Tim Hwang: What if we created a system—they call it “Aeneas”—to scan the historical record and surface interesting parallels for people to look into?

Volkmar Uhlig: If the cost goes down, I think we will see an explosion of knowledge.

Kush Varshney: It wasn’t trying to do the same job that a human would have been doing.

Kaoutar El Maghraoui: Some of the things the study showed were physically impossible to do.

Tim Hwang: All that and more on today’s Mixture of Experts.

I’m Tim Hwang, and welcome to Mixture of Experts. Each week, MOE brings together a group of brilliant minds to explain, distill, and “hot-take” our way through the truly bewildering wave of news in Artificial Intelligence.

Today, I’m joined by a stellar veteran crew. We’ve got Kush Varshney, IBM Fellow in AI governance; Volkmar Uhlig, VP and AI Infrastructure Portfolio Lead; and Kaoutar El Maghraoui, Principal Research Scientist and Manager for the Hybrid Cloud Platform.

We have a packed episode today. We’re going to talk about agents, of course. We’re going to talk about using AI for ancient history. And we have a special segment with Suja Viswesan on the latest Cost of a Data Breach Report.

But first, I want to talk about Study Mode.

Alright, just to quickly introduce this topic: ChatGPT announced a new feature just this past week called “Study Mode.” Basically, it’s a feature where you click it, and it becomes an interactive learning experience. It asks questions, it challenges you. I wanted to bring it up because, listeners of the show will recall, a few months ago we covered a report from the MIT Media Lab that scanned people’s brains. It kind of took over my social media because it was all about how ChatGPT is making us stupid, right? Like, if you literally use AI to write essays, your brain is firing less.

I think this is a really funny and interesting story, in part because this is an explicit attempt by ChatGPT to not have that happen. I first want to do a quick round-the-horn question, as usual. That question is: I’m curious about your favorite study method. Do you use flashcards, outlines, practice tests? Maybe, Kush, I’ll start with you. Do you have a preferred study method?

Kush Varshney: Yeah, I just read the book. I think that’s the easiest.

Tim Hwang: That’s great. Kaoutar, what’s your method?

Kaoutar El Maghraoui: Oh, I usually use active recall. You know, I read, try to close the book, and repeat things to myself.

Tim Hwang: That’s a good one. I usually use that one as well. And Volkmar, how about you?

Volkmar Uhlig: I’m on your side. I’m just reading the book. Luckily, I don’t need to remember things anymore that are relevant.

Tim Hwang: Yeah, exactly. I had to think back for this question. I was like, “Standardized tests—how did I go about doing that?” Yeah, I think active recall was what I tried to do.

Alright, well, I want to talk a little bit about this idea because I think it’s so interesting. People are always like, “Oh, AI is making us dumb,” or “AI is making us smart.” But I think so much of this is about how you actually design the systems. So I’m curious: have you had a chance to play around a little bit with Study Mode?

Kush Varshney: Not yet. But, yeah, actually, this isn’t exactly new, I would say. Claude had their sort of learning mode come out in April, and other people have been playing around with this. So, yeah, I mean, I think it’s a good thing to have this other kind of role that the LLM is taking. The content generator isn’t the only thing; an assistant isn’t the only thing. You can think of it as an editor, a devil’s advocate—all sorts of different roles an LLM is going to want to take. And it’s not going to happen naturally because of all the early fine-tuning that’s been done on these things. So, yeah, having these additional roles is a very good thing.

Tim Hwang: I’m curious. I was talking to a friend about all this, and he was basically a cynic about this whole feature. He was like, “Look, people just want the answer. No one’s really going to use Study Mode. This is just a marketing thing they’re using to push back against the narrative that AIs are making us dumb.” Do you buy that argument? Volkmar, I don’t know if you’re a cynic like he is.

