Good afternoon. Good afternoon and welcome. It's a pleasure to welcome you to this inaugural event in our Dean's Lecture Series and to extend a special welcome to our alumni and friends at UMass Amherst who are joining us from around the world. I'm told that we may even have more people viewing the live stream than in the room today, which feels very on brand for a conversation about technology. Anyway, those remotely, welcome.
I'd like to begin by thanking Don and Rob Manning for their transformational support of the Manning College, including the Donna M. and Robert J. Manning Dean's Fund, which supports this series and so much more across the college. Thank you.
We created the Dean's Lecture Series to provide a forum for bringing together perspectives from across the university to engage with some of the most pressing and complex questions in computing and society. At the Manning College, our work is grounded on our commitment to computing for the common good — the idea that the technologies we build should serve society in responsible and impactful ways. It's hard to imagine a topic more timely or more consequential than artificial intelligence. Its effects are shaping how we teach, how we learn, and how we live. And those effects are only accelerating.
We're fortunate to have someone with us today who has been at the forefront of helping the public understand these changes. Kevin Roose is an award-winning technology columnist for The New York Times and the best-selling author of three books, including his most recent, Future Proof, a guide to navigating the technology future. By the way, I hear there's another book coming out soon. He is also the host of two New York Times podcasts: Hard Fork, a weekly conversation about the rapidly evolving tech landscape, and Rabbit Hole, an acclaimed series exploring how the internet shapes our beliefs and behaviors. His work examines the intersection of technology, business, and culture, with a particular focus on artificial intelligence, social media, and the broader societal implications of emerging technologies.
Kevin will be joined in conversation by my colleague Ethan Zuckerman, whose work many of you already know. Ethan is a leading scholar on the social and cultural impact of technology and a key voice in the field of public interest technology, an area focused on ensuring that computing serves the public good. Ethan is an associate professor of public policy, communication, and information here at UMass Amherst. His research explores how media and digital technology shape civic participation and public discourse, and he is the author of Rewire and Mistrust, two influential books on global media and institutional trust.
Following their conversation, we'll have time for questions from the audience, and then I hope you'll join us for a reception downstairs immediately after the program. Please join me in welcoming Kevin Roose and Ethan Zuckerman.
Well, Keith, thank you so much. And Pat, thank you for taking a break from your vacation to us. Just so that everyone knows, it would be nice... tell them what you otherwise could have been doing today.
Well, I just left Disney World where I was with my family for the last couple of days, but honestly, and you know, please don't tell anyone in my family, it's a nice break from carting their four-year-old around a park.
Usually people get off this stage and immediately say, you know what I'm going to do next? I'm going to go to Disney World. But you... I prefer... yeah, I don't think they're going to be as... it's tempered at first. Today, although who knows, things could get a little spicy. It's an interesting moment. We are, when you're at the end of the semester, it is week 12 of the 13-week semester, which means that for a lot of us the end of work is in sight. You had an interesting conversation recently with someone who told you that there are 300 days of work left. What did he mean and what does that mean for my job and for your job?
Yeah. So I come to you from San Francisco. That's where I live, in San Francisco. We have been gripped by this new religion, this new religion sometimes known as AGI, being AGI-pilled. But this is sort of the idea that we are on the cusp of a transformation unlike anything in human history, probably at least since the Industrial Revolution, probably ever. And people who are AGI-pilled do all kinds of crazy things. They sign up to have their corpses frozen after their death so that AI can reanimate them later. I met one guy who doesn't wear sunscreen to the beach anymore because he's so convinced that AI is going to solve skin cancer by the time it's relevant to him. People don't invest in their 401(k)s in case... it's a crazy place right now.
But one of the conversations I had recently was, I was talking to a guy and he said, 'I think you've got about 300 days left of work.' And I thought to myself, like, oh, this year? Like, oh my god. And then I realized, oh no, he means ever. And you know, I think what he meant was that in his field, software engineering, there are many people who predict that we are no longer going to have software engineers, or at least none that are recognizable doing the tasks that they do today. I think his point was broader. I think it probably was something more like all white-collar knowledge work, which, to be clear, I don't share that belief, so I was merely passing it along. It's sort of a little bit the crazy people in San Francisco belief, but this is very much their... inside the bubble, this is the expectation — not 300 days necessarily, but that we have sort of single-digit years left before large swaths of the economy are transformed, AI is doing all the work, humans are there to supervise at a remove, and just the world as we know it will sort of cease to exist. Not necessarily in a way that means we all die, but maybe we just sort of become relegated to kind of like ne'er-do-wells or something like that.
