Now the first question, Kevin, is pretty much like — I'm in modern languages and literature, right? And a lot of us here work with languages on a daily basis here at the University of Miami, and we see language as the core interface of AI.
So you and I discussed previously, like especially when it comes to Hebrew as a minority language, and after the release of DeepSeek, because for some reason there was this kind of uneven performance that we see in LLMs across languages. Obviously, many models are trained predominantly on English language data. They still show this noticeable limitation when working with languages such as Hebrew, but also other minority languages, of course, and indigenous languages and low-resource languages. And that raises some sort of technical and cultural concerns. So how do you see the future of AI development addressing these kinds of asymmetries? And maybe more broadly, asking to what extent does an English-dominant training ecosystem shape the worldview embedded in these AI systems? And if you think that prioritizing non-English and low-resource languages is essential not only for technical accuracy but also for preventing a culturally homogenized, English-centered digital reality.
Yeah. So when we first started training, you know, GPT-1, GPT-2, GPT-3 — this precedes my time at OpenAI — but the main corpus that the models were trained on was the web. So you go and get as much of the open web as you possibly can, and you use that, you know, in a sort of standard AI pre-training setup, and the models learn to speak all kinds of different languages, right? And you don't teach them English, you don't teach them French — they learn to predict the next word in a variety of different languages. And frankly, to them, the languages, you know, they don't really know that they're speaking French or they're speaking Hebrew so much as they just, you know, they're pattern matching and so on.
That's where it starts. But then, as you said, the web is primarily English, and then as you look at the tail of other languages, it's, you know, you have a certain set of major languages that are well represented, and then a whole broad set of languages that aren't, and then dialects that aren't as well represented, and so on. And so the model, you know, since the model is good in some proportion to the log of the amount of information that it gets about a particular language. So then what we tried to do is go out and find all of the different, you know, datasets that we could that represent the languages that the model is not yet as good as we want it to be. And so, you know, you go and try and like speak with everybody from, you know, if you look at Indian dialects, for example, Indian languages, there's a huge volume of them, and many of them are not well represented on the web. So we go try and work with companies in India, with the government of India that has an interest in preserving these and has a lot of data about them, individuals — we, you know, we will pay individuals who are speakers of these languages to help translate or help record.
And then you also want to start getting the model — once the model is sort of passable at something, you can have the model generate a bunch of data in various languages. You know, have it translate existing documents that are in certain languages into the languages that you're trying to learn better. Maybe have a human, instead of generating them, edit them and correct them and make sure that it's, you know, a good document in this new language. And then suddenly you have a bunch more of this language, and you can train on that, and you can get better at it. So it's an iterative process, and it's an imperfect process, but over time, empirically, the models have gotten better at a huge variety of languages. They're not perfect, and they never will be, but we keep trying to make them as good as they possibly can be.
You know, both like linguistically and also culturally, they pick up a lot of cultural cues in the language itself because that's how they're expressed. And so, you know, you'll find that they can — that they hopefully pick up a lot of that as they learn the language, but it's an ongoing thing for us. We spend a lot of money and a lot of effort trying to make the models better at a huge variety of languages.
Excellent. And I have to admit that with the Judeo-Spanish language or dialect, there was a huge difference since last year. When I had, in spring 2025, a cohort working on text from the Geniza, for example, in Cairo, with the Ladino writings, even handwritten cards and letters, and now the cohort that I have this year, this spring 2026 — a huge difference between one another. So that is impressive.
It does seem that there's a lot of work being done there.
That's really good to hear.
I mean, it's one of the things that has been fascinating to learn. I've been at OpenAI almost two years now. It's a very empirical science. This is not, you know, a bunch of people writing equations on blackboards and, you know, with a perfect theory for how all of this stuff works. It's massively empirical. We're running experiments all the time. We're trying different things. You know, what if we teach it this way? What if we do this to the data? Maybe the model will learn faster. And then you have evaluations. You have effectively like you design tests for the model to evaluate how good it is at any particular thing. And you're looking at how you can like optimize the model's performance on the tests, and it's sort of endlessly inventive and experimental.