Volkmar Uhlig: No, I have two kids, so... Yeah, I think it’s actually great because people learn differently. Some people just need YouTube videos; they don’t want frontal teaching. Other people want to be quizzed. And so I think it’s just adding to the repertoire. It’s almost a tutor, someone who watches you, and then if you make mistakes, the system would actually adapt to your behavior. So it’s almost like Khan Academy pushed into an LLM. I think it’s a great extension to the portfolio. If you give kids the answer, they will not learn. And I think that quizzing method is actually pretty good.

Tim Hwang: Yeah, for sure. You mentioned Khan Academy—that’s where I want to go with this discussion. We can debate the future and how effective it is, but there’s been a lot of discussion, at least among my circles, about how far all of this AI stuff goes in terms of education. I feel like these types of features point the way to the idea of, well, how much learning is just eventually going to be completely doable through AI? Which raises really big questions about traditional schooling.

I know you hesitate to do grand forecasts, but where do you think this stuff goes in a few years? Do we feel like we’re going to eventually have technology that’s competitive with what you might get from a traditional education?

Kaoutar El Maghraoui: Yeah, that’s a very interesting question. It seems to me that already, right now, many teachers and students are using AI. So I think, whether we want it or not... This new mode they have, I feel it’s similar to a cognitive gym. This Study Mode is a step towards a design philosophy that could be like a gym versus a crutch. Because right now, typically how students or many people use LLMs is as a cognitive crutch—it does the work for you to make your life easier. But the idea of a cognitive gym is basically designed to make you do the work, but with some expert guidance and support, to make you stronger in the process.

So I feel like the future of truly valuable AI in education and professional development isn’t about providing answers faster, but about building systems that are expert Socratic partners. Systems that know when to give you a hint, when to ask a probing question, and when to force you to struggle a bit. If this is done right, it’s going to be a huge way to redesign the whole educational system. But this whole idea of the tutor is a much harder design challenge than just building a better answer engine. How do we shift the metrics from “time to answer” to “depth of user understanding”? How do you evaluate these systems? It’s going to be interesting to watch.

Tim Hwang: Yeah, that’s right. I think it’s worth getting into the ethics of this and how you go about doing it. From what OpenAI said in their blog post, they consulted with expert educators and have some magic working in the background to make Study Mode work. Building on Volkmar’s point, Kush, I don’t know if you have an opinion on this. Ideally, you would eventually want an AI to be able to go multimodal for whatever teaching method it thinks is going to be the most effective. Right? I don’t know if you think we can get there.

Kush Varshney: Yeah, I mean, I think the biggest thing—what Volkmar brought up, what Kaoutar is bringing up—is just having a bit more control over this. You don’t want the AI to have all the agency. There’s a finite amount... the AI can have some agency, the human can have some agency, and you just need to balance it. It’s not just doing everything for you; it’s not taking control of the entire situation.

The next topic we’ll get to is kind of along the same lines: how do we have these things adapt to us, making sure it’s doing things in ways that make sense for us? I just wanted to bring up one article that actually came out yesterday. It’s in a maybe lesser-known news outlet called Rest of World, about technology in the rest of the world. It talks about how teaching is happening in rural Colombia, making the point that, yeah, OpenAI or Anthropic are doing these things, but those are closed, proprietary systems. In these schools in rural Colombia, people are using AI integrated with WhatsApp, and the level of reading has gone down in a whole year... all these sorts of things. It’s great that we’re seeing this, but how are we going to bring it to a broader population? Because just keeping it enclosed within ChatGPT is not going to be the thing for most of us.

Tim Hwang: Yeah, I think it’s a really hard problem. Before we move to the next topic, I don’t know if you want to offer some parenting advice. I’ve got a few kids at home, and we’ve been traditionally quite cynical about any screen time whatsoever. Right? So we’ve been very much, “Keep them away from the screens.” But they’re kind of reaching an age where I’m like, “Should I give them access to chatbots?” I don’t know how you approach that with your kids. For our listeners, would you recommend it?

Volkmar Uhlig: Yeah, so my kids are a bit older; they’re teenagers. And, you know, the screen train left the station a couple of years back—a long, long time ago. So what I’m seeing in the education system is that, if you look at a traditional public school system, there is effectively a fight against AI. This is primarily because, I think, the public education system is not willing to adjust the curriculum.

