt Tower: and I'm sure you guys have covered the math stuff, and like graduation rates relative to math, literacy.

[00:22:56] Ben Kornell: Yeah, and also like just finished reading Ted Den Smith's aftermath and running a math company and has been thinking a lot about what does it look like to raise the ceiling rather than just raise the floor. So if you're saying that small models are better delivery mechanisms for schools to meet the unique needs of each learner, I could be convinced if you had a reasonable way.

Of assuring quality control. But what ends up happening is if your only quality control measure is gaming the test system, you're much more likely to get these gym workout routines. Yeah. That drill and kill towards tests, which ultimately longitudinally don't lead to long term student gains. So I think this is where our accountability, if you're gonna have a more of a marketplace and you're gonna open up your delivery models to smaller bespoke players, then you have to have like accountability environment that's much more like real time and much more effective at making quality, transparent and assessing it.

[00:24:05] Matt Tower: And I would add the explicit choices too, that a school like this explicit, like whether a cell phone is in fact banned or if it's like, you know, geofence so that only classroom maps can work. It's both harder to track in some senses 'cause you have more institutional entities, but easier to track in that each entity can be more explicit 'cause they're not designing for the lowest common denominator.

So to be clear, I don't actually know where I stand on that question and I think I appreciate you sort of riffing on it with me. It just feels like so many of the trends, at least on the investment side of this space, are pulling towards that like more smaller schools model that it's worth noting. 

[00:24:48] Ben Kornell: What is holding us back from seeing that future play out And you know, obviously because the US is federated into like state departments of ED and even smaller school boards, like where are you already seeing some of this play out?

Because I think the other thing that we haven't demonstrated in education is an ability to effectively measure, learn and, you know, communicate the results in a way that functionally positively impacts other systems and, you know, spreads our best practices. 

[00:25:23] Matt Tower: I think because of that, that because we don't actually know how to really measure what a good school is, what we default to is time and like how long a school has existed, which is like on some level I, I get it, but it is sort of a weird heuristic of like, oh, this school's been around.

Harvard's been around since 1600. Therefore it must be good. Maybe Harvard's the wrong example, but I think we do often default in our choices to what has existed and how long it has existed, rather than trying to get at answering the questions you just posed, which are like, what makes this school good?

How do I evaluate this on apples to apples against other options I may have? And I don't think we have a good answer for that today. 

[00:26:08] Ben Kornell: Yeah, we're going to keep covering that here in EdTech Insiders. I'd say meanwhile, what also is changing is potentially the delivery models themselves. One of the other headlines that we had in the AI segment is really about open ai and I don't have a specific news headline our listeners will know.

Last week we covered that Sora was pulled from the open AI lineup, but it does seem like there's a lot of smoke here. And Matt, if I say one thing, you've always been our expert on seeing the smoke signals in the future and figuring out where the fire is. You called bye Jews. There's been a couple other ones where, where your crystal ball was quite good.

Where do you think things are at with open ai? Where do you see that going and what's the impact on education? 

[00:26:53] Matt Tower: Yeah, so, and I, I think it's also worth mentioning that there was a big New Yorker article on Sam Altman, the CEO of OpenAI that, you know, had a lot of opinions. I don't think it's really worth hashing out like all of the details into it, but is, you know, probably worth reading for the readers and notably so.

Executive changes that have happened in the past two weeks are the CEO of applications. The former, I, I'm gonna botch the pronunciation of her name, but she's very highly regarded executive who came in from Instacart and was, uh, going to lead the business side of OpenAI, had to take a leave of absence for sickness.

The chief marketing officer also I think, went out on personal leave and the COO was pushed to a different role, like a special advisor type role. So we have four senior leaders at the company who are facing headwinds, both physical, which is super sad, to be clear, and, you know, professional, which is also sad, but in a different way.

So I think my opinion is like the company is gonna change, right? When you have executive transitions at a company of chat, PT of open eyes, magnitude, like the business changes based on who you install in the role they bring their. Previous experiences with them to these roles. So I'm certainly watching who the new senior leadership is going to be.