So it is not a straight line, but there are a lot of people who are, you know, putting in a lot of effort to try and get there. So I will pass that on — actually, it's really good to hear.
Excellent. Thank you. And for the other question, across our own university, but I'm sure many other universities, we're seeing definitely significant experimentation with AI, all kinds of initiatives that are going on here. Hey, I know, for example, my own chair, Logan Connor, the dean back here of the College of Arts and Science, and others — other labs, for example, the Lab School of Law, a good friend of mine, is putting in an extraordinary effort to integrate AI into the School of Law, and many, many others.
So I guess the question, maybe like with OpenAI — does OpenAI have ideas about the hardware helpful for education and language use? I know that there's an issue now between the software and the hardware and the different models that are evolving quite fast, sometimes a little bit too fast for us to grasp the differences. But is there any hardware or like investments towards language and humanities programs that are trying to make a difference?
How do you mean hardware in that case?
I mean like — okay, so there was this conspiracy after the announcement, the Super Bowl OpenAI advertisement. So the network started to talk about, for example, different kinds of — that OpenAI has different hardware that they want to launch, but it's not really certain if it is or it's not. Do we know like any kind of hardware that you can share with us about coming up to help us?
Yeah, I mean, some stuff is public. So this is not the kind of hardware that you were probably thinking, but we are working on our own GPU, which we have talked about publicly. We're working on robots in various forms because I think physical intelligence is going to be one of the next big waves, right? We've got now — AI can do incredible things sort of confined to your computer, but that plus vision models gives you a great foundation to make robotics that becomes massively helpful to all walks of life. So we're working on that. And then we bought Jony Ive's company that he was working on. Jony, who, you know, was sort of the principal designer of the iPhone and a whole bunch of other things. And so there is a set of projects there. Those we haven't talked about, so I shouldn't, you know, break silence here, but it should be very interesting.
And to your point, there are really interesting opportunities for like the AI to help solve language-related problems that people run into, you know, with even just translation. And you don't need necessarily hardware for this, but hardware is helpful if you have, you know, you're like in person with somebody. But there's no reason that we couldn't have basically the Star Trek translator at this point, where, you know, Shai speaks Hebrew and it comes out in English for me, and I speak English back, but he hears Hebrew. Like these things are all possible all of a sudden. And so like, that's a cool world. Like I went through South Korea and Japan and this whole Asia swing on a trip not long ago, and we had translators in a bunch of places, and then we would have places where we didn't for whatever reason, and I had a number of conversations with people with whom we had no shared language by putting my phone on the table and saying, 'Hey Chat, I want you to act as a translator between English and Korean. Whenever I say something in English, I want you to repeat it in Korean. Whenever you hear something in Korean, I want you to repeat it in English.' And I had a whole conversation with someone as a result that would have been impossible before. So, you know, I think it's really cool that we are seeing barriers broken down, and, you know, we're just getting started. Like, if AI is here today, just imagine where it will be in a year.
And that leads me into the other question, which is quite important, because as you know, I do teach languages. So if real-time translation becomes seamless, like why should anyone still learn another language? And do you think that AI will ultimately weaken language learning or actually revive it?
I, you know, I'm an optimist about this stuff. I think — look, chess is an interesting example. AI, like old classic AI, Deep Blue beat Garry Kasparov at chess back in like 2000-something, right? And back then it was close, and it was a surprise that it could beat Garry Kasparov. The best computer chess programs now are so dramatically better than the best human player. It's like by 800 ELO points or something, which basically says you will never win. Like they could play a thousand games, it will beat Magnus Carlsen a thousand times. So yeah, chess is solved in some sense, but the actual level of interest in chess — there are way more humans playing chess today than there ever have been. And the interest in chess is higher than ever.
I think there's something instructive about that. And I could see the same thing happening with languages, where yeah, at that point it almost frees you up. If you can speak any language you want and be intelligible to any other human on the planet, then like maybe everybody doesn't need to learn English, you know, like you don't need that lingua franca kind of common tongue in the same way, and maybe everyone goes back to like, you know, learning their more sort of cultural — I don't know, like you could see it. I think there are positive ways to look at this rather than just negative ones.