We are looking across the board, and we’re also looking at schools which have completely integrated AI into the curriculum. So, I’m in Austin; there are a couple of schools here which said, “You do studying for two hours solely AI-based, no teacher, and then everything else is project work.” Kids are doing crazy stuff. There’s a kid who built a BMX bike park in the city of Austin—needs to raise funding, build it, etc. So I think we are at the point where we finally have the opportunity to explore new education methods. That’s the primary thing.

For hundreds of years, there have been teachers standing in front and a bunch of kids needing to listen. Suddenly we can actually adjust it and adjust it to the learning speed of the kids. My kids are really bored in school; they’re kind of sitting and waiting until the class catches up. With AI, we suddenly have the ability to either go deeper or to just let them sprint ahead. I think, at the edges of the bell curve, you will just get so much more happiness.

Kaoutar El Maghraoui: I think, if I might add to this, Tim, I also have kids that range from teenagers to a nine-year-old, a 16-year-old, and one in college. So, um, it’s important, I think, to introduce the right tools at the right age. For maybe kids 1 to 6, kid-friendly tools like Khan Academy Kids or something like that. But in middle school and high school, things like Quizlet AI or an AI tutor, where it’s more interactive, not just passive learning. You promote active learning where you teach the children how to interact with the AI, ask follow-up questions, learn how to prompt in a more intelligent manner. But also give them times when there is no screen time. So I think it should be a combination of no AI, so you force their cognitive capabilities to build, and also slowly introduce the AI because that’s going to be their world. They’re going to be using these things whether we want it or not. So they need to learn how to use it ethically, in safe ways, and in an interactive, not a passive, manner.

Volkmar Uhlig: This goes along with what came out yesterday: Meta is changing the interviewing process. They actually ask people who are interviewing to actively use AI tools to solve the interview questions. Right? And so I think it shows the shift from, “Are you individually capable of solving the problem?” to “Are you able to use tools to solve the problem?” So I think we are outside of the realm where you don’t use AI; the expectation is that you do. We need to test whether you are correctly using the tools or not, or if you are able to solve the problem without the tool.

Tim Hwang: Yeah, for sure. There are all these interesting processes and systems in society that kind of pre-assume no AI, and it’s very interesting. Everybody’s either in “fight” or “adapt” mode, and I think eventually everybody will have to adapt. It’s kind of about how long they want to put this off and what pain they’re willing to go through. That’s really interesting; I hadn’t heard that Meta was doing that.

Any final thoughts on this topic before we move on?

Kush Varshney: We just need this to be more inclusive. I mean, the best and the brightest are going to have these things at their hands. But how do we make sure that everyone does? We don’t want this just to be a luxury.

Tim Hwang: Good. Yeah, absolutely.

Alright, I’m going to move us on to our next topic. This is a submission from Volkmar, so I’ll tee it up to you as the first commenter. An interesting tweet came out from a gentleman by the name of Greg Eisenberg. He’s an entrepreneur who runs a few companies. It’s a nice little Twitter thread (now on X), and he starts by saying there’s a quiet shift happening in how we design software: we’re moving from UX to AX—“Agentic Experience.”

Originally, I was like, “Another agent thing, I don’t want to look at it.” But the more I read, the more interesting it became. I’ll just quickly sum it up. I think his argument was: back in the old days, we would design applications and interfaces with the idea that the flow of that interface was fixed, right? Essentially, the interface was dumb. You’d start from zero, and then you’d maybe do a little bit of customization, but a lot of it would be about guiding the user through a very uniform experience.

His argument is that, with agents now, we’re living in a world where these interfaces can become a lot more intelligent. They can retain a lot more information about how we interact with them, what we’ve done on that site. So his vision is a future where these interfaces become a lot more malleable, a lot more adaptable, really changing the way we’ve traditionally done UX, which is to think about these fixed flows.

I guess you submitted this one. I don’t know if you want to talk a little bit about why you thought this was interesting and where you think our listeners should focus here.