I know you talk about, you know, coming from the AI bubble, from the land of the AGI-hold. And I will say, you publicized this honestly, how you discovered you were, for the first time, literally while you were an undergraduate, wrote a piece that I think people should read more often, called 'The Unlikelyst Disciple,' about essentially going undercover at Liberty University and just exploring the ethics of the place from the perspective of someone who would be raised secular. I think you may have gone a little bit further, you know, now as someone who is sort of working on this in real time. You had to fend off a ham-fisted search engine today by Microsoft, and you will be... you've gone fairly deep down the Claude Code rabbit hole. What's your relationship to AI and to AGI? What are you using it for? What are you not using it for? What are your thoughts?
So, so I use AI all the time, and this is very controversial and I'm in the minority among my colleagues in journalism, but I find it enormously helpful. And I just finished — well, just as of tomorrow I will finish, or perhaps today — a book about the race to AI. And there is so much about the writing of any work, but especially something as long as a book, that is not about the actual words. It's the vetting and the endnotes and the citations. And I have basically offloaded everything that I'm able to offload ethically and professionally to AI. So I have a team of... that I write with. Obviously I have human editors. I have a human researcher as well, who are wonderful. But for things like research, where I can just now put all of my materials, all of the articles, all the interview transcripts, all my source material into — I use a couple of different AI tools, but into NotebookLM in this case. And then I can just ask it questions. You know, who are the three people in this meeting in 2019? Or what's the best argument against this line of technical research? What might this person have been thinking when they said this thing? It's been enormously valuable just for me to have this sort of litmus test of all my source material.
And then I have a team of agents who review my work as I feel like it's pretty good. I have a fact-checker that goes through chapter by chapter. My strategy with that, considering an AI, because a couple of years ago it was making stuff up all the time and we were having to fact-check it — I always fact-check it. It catches some things that are actually mistakes that I made that I can then go back and fix. Obviously also working with human fact-checkers. So all of this is kind of experimental and new, but I found it enormously... what I would probably say would be something like a couple of years of research on this. But I've established two bright-line rules, or my editor said I had to talk about this in the foreword to the book, which is: one, the words. I am not having this written for me, because I think that would be an ethical conundrum and something that I am loath to do when they are tired. But also because it's just not that good. It produces stuff that sort of looks good and coherent, but you... in like one layer, it's like some of the ideas just don't connect. And then the second rule, other than sort of I write my own words, was that I'm just not feeding any sort of sensitive documents into a cloud-based AI system. So if I have a confidential source, if I have something that's legally sensitive, I keep that pretty well walled off from there.
So, so how do you navigate the gap between your frightening beliefs that there's 300 days of work left — really receptacle about it, but clearly part of the bubble of the world that you're coming from — and very real... you don't say, I'd write not just games as an... It seems like we're going through this revolutionizing moment where AI has done very, very dumb things. I think almost anyone who has coded in a serious way, in any way, can see that a lot of that job — I mean, you know that a lot of that job can be as well done by someone that codes as a... by AI. It may be possible where there are other fields, perhaps the right fields, that are immune to the same strategies of scale, to the same strategies of particular forms of reinforcement. We are at a... we're a diverse university. You've got people studying everything from nursing to turf science to philosophy to a lot of computer scientists here. How many of these fields do you think are going to get transformed? Do you see a moment at which AI becomes good enough to be replacing your own writing, or do you feel differently about this transformation?
That's a great question. There's this concept that people in Silicon Valley talk about called jagged intelligence. It was a term sort of popularized by a researcher named Arvind Narayanan, and what he meant by that is essentially what you just said — that there is, there's like deep spikes in AI capabilities. So for example, AI is now, you know, in coding competitions, it is better than almost every human coder at those rare competitions. It is outdoing many math PhDs and physics PhDs in their chosen fields. It is doing incredible things with agentic systems that can go out and work autonomously for hours at a time. But if you, you know, give it simple tasks, it sometimes fails at things that would be relatively simple for humans. So what we have is not sort of a perfect representation of all human intelligence. I think until now, or at least at some of the early AGI-belief people, they probably would have had this sort of burst of capabilities explosion that would sort of rise, that the whole water level would rise. And I think what we're seeing now is that the gains in some fields are just much bigger than in others.