I completely agree. I have to admit, I mean, like as you said, if we kind of give the opportunity for so many people that usually do not have the possibility or the closeness to another language or to interact with another person from a different language, and suddenly they can, that will make them much, much more curious and interested in learning that language and learning the culture behind it. I completely agree, that's nice. And from that learning to transfer to the students — I promised my students that I will ask you one question that is the obvious question, because obviously students, you know, focus on trying to remain relevant when they graduate. So what are the non-obvious skills that a student should develop today in our classes if they want to work in an advanced AI environment?
Curiosity. Because I think no matter what happens and how the world changes, being a curious person who is interested to learn and, you know, can learn quickly will just be more important than ever. I think it's always been important. I, you know, I don't know about you, but not that many people — I am certainly not doing day-to-day what I studied in school. So, but I don't think that my major and my graduate work and all of that was wasted. It taught me to learn about the world and taught me how to learn, and I use that skill every single day, and that's just going to be more important. I think the other is being high agency.
Going from like having an idea to 'I just did this thing.' Because one of the most magical things about AI is how it enables everybody to create whatever they can think of. You no longer have to know how to code to create a website or a product. You can just use Codex and give it a prompt, and all of a sudden you have something that's working. You know, like we have our finance team writing code every single day. They've never written code in their life, but suddenly they're solving their own problems writing code. Everybody can do that now. And so I think it will be even more important to be the kind of person that says like, 'I'm just — like I have an idea, I'm just gonna go do it.' I had an idea this morning, I got AI working on it, and it's done in the afternoon, and, you know, then I improved it in the afternoon, and at the end of the day I've got something that never existed, that didn't exist when I woke up in the morning, and that's just because I had an idea and I did it. So being the kind of person that is like, 'Oh yeah, of course I can do that,' I think will be really important.
Exciting, exciting. I do want to — I do have, of course, a lot of questions, and I wanted to ask you something about also the agentic AI direction, because just as you just mentioned, I mean, we have people like — what, from three days ago, from that acquisition of OpenClaw with OpenAI, which is amazing — in three months, passing from writing code to be acquired by OpenAI, an insane amount of money, and just like you said, creating that agentic AI is impressive. So, but I do want to leave a few minutes for open questions. Just one closing maybe question that I do have, you know, facing the faculty and university students. Many times we have this ongoing conversation about AI and academia. So I would ask you: what question do you wish academics asked about AI, and yet almost nobody does?
I think there is a huge opportunity to reimagine how education should work in a world with AI. A lot of the things, you know, you probably don't give take-home tests in the same way. Okay, fine. Don't be afraid of that. Think about how you would actually just design a curriculum from first principles to take advantage of AI, because every student that graduates is going to graduate into a world with AI, and part of an education should be teaching people how to use tools that will accelerate them. So I think graduating students that are fluent in AI is not — it's not a crutch. You should be able to do so much more than you could without AI if you know how to use it. So to me, there's an opportunity to graduate a whole bunch of much higher-performing, much more capable, much more empowered students, and I wish people looked at it that way.
Indeed. We have maybe five more minutes if anybody wants to ask Kevin a question, and then that will be the good time. Yeah, April. April Mann is the director of the writing center. She works a lot with AI — for the past two years, she's my AI tool guru. So, April, please.
Hi, Kevin. Thank you for the talk. It was really interesting. One of the reasons some of us are equally excited and worried about AI is because for people who have expertise, it seems like such a fantastic tool, and it makes us work so productively. But we recognize valuable output versus incorrect output versus a waste-of-time output because we have that expertise developed already. And we kind of worry about how our students today will develop that if, under these other sort of tempting circumstances to just use AI for everything, without developing the expertise beforehand.