Volkmar Uhlig: Yeah, I think the core change here is to rethink how we are interfacing with the machine. As you just pointed out, this was the work of a designer who had to figure out the most common flows. I think AI is really interesting in two ways.

One is: usually you implement 90% of the flows, and 10% is exception handling. If you look from an enterprise perspective, the 10% is your labor cost because you have all these people sitting around dealing with cases which haven’t been implemented. Now you can suddenly go and say, “Well, I’m fundamentally changing the interface. It’s dialog-based.” Or even if it is an application where you interact with the screen, the screen elements could be generated directly by AI. You have an adaptive user interface because the AI thinks ahead—what’s the next step?—and it gives you the choices. Maybe you don’t want to type; maybe you have a bunch of icons you click on.

The other thing is: if the user interface goes away from pre-programmed flows, the speed of new user interfaces will be incredible. Now you just describe what your fundamental business problem is, and the AI can fill in the gap. On the flip side, the implication is: how do you prompt? Because the AI could get lost in the weeds. So I think it will be a new skill set: how do I put guardrails around the model and the possible questions and answers it gives?

What I can’t wait for is that I don’t need to re-enter my name, my credit card, and my billing address because there’s absolutely no reason I need to do this every single time. The system can incorporate information already collected about us. It’s kind of like when you go to the bakery on the corner and they already know what you want. I think that’s something we can now create in the digital realm. The user interface will become much more personal, which is nice—it’s adapted to you and not to the mass.

Tim Hwang: Yeah, this world starts to look very interesting and different. I was reminded recently: my wife was in the other room and said, “Could you pick up my phone? I need to text a friend.” I picked it up and was like, “This home screen makes absolutely no sense to me.” All the apps are in different places, lots of different configurations.

I think one really interesting world is that, in the future, you might sit down at someone else’s computer and just realize, “Wait, what application am I even using here?” because it would be so customized to their use.

I don’t know if you have responses to this. Are you excited about this world of customization, or do you have, at least in my case, a little bit of hesitance that it might actually be really baffling because everything’s going to be very customized in a way that at least I didn’t grow up with?

Kaoutar El Maghraoui: Yeah, I think I have mixed feelings here. Part of me is like, this is going to be exciting—this super customization that’s going to make life easier. I talk to these agents, and they know me so well, so I don’t have to explain myself every time, re-enter all my history. But then, the flip side is: when they know too much, what’s the implication? What are the security implications? Like you’re saying, when we interact with others or try to look at other people’s phones or experiences, we’re going to be lost. Your world is going to be super customized, so you’re going to be lost in other worlds.

But I see this as a big shift we’re noticing. This move from UX to AX is one of the profound shifts in human-computer interaction since, I think, the graphical user interface. It seems to me we’re moving from a world where we are operators of tools to where we are managers of agents. It’s not about having menus and buttons; it’s more about having conversations.

For decades, we had this invisible UI that was so intuitive the user didn’t have to think. But in the agentic world, I think it’s the opposite. The agent’s reasoning must be transparent. If I’m asking an agent to spend thousands of dollars of my money, I don’t want magic here. I really want a clear plan and the ability to approve it, and trust. So the core design challenge here, I think, is moving from “how do I arrange pixels on a screen” to “how do I design a relationship between the human and the agent?” A relationship built on trust, which requires several things: competence (does it do the job well?), transparency (can I understand what it’s doing?), and control (can I intervene and correct when things are going wrong?). That is a big shift we’re noticing right now.

Kush Varshney: Kaoutar summarized it quite well. This is a new paradigm of interaction. Something my team has been pursuing recently is this concept of mutual theory of mind. Pretty much what we’ve already heard: when two people are interacting, if I know what the other person is thinking and they know what I’m thinking, we can work better together. This goes to second and third orders. Some of you might have seen the movie The Princess Bride a long time ago; there’s a whole scene about thinking of what the other person is thinking, back and forth. This is exactly a way to make things more productive in a relational way.