We've seen in their product roadmap that they focus much more heavily on coding to match anthropic and enterprise sales. Again, specifically to compete with anthropic. Anthropic this week announced that they had grown from $9 billion in run rate revenue to 30 in the past, I think like three months, which is hard to comprehend the magnitude of, of that growth.

We saw some jokes about how it might have been driven by meta employees competing to spend the most tokens, uh, which is a whole, whole other category of, of like, I don't know about that. But the moral of the story for me is like, I do believe OpenAI and chat PT specifically are going to change. Based on the folks who step into these new leadership roles, whether that change will be good or bad.

I don't know. I think it would be pretty hard to say, but I think they're certainly feeling the heat from Anthropic specifically and probably on a different access, Google who just continues to pump out new models and, and new capabilities deeply ingrained into the Google suite of products. 

[00:29:22] Ben Kornell: Yeah, I mean, a couple of points that really resonated with me that you mentioned, you know, the revenue piece, anthropics revenue, a quarter of the expense.

They've surpassed open AI's revenue. Yeah. That's meaningful as a barometer here. Yes. Even though open ai, if it didn't have anthropic, everyone would be saying Open AI's growth has been like, you know, stratospheric and unprecedented. Like you said, I think the arrival of the new executives signals focus, and I think it's going to actually mean less noise in the space for ed tech companies to compete with in K 12.

It doesn't seem to me like OpenAI has a lot of interest in pursuing K 12. Formally, they announced partnerships with the teachers union haven't heard much since then. They announced a few partnerships with a few schools, or a few districts haven't heard too much. And you know, with higher ed, it doesn't seem like it's necessarily chat GPT for your classroom instruction.

It's much more like you are a large business. You should be using this as your AI provider. 

[00:30:30] Matt Tower: It's just chat gt. Like it's not cha gt for your classroom. It's chat GT like period. 

[00:30:34] Ben Kornell: Yeah, 

[00:30:35] Matt Tower: full stop. 

[00:30:36] Ben Kornell: That's right. And in that game, they're in an uphill battle with Anthropic. So my biggest point would be trust.

This is actually the whole thing going down with the Department of Defense was the first sign of this. And then if you read the article in the New Yorker about Sam Alman, there's really no new information there. It's just putting it all together and saying, is this person trustworthy? And as like soft as that sounds, what an important currency that is in this moment.

And Google has trust. In part because they're reliable, they're big, but they're also everywhere and all your shit is in them. And so they're gonna play on that advantage that, look, you already trust us with everything anyways. We're just adding this as new features to stuff you already use. And Anthropic has thoughtfully and skillfully engineered a trust paradigm, primarily starting with their B2B partnerships.

But now all the coders trust Claude Code. And I would say the level of pissed offness with people at Anthropic is not anything they're doing with the models or any company drama. It's that the tokens are, they're running out of tokens, they want more. And as much as that, you know, you never wanna run out of your supply for, uh, incredibly high demand.

Leaving your customers who just are voracious and want more, that's a good problem to have. And so I do think OpenAI is kind of caught in the middle here. And it also, your, your last comment about meta and the like token races and things like that, it also just goes to show what a joke that's become and how sad because you know, for our EdTech friends out there, there's so much that early days, the meta and Facebook and CCI were doing for our EdTech universe, I still, I love everything that Learning Commons is doing, but there's no direct credible AI support for education coming from Meta the company really anymore.

And I think that that is, having less of that competition is just going to be a bad thing overall for the quality of the models. 

[00:32:46] Matt Tower: Yeah, I mean, for me. When I look at the big tech companies and their relationship to education, at a very basic level, just look at how, at their staffing, right, and how many resources do they have committed to the education industry.

You could look at that by like token allocation, but I, I look at it by headcount and it's pretty clear that the stack rank is Google, Amazon, philanthropic, OpenAI, meta. And Microsoft's somewhere in there. I, I need to check on their headcount, but like, you know, that's a pretty easy heuristic to gut check via LinkedIn.

Unfortunately, you can't use tokens to do it on LinkedIn, but you, you can get there pretty easily 

[00:33:25] Ben Kornell: someday. So our last topic is really around what this future looks like for workers and for our college grads. We had a series of posts about basically people graduating into joblessness as well as these like niche areas where everyone's getting employed, like accounting.