I think that's what people mean when they say 300 days left of work. That is the field that they're talking about. So, so I think in some sense that's a tech-people problem. It's just easier to build these systems to get quite good in domains where there's a right answer. It's too big... you can train a model to go out and write a bunch of code and then write unit tests to see whether the code works. And you can do that over and over again using this technique called reinforcement learning until the model has stopped making errors. That's not possible to do in every domain, and not every domain has clear right or wrong answers. I think in a funny way, this is actually kind of a little trick. We were talking before about computers learning how to play chess, and you know, we ended up quite a ways with Deep Blue facing off against Garry Kasparov, and that's a clear event where you can literally look at it as, yes, you won, good job, pat on the back. Coding feels like something where we could often give the pat on the back. Why is writing an article different if the output is from this? Oh, it's more subjective.
And I think it's also connected to this concept of self and expression. When I use an app on my phone, I don't think that like the code that is running in the background is an expression of the programmer's authentic self. Like, it's a functional thing. And coders get very, you know, sometimes they get very upset. They think they, you know, they do express themselves. But I think for most users, they just care whether the thing works. And I think writing is not like that. We've seen now, actually, romance novels are sort of the first genre of fiction that has fallen to AI, or is in the process of falling to AI. And romance novelists are out there, dozens of books a year using AI, but they're not telling people for the most part that they're doing this because they found that the moment you say this is AI-generated or even partially AI-generated, nobody wants to read it. So there's something about at least the illusion that we are hearing from a human and the inner thoughts and experiences of a human that is somehow rooted in a person that makes us believe in it and value it more than we would if we just thought it was AI. Even if the text generated by the AI is better on some sort of discrete level, we would still want that to come from a human because part of what we're doing with a book is trying to connect with the person.
This is probably the moment at which we should make our disclosures. I've gone full on on this. I've got a... you got the headset with Anthropic's new models going into your head. This was what you requested. It's in your neural... you got models running inside there. So, your last book before this new book about AGI was called Futureproof, and it was this idea — and this was a while ago, back in 2021 — 'Nine Ways to Live in the Age of Automation.' You were telling people at that moment about this idea of sort of leaning into their humanity. And you, I think, just made a fairly effective case: you need to understand that I am a real human making the judgments about the words on the page. I am putting my words into the podcast. How does that apply more broadly to the folks in the audience here, particularly to the folks in the audience here who do, to one extent or another, write code — whether it's a form of expression, whether it's a form of gig work, whether it's the job that they have right now?
Yeah, I think interpersonal skills are about to become much more important in computer science, for reasons that are actually, like, I'm not predicting, but I was telling your faculty earlier, I was at a startup recently and you walk into their office and everyone's got their four-monitor setups and then everyone's also got like a little microphone at their desk because these programmers, they no longer code with their fingers, they code by talking. They just tell their agents, 'Hey, solve this problem,' or 'adjust this thing,' or 'run this test,' or 'build this RL environment,' and the agents just go off and do that. And it sounds like a call center, but what they're doing is communication. They're not communicating with each other as much, but they are still communicating with each other. More of their time when they're actively working is spent communicating with these agents, and they're finding that actually the skills that make you good at communicating with humans, you also need to use them with these AI systems because you're going to get better work out of it if you give them your instructions, metrics for success or failure. If you come to them in an empathetic way, for whatever reason, it seems to produce better code. So there's an interesting kind of scenario unfolding in computer science and in the tech startups in the Bay Area, which is like people are now starting to learn empathy for the machine.
And it's becoming... and I mean, it all translates out of here. Yeah. Six don'ts. It's true. Like, they're trying to not burn them out. Like, there's obviously, agents don't get burned out once, say, as humans, they don't like being on vacation, but they do sort of, the quality of their work degrades if you give them a task that is too much for the context window. So there's this interesting dynamic where now every engineer is kind of a manager, and they're realizing, okay, oh, we actually have to teach this new kind of management to people, even if they were sharp individual contributors.
So interesting to know there's a direct parallel to this. I'm a professor at a university where a big chunk of what we do is research. We have very powerful, flexible research assistants. They are a bit of a scarce life for these days, having a lot to do with funding challenges in the environment and such. But this is a conversation that I know a lot of fellow professors are having, which is: I can hand this task to a grad student and I can help them learn how to do research. I can help them learn how to formulate queries and send them out there. Or I provide code and I can probably get the results faster and I don't have to hear about the fact that they got locked out of their apartment. This is a real moral challenge for the world of research, and the broader challenge is if it is getting easier, easier, easier to do work on your own without some mentorship, how do people learn? How do you learn to ask those research questions? How do you learn to do that really detailed sort of work that I'm guessing you and I would agree probably still is going to require a human touch, at least for some time.