I think there are going to be all these things from human relationships that come into it. It’s not just going to be the AI adjusting to us, but us adjusting to the AI. We’re going to be introducing ourselves to AIs; AIs are going to be introducing themselves to us. There’s going to be thinking about, “What’s the level of conversation I should be having? Are there particular information processing styles I should be appealing to?” in both directions. I think that’s going to be such a nice change for us. It allows so much more tinkerability; we can make these things authentic for what we need. But, yeah, the risks are there, as Kaoutar talked about, as Volkmar talked about. So, yeah, just balancing all of that.

Tim Hwang: Yeah, for sure. There’s one hypothesis I had thinking through this post that I thought was fun. We can almost draw an analogy to the rhetoric of the 2000s: “Oh, we’re going to live in a world where everybody can have a blog; there’s going to be blogs on every possible topic.” At the time it was called the Long Tail. “The most popular shows will become less popular, and then there’ll be lots of micro-shows everywhere.”

One reflection from that era was, well, that didn’t actually quite happen. On YouTube, there’s still a small group that gets a huge amount of attention because it turns out people have some shared interests, and there is usually breakout behavior.

I’m curious if you think that’s maybe one outcome for all this: you set up all this gigantic tooling for your interfaces, allow them to drift and customize every way you want, but it turns out humans are very similar, so you actually end up with a lot of interfaces that are really quite similar for the vast majority of users, and then a very long tail of truly bizarre interfaces. Do you think that’s maybe one outcome we end up with?

Volkmar Uhlig: I think, if you look at the web, right? The web had tons of different versions, and now it’s all standardized; it all looks the same. It’s primarily so you can go from one webpage to another, and your cognitive load is low. I think we will have some that become more dominant, and then everybody will just follow. But I think this is exactly where we need to go through an experimentation phase, which is why I’m really excited about this. Finally, after 40 years, it’s not a mouse and an input field; we can actually rethink a user experience.

The last major shift was the iPhone—the input device changed from a mouse to a finger, but otherwise not much has happened. This morning I drove in, and Tesla now has voice in the car. I use it quite often. I used to use it on my phone; you just have conversations. I’m like, “Okay, I found this paper, can you summarize it for me?” So I’m having a dialog with the car now, which is kind of silly, but it’s 25 minutes of driving, so I can actually do stuff while driving.

I think we will get into user interfaces where there’s an expectation of how a flow looks if you book an airline ticket or buy something, but you may branch off. It knows stuff about you, so it will cut out steps you don’t want. A new way of interfacing with the machine will emerge. We’re currently in the experimentation phase, but it’s nice that there’s actually something new.

Tim Hwang: Yeah, I agree. It’s a new breath of fresh air. I met someone recently who allegedly influenced the “pull to refresh” interface on your phone. She was like, “Yep, that was me.” I was like, “That’s really crazy that someone had to come up with that.” It’s interesting to think we’re in a very similar place for AI now, where all those tropes or design patterns need to be built out.

Kaoutar El Maghraoui: Yeah. I think what also excites me a lot is neuromorphic interfaces, where your brain is actually interfacing, or your eye—not just your voice. Other things will be used as part of brain-computer interfaces or prosthetics and biofeedback devices. There’s a whole world of new things that will emerge, some of which might be very useful, especially for people with disabilities. I hope that’s going to open up a lot of things they can’t do today. So that is really exciting.

Kush Varshney: Yeah, I was just reading a paper yesterday from some folks from Korea. They’re thinking about what the next iteration of interaction research is, looking at expanding. We’ve kind of gotten to a point where we have a definition of human-centered AI. The next thing is maybe group-centered AI because you can rely on cognitive psychology at the individual level, but now you’re going to have these things in human-AI mixed teams. It’s a social psychology question now: what are the new considerations? How do you make sure everyone’s voice is heard? All these things are going to be part of it. So I think that’s another angle for exciting research to come.