How is accounting still hiring people when AI can do almost all accounting? But it's a fascinating time where basically the job board, especially the entry level job board, is getting totally remade. You know, higher ed organizations are trying to rapidly shift and evolve so that their students can get the jobs that remain and the new jobs that open up.

What's your take on all of this upset and hubbub about college graduate joblessness? 

[00:34:15] Matt Tower: Yeah. So as a millennial, I am still bitter about being called a job hopper by all the boomers. And this, this was a, a big theme in the 2010s was the millennials couldn't stick to one job. They were the worst. They were just constantly begging for promotions.

And when they didn't get 'em, they went to another job. And, you know, when somebody finally sat down and looked at the data, it turns out millennials and boomers job hopped at almost exactly the same rate as like when boomers were in their twenties and millennials were in their twenties. It was almost exactly the same, like tenths of a percentage point difference.

And the narrative was made up by boomers who had forgotten their twenties to be, you know, a little bit spicy about it. And I think what we're seeing here is a similar narrative emerge of like. You know, yes. Some of the jobs that were available to the two of us and most of your listeners when we were graduating college no longer exist or will not exist for much longer.

You know, accounting is, who knows? Computer science is more interesting to me because, you know, I think it could turn into a lot more computer science majors. 'cause everybody's a developer now. It's a different skill set to be a developer. I think the jobs are just gonna look different. I, I read an article two weeks ago, three weeks ago about all the teenage millionaires from Roblox.

They're building Roblox games and making millions of dollars selling Roblox games to other people. You and I, I don't think could even contact you have, you have kids who, who might be more into Roblox. So you might have like some inkling of how that happens. But that wasn't a job that we looked at coming outta college.

And yet it exists and it's obviously not everybody making games on Roblox is making millions of dollars. But the point is the jobs are gonna look super different and super new and. Again, like just because the jobs we wanted coming outta college aren't there, doesn't mean there aren't no jobs. So I think this is gonna turn out to be a big nothing burger of a story.

[00:36:12] Ben Kornell: It may turn out to be a nothing burger as the things play out because you know kids are adaptive. I think what the story that I'm most interested in is how do institutions change, evolve and adapt to support their students to be most successful. And I think that is going to be a real story. Those that do and that figure out how to prepare their kids to be more entrepreneurial, leverage dynamic skills, be constant learners, they're gonna have a higher success rate with their graduates.

And there is a way in which the ROI on higher ed is under scrutiny more than it ever has been before. Those that win in the game of preparing their kids for the future will also win in terms of scaling their organizations and institutions. Unfortunately, that's about it for our show. We have to wrap here.

But we will all see you on the surfboard coming up this weekend. And for those of you who haven't signed up yet, for the EdTech Insiders happy hour, we're gonna be releasing 50 more slots Sunday night. So sign up right now. This is your last chance and hopefully we'll see you there at our A-S-U-G-S-V Happy hour.

Thank you all for joining, and thank you Matt Tower. As always, great to have you as a co-host. 

[00:37:31] Matt Tower: Thanks, Ben. I appreciate it. Talk soon.

[00:37:34] Alex Sarlin: For this episode of EdTech Insiders. We are here with Yoon Yang. He's the CEO and Co-founder of piv, which applies AI to Education. Formerly one of Korea's youngest AI research scientists at AI Tutoring Company rid.

He published LLM focused AI papers from his first year at uc, Berkeley, where he earned an EECS degree before founding pensive to tackle higher education challenges with ai. Yoon Yang. Welcome to EdTech Inside. 

[00:38:03] Yoon Yang: Thanks, Alex. Thanks for having me. 

[00:38:05] Alex Sarlin: I'm really happy to talk to you today. So that is a very impressive bio.

You started as one of the youngest AI researchers. You were publishing LLM research very early on. Tell us about your education and if there was a moment when you realized, oh, this technology is so powerful, it could really be applied to educ. 

[00:38:23] Yoon Yang: So I was born in a small island in South Korea, and there were multiple situations where my entire family was moving to an entirely different education environment where when I was going to a fourth grader, I moved to Seoul, which is one of the biggest cities in South Korea.