Yeah. I mean, what we do in a sort of fully human setting is we just have feedback, right? We have training, we have feedback sessions, we have peer reviews, we have performance reviews. That happens once a year at a lot of places. What if you could do that every day with your agents, or twice a day? I say, 'Hey, how am I doing at giving you the kinds of instructions that are producing good work?' And obviously they're going to, you know, flatter you if you don't instruct them to. They're going to say, 'Oh, you're the best manager we've ever had. You're so amazing. I'm so lucky to be working for such an empathetic genius.' But with some skillful prompting, you can actually get them... I mean, one of the things that I love to do with my agents when I'm writing is I have one agent write a really positive review of whatever I've just written and one write a completely scathing takedown, and I then have sort of an agent who synthesizes and says, 'Well, this point looks a little more on the nose.' And I find this very helpful. So it has just, like, tightened my feedback loop a lot.
But I think you're right. There's like, there's some piece of learning to be a good colleague, a good manager, a good student, a good teacher that is just about showing up and putting in the work. And I don't know that you can fully replicate that with agents. So I think there is still a valuable role in training people how to give feedback, how to receive feedback, how to work on a team, how to express leadership. Like, I think those are all very, very valuable parts of what students should be learning.
We were talking earlier about one of the problems in academia at this moment, where when you're studying AI, the academic publishing cycle is so long that when new research comes out, it may actually be out of date. So I'm almost certainly going to cite some out-of-date research here. But there is research that suggests, looking at software that's been put into public repositories, things like GitHub, researchers have gone and sort of looked to see who seems to be making the best use of these models. And what the researchers found in that particular study was that people who were experienced programmers, people who really knew how to code in and out, seemed to have a real productivity increase using these tools. People who were much newer to the game, were starting to use these things, didn't seem to be using them very well, looked like they were actually having a skill reduction in using some of these things. Now, you may be ready to tell me that the new tools are so much better than this that it isn't going to happen anymore. The question I'm trying to cue up is that we have a real challenge as educators helping people become computational thinkers. How do we get people to think in terms of how a machine solves a problem? Historically, we have done this by asking people to do the very painful task of coding a computer for hours and hours on end until it literally reshapes their brains. Now, you can get your way through almost any intro coding class by feeding your assignment to Claude Code, and it's probably going to do okay. And unless we do some of the things I've seen some of my colleagues do, like have people do blue-book coding exams, it's hard to know whether we're going to be able to teach that sort of computational thinking. Is there hope for us around that, or should we assume that these tools are getting to the point where having that sort of knowledge may not be as important as I think it is?
I don't know. I'm not a computer scientist, and I always try to express my own doubt about my intuitions when it comes to fields that I'm not an expert in. My impression is that the changes in coding are just sort of moving several layers up in abstraction. Right? So we used to teach people assembly, right? How to, like, code literally into the computer, and then we abstracted that to programming languages, and those programming languages got more abstract over time. And then we had, you know, object-oriented programming and, like, all these changes in computer science have basically been moving from, like, something that talked directly to the computer to something that sort of talked to a set of abstractions that then talk to the computer. And so I think on one level what's happening with AI right now is just, like, another layer of abstraction. It's like you still need to understand something about what is going on.
But for example, I don't code, and I've built probably 20 apps in the last six months. And parts of it, like, I could have saved myself some time if I had known something about coding, but other parts I think I was probably actually more effective at than someone who's a professional coder because I was just coming in completely blind and I turned sort of the planning and thought process behind the code over to the coding assistant. And so I was not coming in with my own preconceived notions of, like, 'Oh, we're going to need a SQL database for this and you're going to want to write it in this language because this language is better for this kind of thing.' I was coming in as, like, a total neophyte, just saying, like, 'Hey, I want an app that does, like, these six things. Can you kind of figure it out for me?' And it would do that. And so I have learned some computational thinking just by looking at what the model is doing on my behalf. And you can actually now inspect the so-called chains of thought, which I find fascinating.
Well, this is my favorite part of using AI right now, is that you can actually see the inner monologues of the AI systems as they're sort of trying different things. 'Oh, that didn't really work. Let me back out. Let me try it in a different language.' I find that I'm learning a lot from just watching that work. So, is that useful, applicable knowledge? Maybe not. But I think it is teaching a kind of computational thinking that is more attuned to the realities of the tools.