Tim Hwang: Yeah, for sure. I was just thinking when Volkmar was talking earlier: there’s a funny world where everybody has their own different interface in the future because we can do that now. But agents still need to talk to agents, so maybe MCP (Machine Control Program) will be the standard. The remaining standardized part of the web will be only what agents can see, and everything else will be very customized when it’s getting to a human-readable state. The complexity of that will be very interesting to navigate.

Alright, I’m going to move us on to our last topic. A super fun paper came out. If you’re like me and enjoy reading papers in Nature, this is one for you. It’s called “Contextualizing Ancient Texts with Generative Neural Networks.”

It’s a fun paper which basically says: look, a lot of the work of historians, particularly when they analyze ancient texts, is they look for what’s known as parallels—texts that have shared phrasing, shared function, or shared cultural settings. The idea is you do research by putting these next to one another and trying to make inferences. “This was written in this place at this time, and this was written in that place at that time, but they share these commonalities. Maybe they had a trade relationship; their languages are similar.”

This group of researchers said, “Well, pattern matching for parallels is something generative AI seems to do really, really well. What if we created a system they call Aeneas to scan the historical record and surface interesting parallels for people to look into?”

The results are pretty interesting. Out of the candidate parallels identified, historians found them to be useful research starting points in 90% of the cases.

I think this is a really fun story. Kaoutar, I’ll throw it to you. Normally we talk about AI in the commercial space—what’s happening in B2B, B2C. We also say AI is really going to be good for research, usually meaning pharma or math. But this is an application of generative AI to a pretty different domain, and the results seem impressive.

Kaoutar El Maghraoui: Yeah, I really love this direction. I feel like we’re in a world saturated with conversations about AI for profit, productivity, or power. This is a story about AI for posterity. AI at its absolute best is not just replacing human labor, but recovering lost human thoughts. This project is a powerful counter-narrative to all these fears around AI. It really shows that AI can be a tool for connection, not just optimization. It connects us to the voices of philosophers from thousands of years ago—voices we thought were permanently silenced or lost.

I think it also provides a blueprint for the future of the humanities. The combination of advanced imaging and pattern recognition—AI is creating a new field of digital archaeology where we can apply the same techniques to faded manuscripts, damaged artworks, etc.

But one key question is: how do we ensure these incredibly powerful, often expensive, AI tools are made accessible to researchers in the humanities? They’re often less-funded departments. This could unlock a lot of historical or cultural mysteries. That could be the next challenge for AI.

Tim Hwang: One thing I was thinking about, building on Kaoutar’s point about accessibility, is the technology as a whole. Imagine the landscape of all possible research problems. We’ve often assumed AI will accelerate the hard sciences first—material science, finding new proteins. But it may be fundamentally lumpy. I imagine a world where this “parallels” task is something AI has been able to do for some time. There’s almost a world where, even before we get super-accelerated science, we might have an explosion in historical knowledge first. AI is going to have these weird effects on which parts of our knowledge move quicker than others. Those outcomes are very interesting to me.

Volkmar Uhlig: I agree. I think we are touching or scraping right now on AGI in the end. If you look at this, you can let this thing run for a while, dig around, form hypotheses, and then based on those, form further hypotheses. This is where machines... we have so many humans who want to do archaeology, but as Kaoutar said, there’s limited funding. Suddenly you can fund this with just energy. Right now, it’s just energy and a bunch of GPUs. We can accelerate our knowledge about historic things just by having a few people with a huge body of NVIDIA chips behind them. Suddenly you can get to new knowledge.

There’s always this chain of thought. You have something new invented, and then you know what follows next. Suddenly there’s the possibility to explore crazy ideas we haven’t been willing to fund because you need to apply for grants. It will open up a lot of understanding of the past, even in domains we traditionally don’t consider valuable. But if the cost goes down, I think we will see an explosion of knowledge.

Tim Hwang: Yeah, that’s super interesting. It sounds like you’re arguing there’s a cognitive dividend that applies to these fields. Usually, it’d be really hard to get grant funding, but in a world of open source or even less sophisticated models, you might get very far. There’s a benefit to all these fields that might not otherwise get funding to push research ahead.