And then for high school I went to a private high school as well, and boarding school, and for university for the first time ever in my family, I was crossing the Pacific and coming to uc, Berkeley to attend college. So every instance of my education journey was defined by. A lot of different changes and different environments and there I firsthand noticed how different education environments and different really resources can shape one's outcome and opportunities.

So that's when I realized if I were to found a company in the future, I want to tackle problems in learning and problems in education. And I'm deeply interested in how we form a talent and then grow our talent as well. So that's why I joined RID in between my undergrad where COVID hit and I didn't really realized much value of attending university during very much online situation.

So I decided it took two years of gap year and actually went back to Korea and worked as a research scientist in a company called Grid, which was back then, even before Tragedy PD came out, was working on AI tutoring for English learners. So that's when I realized a lot of opportunities can be shaped using ai.

We were generating new questions, we are recommending what kind of concepts learning need to solve, and my domain expertise inside RID was a field called knowledge tracing. Knowledge tracing is a field where given a sequence of the learners trajectory, what kind of questions they got, correct, what kind of questions they got, right?

Essentially, the AI model will predict the proficiency of that student and try to recommend the best content that the student will learn from. So essentially that's when I realized AI really has a lot of potential impact in education. And coming back to Berkeley, I was more deeply interested into actually building a application that can be widely used by institutions and learners.

[00:40:34] Alex Sarlin: Fantastic. And yeah, that concept of knowledge tracing is deeply linked to many of the things we talk about in ed tech a lot of the time of adaptive learning, personalized learning. When you think about the recommendation engines, how to get the right trajectory for the right learner, that concept of knowledge tracing is really at the heart of it.

Tell us about how knowledge tracing and AI go together. Why does AI make knowledge tracing so much more powerful? 

[00:41:00] Yoon Yang: So when you think about what makes a great human tutor, a human tutor would remember the past interaction with the student when the student's solving a question, okay, this student is solving a incursion question, but I know that this student was iffy on basic function concepts.

So let me actually bring that back. And then before I give any hints on this recursion question, I can solidify the concepts on the functions which serve as the foundational concepts solve recursion questions. So the point of knowledge tracing is that the AI models can actually go back. 10 interactions back, 100th, interactions back all the way to the first time they actually met.

And utilizing all these long sequences and long contexts, you can actually model the students' interactions or model students' knowledge proficiency and propose the right way. So it's a bit like a human tutor, but in some sense it can actually go back much longer and provide more context about a student because.

If you're tutoring one-on-one, you may have a deeper relationship with a learner, but if you're teaching hundreds of students at the same time, like a university environment, you may not remember what UN was good at, and you have to basically kind of keep asking questions, which essentially is a shortcut of.

How AI would do knowledge tracing. 

[00:42:23] Alex Sarlin: That's really interesting. So piv, your new company, is using some of these techniques and some of these technologies to do grading among other things for higher education. Tell us about PIV does and how you're using some of these really interesting technologies to turbocharge and improve learning outcomes at universities as well as efficiency.

[00:42:43] Yoon Yang: So 

Pensive and an AI learning platform used by higher ed institutions. The main value of Pensive is primarily Twofolds Pensive health instructors to save grading time using their grading style so that they can focus more on student interactions. We call this switching screen time to student time. And for students, we provide a personalized course, specific AI tutor so that students can really understand the course, learn the course, using the professor's way of teaching, so that rather than getting general advice from generic language AI tutors, they'll be getting course specific advice and tutoring for that course.

[00:43:24] Alex Sarlin: I sense an interesting through line of what you're talking about for each of the things that Pence does, it is able to, you said, emulate the instructor's grading style. So it's learning from the instructor about how to give meaningful feedback to the students. It's learning from the students' background about how to tutor them in a way that serves their particular knowledge.

And it's course specific, right? It's designed with the course content in mind. Really interesting combination. Tell us about why there's so much value in understanding each of the pieces of, you're talking about the content, the teacher and the student. It's sort of the triangle of education. Tell us about how it works within each of those systems.