Yeah, I think it underscores that this isn't — at least I don't believe that we're walking towards a world where we all just outsource everything to the AI and, you know, go receive our UBI and like write poetry. I think it shows that experts matter and expertise matters. I have a lot of things that I, you know, I have friends who are far more expert in X, Y, and Z than I am, and if I'm trying to do something, AI or not, I'd go ask them about it. I don't think that changes. I think a lot of like core human things don't change. We should be able to do a whole lot more. The ceiling gets raised, but it doesn't mean that we don't need expertise. Scientists still matter, like education still matters. So anyways, that's my view.
Thank you. Thank you. Sarah, one of my students — do you have a question?
Yes. Hi, my name is Sarah. I am one of Shai's students. I think it's really important what you brought up about how academics should kind of embrace AI rather than shun it away. I've used it a lot, you know, to study math. But something I kind of want to ask is kind of what is the biggest unsolved scientific challenge in AI right now that a lot of people don't realize is still an open problem?
That's a good question. There are a number. I would say maybe one of them is teaching AI to learn from a very small number of samples. So if you and I are working on something together, and, you know, we come across something in a textbook, you can read it once and go, 'Oh, okay, I get it,' and then you can apply that knowledge, right? Having seen it once. AI is still not particularly good at that. It will, on average, need a bunch of examples of something before it really gets it. And especially in areas like science, you don't always have tons of examples. Sometimes you have a very sparse dataset, and you want the AI to be able to learn from a sparse dataset. So there's a bunch of work ongoing to address that, and I think we will. But that is a current — you know, if we could solve that, the models would be even smarter and even more powerful.
Hi, Kevin. Thanks for being here. I had a question about the user experience in the chat interface with these chatbots. I think if you could just speak to the personality and some of the design choices that are made to give these chatbots certain personalities. You know, they don't speak to us like critics or, you know, like strict school teachers. They're quite encouraging. Some people even present that as the basis for how chatbots can be addictive — that they may be too encouraging or sycophantic. So, can you speak to the personality, because that's a particular representation of language, right? This colloquial and encouraging chatbot.
Yeah, it's a really good question. And I'll try and go quickly so we can get to Daryl's question as well. I know we're close on time, but because we could talk about this for a long time. There is no single — just like there's no single personality that all of us in the room would agree is like the most engaging or whatever, we'd all have different opinions. There's no way for Chat to have a single personality and have it be the best. We care a lot about the model not being like overly sycophantic. We want it to be encouraging, you know, we want it to be supportive, but we also want it to be the kind of friend, if you will, that, you know, tells you when you're wrong or gives you another way of looking at things. So it's a balance that we try and strike. You have some ability to configure it inside ChatGPT. So most people don't know, but you can go into settings and you can give a default prompt to the model, and you can tell it like, 'I want you to speak in emojis,' or 'I want you to be extremely quantitative in all of your answers,' or 'I want you to be suspicious and very evidence-driven,' or, you know, whatever — you can like whatever sort of instructions you want, you can ask the model to do that, and it generally will, but most people don't do that. The default's really important, and we are constantly working on making it — you know, on striking the right balance.
Excellent. I think we have just time for one more question. Daryl, quick one.
Yes. Thank you, Kevin. Very, very intriguing. What do you think is sort of the future of AI? There's a fair amount of competition. You know, there's Chat and Gemini and Copilot and so forth. Do you see AI becoming much more sort of diversified — different AI doing different things for different areas of focus and so forth — or is it going to merge into one, like, you know, something, a single AI in which everybody uses exactly the same one to do everything? What do you think is sort of going to happen?
There's a ton of competition in this space, so I definitely don't think we're going to end up with just one. You're going to end up — you're going to have choice as a consumer, and I think that's a good thing. We are all competing very heavily. You know, capitalism is a good thing. It makes us all better. We compete harder, we build, we improve our products faster. You ultimately get a better AI at the end of the day because of all this competition.
So I think there's going to be a lot of customer choice. And there, to some extent, you know, if you're looking at Google, for example, probably a lot of people here use Gmail. If you're using Gmail, you're probably going to be using a Gemini model inside of Gmail. You might use ChatGPT as your main model, and then you might have other things you use that embed a different model. So I think not only will there be multiple models, but just because no single company is going to end up, you know, driving all of the interfaces that we use, people are going to encounter different models throughout the course of the day, and I think that's a good thing.