Volkmar Uhlig: You can do the same thing. We’re seeing in code generation a 10-to-1 ratio. One good engineer using an LLM can produce the code of ten people. Why is that true in coding and not other disciplines? Coding is just a very profitable business; that’s where it gets applied first because there’s high competition for that skill set. Of course, it will be deployed in every skill set. We will get similar productivity gains. People building models are programmers, so they make their tools first, but it will roll into other disciplines.

Tim Hwang: Yeah, for sure. Of course, I’ll ask the obvious question. Some people listening will say, “All the archaeologists are out of a job. All the ancient historians are out of a job.” Do you buy that?

Kush Varshney: Yeah, I don’t think so. The fact that there was synergy shown in this paper is actually unique. In 95% of human-AI team studies, you don’t get any overall benefit from the combination. Here, because the task itself was contextualization, it wasn’t trying to do the same job the human would have been doing, but freeing them up to do other things as part of their workflow. That’s part of the story.

The other thing interesting about this is that Latin, which this is about, has a lot of resources. Human historians have been studying it for a long time, so there was a large corpus. But when you look at lower-resource ancient languages—a great example is the Indus Valley Civilization. Their script is still undeciphered because there’s such a paucity of content. Actually, earlier this year, there was a $1 million prize announced: “Can you use AI to decipher this script?” That’s going to be a completely different endeavor because it’s not using all the human accumulation we have; it’s to discover something completely new.

Tim Hwang: Yeah, for sure. It goes to an important subtlety: we have to think about where the bottlenecks are in the research. Here, a lot of it is just not having enough grad student bodies to throw at the dataset to find these parallels. So you’re unlocking opportunities by enhancing or substituting that labor.

Kaoutar El Maghraoui: But I think additionally, some of the things the study showed were physically impossible to do. For example, the scrolls were buried and carbonized into fragile, solid lumps. Physically unrolling them would destroy them. For centuries, the content was a mystery. But now, with technology involving 3D scanning, using high-resolution scans to create detailed 3D maps, and then using AI to train a computer vision model to detect subtle differences in texture corresponding to ancient ink, and doing virtual unrolling... These things were physically not possible for researchers. The AI here is unlocking things.

It’s not about replacing historians. It’s a massive collaboration between computer scientists, physicists, and experts in ancient texts. AI is augmenting human expertise, allowing scholars to do something physically impossible before. That is pretty powerful.

Tim Hwang: Yeah, that’s really cool. It makes me think about putting together an engineering core that goes from field to field, looking for these bottleneck areas. Once you start looking, they’re kind of everywhere. The traditional problem is how to drag engineers to work on these things, but it’s such a cool topic.

Alright, I’m going to close this up there. That’s all the time we have for today because we’re going to move on to our final segment with Suja on the Cost of a Data Breach Report. But as always, Kaoutar, Kush, Volkmar, thanks for joining us, and we’ll see you soon on Mixture of Experts.

Kaoutar, Kush, Volkmar: Thank you.

Tim Hwang: So I’m thrilled to have Suja Viswesan joining us today. She’s the Vice President of Security and Runtime Products, and she’s joining MoE for the very first time. Suja, welcome to the show.

Suja Viswesan: Thank you. I’m really excited.

Tim Hwang: Yeah, definitely. We wanted you on because we covered the Cost of a Data Breach Report in 2024, and I understand you were deeply involved in helping the 2025 report get together. We’d love to dive into the data.

I wanted to start with one number that stuck out: 97% of organizations reported they had either experienced an AI-related breach or lacked proper AI access controls. This is wild because some of this AI-related data is among the most valuable things companies own. I’d love to hear more about that number. Are you shocked by it? It certainly stood out to me.

Suja Viswesan: It is shocking, and also kind of expected because the challenges organizations typically have are multifold increased because of AI. It’s like one of my colleagues used to say: it’s like COVID. You wanted vaccines to help, but you still had to wash your hands. The basic hygiene that’s not there gets exposed and exploited in this AI era. That’s why I said it’s not surprising, but at the same time, it is surprising that the other number is only 63% of organizations don’t have enough AI governance policies to address these problems.