[00:44:01] Yoon Yang: Yeah. One thing that is really interesting about universities is that faculties are essentially top in their fields, and they have been teaching that course for multiple different decades. They are the experts. They are the course architects who has the most knowledge on how the curriculum should be ran and how should they teach and instruct students.

We believe our job is really amplifying the power of faculty and really bringing that faculty's empathy knowledge on the curriculum to individual students as if they're getting one-on-one attention. That's why we think injecting these context around the course is super important for students and learning perspective.

When you are approaching grading, you may think that. Grading is a very objective thing. If you give a student's answer to a hundred different faculties, they're all great, actually, very differently. They have their own preferences. They have their own learning objectives. So every single faculty has their own philosophy and grade in their different ways.

So our job at Pens is not to really give a uniform grade to students, or rather learn from the faculty. How should we construct rubrics? And then the faculty will be using our AI to generate the rubrics. Actually, let's test this rubric for a few samples. Oh, it seems like this rubric is not catching these mistakes.

Let's revise this rubric set, and then the AI would learn from that rubric, learn from that sample, and calibrate itself. To really learn from the faculty's grading to grade as if the faculty's grading, the entire grading set. And that said, faculty still review every single submission intensive grades. It facilitates grading by providing a clean transcription, providing a summary of the students' work and AI giving feedback on the detail.

And I think it's the same applies for the student side where, because every course materials are very intentional, there's a lot of the time where chat, GPT or generic AI tutors would shortcut student learning, either directly jumping in by giving the answer or. Pulling the student in a different, very generic way of solving this question, which is not intended by that faculty.

So when we build an AI tutor, we make sure that the tutor is very aligned with the curriculum, the material. It actually ingest the student's previous performance in the platform as well, because Penns is and grading platform. The tutor knows what kind of mistakes that student made in the previous midterm and what points the student earned from the past homeworks, which is super valuable in terms of knowledge tracing standpoint to help the students out for the future exams.

[00:46:41] Alex Sarlin: Yeah, it's a very powerful vision, and you're coming off a closed beta that pretty significant population size. You've been working across over a hundred institutions. Penn has already graded over 3 million questions. Tell us about that closed beta process and what you learned during that process, and maybe what surprised you about this as you started seeing ai, especially in relationship to how instructors react to AI that tries to learn from them and emulate their grading style.

[00:47:10] Yoon Yang: Right. We definitely saw a shift in perspective gradually when we started pens two years ago, and even back then, 2024. AI came a long way in two years. It just changed everything and we could see the general perception of faculties changing as well. One of the biggest surprise was actually how supportive the faculty was and how hands-on they want to be involved in the product development process.

So it's not a high to say that Pensive was built together with the instructors and faculties where they would give us a lot of feedback. They would try out different things, even though the platform was constrained to provide certain way of grading. They was trying different things and giving us feedback.

So really, if I were to pick one thing that surprised me the most is how. Enthusiastic faculties are on incorporating AI and really incorporating their workflow. That said, faculties always say, and even the ones who says, I hate grading, I never became a faculty to grade. They still really want to be hands-on because they want to make sure that every single student is touched by their knowledge and by their instruction.

So working with those faculties has been really, really a fun journey. These are faculties from some of the top of the minds in the entire world on their fields. So it has been a fun ride. 

[00:48:35] Alex Sarlin: Amazing. Yeah. It is a positive surprise about the enthusiasm you saw, partially because when you look at some of the polls on ai, sometimes educators and higher education faculty get concerned about AI and they say, oh, it can do all these things.

What is my role vis-a-vis the AI in the future? Is this trying to replace part of my role? And that is so antithetical to how you think about things at piv. You talk about amplifying faculty's impact, amplifying their knowledge. Tell us a little bit about what that looks like, because I think that's a perception that we've seen, I think, grow in general over the last few years, even if it's not really what ed tech industry is intending.

[00:49:15] Yoon Yang: Right. So when we see how faculties are actually using their time and their TA's time when they're teaching these large classes in universities, I. We're actually recently seeing more and more TAs bombarded with grading work just because mainly caused by the budget cuts by federal funding cuts. The state funding cuts are really driving faculty and TAs to really focus 80%, 90% of their entire teaching time on grading.