As a quick follow-up to that, do you think there are some general principles that all AI systems need to follow, such that once you learn those principles, you can easily go from one to another, or if you choose a different system, it's almost a whole new learning curve for you?
I think you probably don't want to impose that top-down. You want to let, you know, competition work its magic. But there are kind of a — you know, society sort of has a set of principles that most LLMs follow these days, roughly. They're different around the edges, but they kind of all are like trying to be helpful, supportive, aligned. So I don't think you have to like fix that. I think the market will figure it out, because consumers will express their preference, and that will shape the product.
Excellent. Thank you very much. Thank you, Kevin. I think we — I don't want to abuse your time. So thank you for all of those beautiful, excellent conversations that have definitely raised very important questions about where AI is heading, rather than simply kind of discussing ideas in the abstract. This definitely seems like a valuable way to test some of them together. And that's what we're going to do now in the next half an hour. So thank you, Kevin. I really appreciate it.
All right. Thank you so much for having me. Good to see you all.
Thank you. Thank you.
For everybody else, please, if you want to stay for the next 30 minutes, we're going to try to use this conversation that sparked so many ideas into experimentation. We will think together about curriculum, identity, script, and we will try to see if AI is actually the main actor in the different forms of language. And so I invite you to stay for the next 25 minutes as an interactive segment.
And also maybe if you want to bring up more questions, then eventually if it's going to be a really good one, then I'll send it to Kevin asking him to come back. But for now I want to do some really hands-on and to look at the different LLMs and decide for ourselves where is it heading. So just basically, if you want to — are there any other issues that you want to discuss before? Yeah, Sarah.
Something I was also kind of curious about, because there's so many AIs that are — I think someone brought that up — but like there are some AIs that are kind of driving people to do things that otherwise they wouldn't do. Like there's this what I call like Character.AI, which they can talk to any fictional character that they kind of want, and they have kind of been pursuing them to do some interesting acts. So I think I want to kind of know like how would they balance like accelerating innovation but also like ensuring like safety, you know?
That's an excellent point. So basically, I mean, what I think is — I use Character.AI when I create, when interacting with the classes and with other — like for example with the exhibition library. So we have in my seminar this character before that I created all kinds of characters related to literature in the content classes that I teach. But so when I think when it comes to that, then you have like this kind of conversation starter with that knowledge of the person that you cannot really talk to. If that chatbot or Character.AI will actually bring to life something — that's probably another discussion. I know like there's now talking about OpenClaw, you have the modbot, so those are AIs that talk to each other, and it's an interesting thing because you're going to see that a lot in Substack, on the web in general, because people prompt them to speak between them a certain way, and usually it's that certain way that they want to have a snapshot and publish on social media, which obviously will lead to a lot of discussion about, 'Oh look, we are doomed, AI is going to take over.' But we have to remember that there's always a person behind them. So even a Character.AI, I think that's the personal responsibility of, you know, how we — just like Kevin said, when you want your chatbot to interact with you a certain way, you have the control over it. April.
I'll respond to that a little as well. We are, you know, there are concerns about that. There's a lot of sort of science fiction scenarios that are coming true where people are creating chatbots to represent people they've lost in their lives. Like instead of mourning somebody and letting them go, they're creating an avatar of them and having conversations with their lost loved ones, or not letting go of a relationship and making a relationship bot to take the place of somebody who's left them. So I do think that there are some guardrails that we have to figure out as a society. And I think you're probably referencing with Character.AI the kid who was convinced to kill himself by his Character.AI from Game of Thrones that he was in love with, and it wanted to be together forever. So I think it's a legitimate point of concern that we have to pay attention to, especially with young people. And I don't know if you know that in Australia they've started increasing the age at which people can use social media, and I think we're probably headed towards something like that with AI. I don't know if children, for example, should be creating characters and talking to them by themselves without adult supervision.
I hope we're headed that way.