And I agree with you. Like, we have at least two categories of people, possibly three categories of people. Category one are the professors and the faculty who are often trying to solve giant research problems. And the world that you're describing, in which you can throw a lot of documents into Google Notebook or another RAG-based system and be able to sort of query them, that's pretty exciting. It's pretty sexy. There are folks like Justin Gross sitting in the front row there who I know are trying to think about how data scientists do incredibly sophisticated work both with and without AI. There's also a lot of people in this audience who are wondering whether they're ever going to be able to get a job and don't have tenure, and they are looking at the fact that Meta just laid off 8,000 people. They closed 6,000 open jobs. At the same time, Microsoft is offering early retirement to about 9,000 people. So those folks are very interested in what world are they going out into. We also have in the front row some people who are really interested in how you keep a university open if the people who wanted to come here to get jobs now start wondering if those jobs are going to continue to exist. So can you make categories two and three happy, or at least offer them some sort of consolation?
Yeah, I mean, look, I am very glad that I do not have to solve these problems for a living. I have great admiration for the university officials and administrators and faculty whose job it is to navigate all this. What I can tell you from my perspective is that I think that there are a lot of institutions that are right now trying to kind of adapt the old ways to the new reality. And I think that that's going to be a hard thing to do because these models are getting so much better every single year that you are in essence trying to lesson-plan, curriculum-plan, and plan the structure of a university around something that is going to be obsolete or on its way to obsolete by the time that you finish. Take it from me. I wrote a book about AI the year before ChatGPT came out. It is very hard to make predictions about the future when you're on an exponential like we are with AI right now.
I'll tell you what I would do if I were starting a school. And obviously, I'm glad that I'm not starting a school. You should never send your children to my school. I would devote about half the days to just pure offline learning, like pure sort of Socratic — like, let's read a bunch of books and let's talk about them. I would have no devices anywhere near it because I think there is, to me, the danger of sort of trying to hybridize everything is that you end up with the worst of both worlds. You end up with, like, the technology that is disrupting all of the work that's happening in the classroom and you're also not learning anything. So I would, like, half the days, or maybe half of each day, would just be like education as it existed in the 17th century. And then the other half would be like giving everyone unlimited budgets for these AI tools and just having them build stuff because I think there is really no better way to learn about the technology than by using it. I have found my own learning has come mainly through doing.
And so that would be the other half of the day, is like, let's stay pretty close to the frontier, like, use the best tools available, figure out what they can do. Because right now in the job market, these are the categories of people who are getting jobs, right? It's like the people who have skills that have nothing to do with programming computers — there are still lots of jobs for them. And then within computer science and programming, it's the people who are really skilled at getting the most out of these tools. They're now... they're no longer... you may remember the phrase '10x engineer.' This was sort of the old term of art in Silicon Valley for, like, someone who gets 10 times as much work done as the next person. That's no longer impressive to them. Now they're talking about 100x engineers. And some of those are senior people who have gotten really used to the tools, but some of them are, like, 17-year-olds who are just, like, really, really obsessed with these tools and are using them every day and are figuring out how to, like, do amazing things with them. So I think that that would sort of be the way that I would push education, is sort of away from this kind of messy middle and toward these more sort of extreme-sounding ends of technological adoption.
I think you may have just updated Maria Montessori for the 21st century, at least for the Bay Area. So if you get sick of this New York Times gig, we can probably start up a school for AI-driven engineers. My kid goes to a Montessori preschool, and all preschool teachers should make $1 million a year. That is my actual view. That is a thankless and hard job. We have a department of education here who I think would be very, very happy to hear that.
When we were talking before, you actually sort of broke down a distinction that sounds a little bit like how you were talking about breaking down that day. There is a set of jobs that are going to be automated. There's a really interesting philosophical question about how broad or how narrow that set of jobs is. My friend David Jensen was making the argument earlier today that it's a narrower set than you think and that the hype people are very good at it. There's other folks who say, no, if we got enough data we should be able to do it, and that might mean biology and then law and so on and so forth. But you were also acknowledging a human set of jobs where you think we're going to be resistant to automation, or automation is going to be really hard to do. Again, keep in mind you're talking to educators. Keep in mind you're talking to people who genuinely are doing this for love, since most of us could go work for Anthropic for more money. How do we think about educating people on sort of...
of both sides of that equation?