And I really think that if we could provide more face-to-face time with students, which I think is very uniquely human, we get much more opportunity to align students' learning objective and give them more empathy on how they should learn in their learning journey. So. By amplifying faculties, we mean when faculties use pensive, they direct our rubric generation AI to generate the rubrics based on their style.

When the AI generates the initial works, they would keep iterating by grading a few samples, tested again, few samples tested again, until they see there's a right rubric that they can actually grade. Now, with that, they would ask the AI grader to grade the entire students for the first pass, and then the Pensive system would let the faculties know which are the highly confident ones, so that they can just quickly spot check the submissions, which are the low confidence ones.

They would have to spend more time. So overall, this process. Every single step faculties are engaging with the AI on how the AI should run. At the end of the process, faculties end up spending on average three times to five times more time compared to using other normal grade green software platforms or grading by hand, which are really free up their time on opening more office hours, opening more tutoring sessions.

One really positive anecdote that we have is one of the uc, Berkeley's largest course, 1000 student data science course, was able to start a group tutoring session hosted by the TAs because is freeing up those grading load on their plates. So now every TA gets to teach a group of students, which is a win-win on both sides for student.

Even though they're enrolled in a massive 1000 student course, they're actually having a small group tutoring session now with a TA on the TAs. They never sign up for, um, to a TA job to grade a ton of papers. They actually get to learn how to teach better, how to interact with the students. So I think those are the positive benefits of really amplifying the faculties.

And I really resonate with this because when I was attending Berkeley, I was a tutor teaching computer science. So I was in this exact same position of teaching a small group of students trying to understand how, where they're struggling. And oftentimes we would unlock a very hard concept to understand in 10 to 20 minutes of just one-to-one FaceTime, where the student has been struggling for multiple hours because they couldn't really get a human support.

[00:52:25] Alex Sarlin: I think connecting the budget cuts in this moment to the idea that TAs and faculty are getting more and more of a workload from grading is a really, it makes a lot of sense. I had not thought of it quite that way and makes a lot of sense and Yes. You know, you mentioned a couple times very few people went into teaching or being a TA just to do the hand grading of many, many identical papers, so I, I can definitely imagine that there's a huge benefit there and I love your example of, you know, the group tutoring that it becomes enabled by the time that is saved from the individual grading.

That's exactly the kind of. Effect we'd like to see. So as pensive and AI in general continues to sort of grow its capabilities, you talked about grading, you're working on, you know, a really powerful, very personalized tutoring system. How do you see the role of AI evolving at, at large, within higher education over the next few years?

You know, it's obviously completely inserted itself into the higher education ecosystem in so many different ways from students, from faculty, from administrations, just in three years so far, in the next three to five years. What kinds of things do, should we expect to see? 

[00:53:31] Yoon Yang: I think on the assessment side, AI is really enabling the faculty to think about very creative, original assignments that are much more aligned with the concepts they want to teach.

When we do discovery calls, a lot of faculty say, ironically, we're going back to Blue Books, right? Even in the era of this digital learning, because so many students are cheating with AI delegating their thinking and for them. Pensive is a tool that allows a different way to assess students. Some faculties offer an oral exam, some with an ai, some faculties.

They used to give all MCT questions, multiple choice questions to students, but now because they are free, they have more resource to grade with P, they actually open back or open up, uh, free response questions and long form questions, which are much more pedagogical and test students understanding. So we see more and more projects based learning and different types of assessment rising due to ai.

They basically think in the next three to five years, there'll be different way of assessing students' knowledge, scale, giving feedback, and so forth. Another angle I'm seeing is as quickly AI is helping instruction itself. It's also disrupting what to instruct to students, which is very interesting because one of many purpose of higher ed is to train the future of workforce.

Now, there are a lot of doomsday articles on, there will be no juniors being employed in the future. New grads and fresh grads are very frustrated on what kind of jobs they could have get. My perspective is these roles are shifting and transforming work is something that we've drive value tremendously that I don't see a generation of not working, but rather those work will be transformed to be more meaningful and more valuable.