Tim Hwang: That’s right. Is that the right way to think about it—companies already don’t get the basics right, and the age of AI is accelerating those issues?

Suja Viswesan: Exactly. A problem that took six months or even 2-3 months to manifest can now, because of AI in bad hands, expose all these problems much faster.

Tim Hwang: Yeah. A few years ago, I was saying, “Pretty soon you’ll have AI-enhanced attackers, and the space will be more dangerous.” Can you give our listeners a flavor of the kinds of attacks we’re starting to see? When people hear “AI-enhanced attacks,” they have a fuzzy vision. In practice, what are we talking about? Things like phishing?

Suja Viswesan: I think one part is phishing, definitely. It used to take 14 days to craft a phishing message. We had initial controls—looking at language, grammar. All that is changing. Now it takes minutes or seconds to craft a very personalized, diligent phishing attack. It has upped the game for attackers. Defenders need to work on different protocols to determine this.

The second part is data challenges. You talked about the IP, the crown jewels of a company, exposed through AI. Without proper governance, they can be easily exploitable. So it’s a two-way prong. We also have AI on the defender side to help. If you look at the numbers, the cost of a data breach is going down because defense is upping their game, but offense is also playing that game. It’s like a race.

Tim Hwang: Yeah, that’s right. I wanted to ask about big trends that stood out. You mentioned one: AI defense is maturing, so breaches are less expensive. Anything else people should pay attention to?

Suja Viswesan: Extensive use of AI in security is giving a lot of cost savings. We saw in the data breach report about $1.9 million savings because they could use AI to protect. That’s a big positive side. Yes, it’s a scary world, but we’re seeing positive sides on how to adopt AI to defend and increase protection.

The other part is the “needle in a haystack” problem. With AI, you can elevate your security analyst’s life. They can spend time preventing attacks instead of the hard labor we can now offload to AI.

Tim Hwang: Yeah, that’s right. Are you ultimately optimistic? Next year, when we talk about the 2026 report, will the cost/impact of breaches continue to decline? Is defense becoming dominant, or are we bottoming out?

Suja Viswesan: I think the cost will go down, but the volume might go up.

Tim Hwang: More breaches, but less per breach?

Suja Viswesan: Yeah, because of existing hygiene issues. Your data security posture—for both structured and unstructured data, your identity and access, your secrets posture, encryption posture—all need to level up to reduce the attack surface.

Tim Hwang: I’m curious: are there differences between types of enterprises? Is data governance a bigger problem in some, while others lack AI-enhanced analysts?

Suja Viswesan: Critical infrastructure companies—financial, health—are highly regulated, so it’s important for them. It’s a one-step function. For those who weren’t highly regulated, they didn’t care before. Now they have to care because of ransomware, data breach; they lose trust with customers, which is hard to rebuild. That’s the difference between regulated and unregulated industries.

Tim Hwang: One last question. A lot of listeners might read this and say, “I don’t have an AI governance policy.” Do you have a recommendation for how to get started? What should they do tomorrow?

Suja Viswesan: The first thing is look at your data—easier said than done. For structured data, we have clear regulation and governance. The unstructured side is what a lot of AI is trained on to increase productivity. That’s where we need to look at lineage. Does your governance and access policy travel along with this unstructured data all the way? Looking at that helps prevent problems.

I would start with: Do I have lineage and clear accountability for unstructured data? When it gets sharded and put in a vector database, do I know where it came from and who can access it? That will inform the model on how it behaves. That’s where I would start.

Tim Hwang: That’s great. Well, Suja, thanks for coming on the show, and hopefully we’ll have you back next year.

Suja Viswesan: Thank you, Tim.

Tim Hwang: Thanks to all of you listeners for joining us. If you enjoyed what you heard, you can get us on Apple Podcasts, Spotify, and podcast platforms everywhere. We’ll see you next week on Mixture of Experts.