And I think one of the biggest challenge in higher ed right now is how can we actually equip learners ready in this AI native workforce? So not just using AI to. Transform what is working to be more efficient? I think gradually it'll be more enabling faculty to think about how should we teach students to use AI really well?

To think with ai, one of the things we have to be cautious on is always students aren't yet fully formed on their critical thinking abilities. So how can we prevent them delegating their thinking to ai, hence the name of pensive. We want to make students to be more pensive, more thoughtful, more introspective.

So I think that's just general direction will be heading. 

[00:56:17] Alex Sarlin: So we've had this really successful closed beta. I think this idea of personalizing learning at the instructor level, at the student level, at the curriculum grounding level, incredibly powerful. What is next for pensive? Have you gotten any funding to grow the company?

[00:56:32] Yoon Yang: Yeah, actually, uh, we just announced our seed round. We raised $6.8 million funding from great investors who are mission aligned, including Mayfield Reach, capital Scouts from Sequoia and recent or it's, so it's great to have them on board to our mission in the journey. We'll be using those funding to really deepening our current pension product on helping instructors more, improving our grading platform, improving our tutoring platform, but also serving other types of institutions as well, such as helping K 12 teachers to grade better on their essays and STEM questions.

Right now we are focusing on STEM fields, but we do want to expand on different types of assessments and different fields of study, and also helping the learners. One of the biggest pain points we're noticing is how is the content I'm learning right now connected to this AI native world and the world of ai?

How can I actually be surviving and succeeding in the workforce? So also helping students more on letting them know and really tutoring and mentoring them on how can they actually, what should they actually learn? How should they approach these concepts and connect to the future opportunities? 

[00:57:46] Alex Sarlin: Amazing.

You're touching on so many of the top issues of the day, right? What does career readiness look like in the age of ai? What is the future of assessment? How do higher education and K 12 work together to help students prepare for this very unknown future? It seems like a very bright future for pensive, and I'm really excited.

You mentioned the name, and I have to ask, I'm sure you've gotten this question before. Was there any Harry Potter reference in the name that people who have read those books, of course, remember the Pene, this sort of a play on the word pensive, where you can see the future or or journey through. Was that in your mind in any way as you named the company?

[00:58:23] Yoon Yang: That was one of the factors. I'm a Harry Potter fan. The biggest factor was delay to thoughtfulness and being introspective, but at the same time. As you mentioned, I couldn't, I couldn't keep thinking about, 'cause in a way how Penn c with the first spelling is depicted on the novel, connects to our mission as well.

'cause we are essentially collecting how students are learning and really helping them introspect and think about their learning path as well. So it was a good catch. 

[00:58:52] Alex Sarlin: Yeah. You, you take memories out of their head and use them for the future. Brilliant. 

[00:58:56] Yoon Yang: Right, 

[00:58:57] Alex Sarlin: because it's a great metaphor. Well, thank you so much.

Yoon Yang is the CEO and co-founder of Pensive. It applies AI to education. He was one of South Korea's youngest AI research scientists when he was at the AI tutoring company rid, and they are just celebrating a $6.8 million seed round. Congratulations and really looking forward to seeing what you do next.

[00:59:17] Yoon Yang: Thanks, Alex. Thanks for having me. 

[00:59:19] Alex Sarlin: Thanks for being here on EdTech Insiders. Thanks for listening to this episode of EdTech Insiders. If you like the podcast, remember to rate it and share it with others in the EdTech community. For those who want even more, EdTech Insider, subscribe to the free EdTech Insiders Newsletter on substack.

This season of EdTech Insiders is brought to you by Cooley. LLP Cooley is the go-to law firm for education and EdTech innovators offering industry informed council across the pre-K to gray spectrum with a multidisciplinary approach and a powerful EdTech ecosystem, Cooley helps shape the future of education.

This season of EdTech Insiders is brought to you by starbridge. Every year, K 12 districts and higher ed institutions spend over half a trillion dollars, but most sales teams miss the signals. Starbridge tracks early signs, like board minutes, budget drafts, and strategic plans, and then helps you turn them into personalized outreach, fast, win the deal before it hits the RFP stage.

That's how top Ed tech teams stay ahead.