Yeah, I think it's the right question. I think the thing that I have been trying to grapple with is sort of the difference between jobs and tasks, because there are occupations where we could have automated them a long time ago, right? Like everyone's got a coffee maker in their house. That is essentially a form of labor automation. You know, a person used to have to make the coffee, now a machine does it. But we still have coffee shops. Why is that? It's because maybe the coffee is not the thing that you are paying for. Maybe it's the experience. Maybe it's a quiet place to get some work done or a nice place to have a conversation. Maybe you just like when the barista, you know, says have a nice day to you in the morning. There are these little coffee kiosks. Anyone been to the San Francisco airport and seen the coffee robots? So they've got, like in every terminal now, they've got these robots that'll sort of descend with the giant claw hand and make your latte, and they're not popular. Like they have had trouble keeping them open because even though they are faster and maybe even a little cheaper than the Starbucks that's just, you know, 10 ft away, like people really want that human connection. There's something about that that feels good to us. And so I think there will be jobs where even after all of the tasks inside the job or a large fraction of them have been sort of eclipsed by AI, like we will still want that. I was giving the example earlier like a therapist, like it will probably be true that shortly AIs will be better therapists than humans along kind of any conventional metric that you would use to evaluate them. They're a lot cheaper. They're always available. They've trained in every modality because they've been trained on all of the data on the internet. So, you can make a case that the therapist is doomed. But I think there's a subset of people and probably a large subset of people who will just never want to go to an AI therapist. It just, even if the advice they're getting would be identical to or superior to the advice a human therapist would give them, that just feels bloodless. It feels bad. We don't like it. There's something about us that is hardwired to love human connection. So I think that there one place where the sort of AI bubble often gets it wrong, I think, is insisting that as soon as a job can be automated it will be automated because they see that for themselves and their companies. But I think that is not true in the economy as a whole.
I'm gonna open this up in just a moment to the audience and I'm gonna roam around with a mic so that the rest of you can have some of the conversation that I'm having with Kevin. But I want to give you sort of a three-parter, and it's sort of three versions of the same question. Obviously you're writing your next book on artificial general intelligence. You've been following this as closely as anyone else. As you think about this transformation that we're in the midst of, what excites you the most about it? What scares you the most about it? But here's the hard one. What do you think we're missing? You have written about an incredible range of subjects. You are a much better technology columnist than I am, but we both have the challenge of trying to figure out what to write about in tech. And I've discovered that 85% of the columns that I write these days are about AI. What aren't we seeing? What aren't we writing about? So, what excites you? What scares you? What are we missing?
What excites me is that I think if we manage to solve all of the big problems around AI, obviously there are tons of big problems, everything from addictive chatbots to environmental harms to just the existential risk that all this could kill us. Like, I think this could be really good if we get all of that right. I was just at Disney World and I went to, they have this amazing ride called a carousel of progress. Anyone ever been on this? This is like an old-fashioned ride where you go in a theater and it sort of rotates around and they show you this sort of series of animatronic tableau of like life in 1900 and life in 1920 and life in 1940 and like each time they have a little bit more technology. You know, at one point they have an ice box and then the ice box is a refrigerator and then they have electricity and they're not like, you know, doing the laundry by hand and it only takes them, you know, an hour to do the laundry and it's like, oh my god, life has changed. I watched that and I thought, well, we need something like that for today. We need someone to give us a positive vision of how our daily lives could be better 10 or 15 years from now. Because I do think that technology generally improves our life. I am not one of these people who like wants to go back to an agrarian economy. Like I like my, I like my, you know, vaccines and my iPad, and so I do think that on the whole if we manage to like solve all of the contingencies, this could be really good for us. It could free us up to do the things that we love. We could cure a bunch of diseases. I'm very excited about what's happening in AI and medicine. My own father died of a rare form of cancer that I think could have been treated with better interventions. So I'm very excited about the potential for AI to help in medicine. That was the positive side. The what worries me is that we are just not ready for this. Our institutions, our government, we are not making the kinds of changes that we would need to make if all of the things that the people in San Francisco are saying about AI are true. I worry about existential risk. These systems are behaving in ways that their creators did not anticipate and do not know how to control. Some of you may have seen the recent coverage of Claude Mythos, which is Anthropic's new model, which is finding zero-day security vulnerabilities in 27-year-old software and appears to be wreaking all manner of havoc. And so they're releasing it to defensive cybersecurity teams before they're releasing it to the public. I think that stuff is very underappreciated. I think bio-risk is very underappreciated and I just don't think our government has a good track record of regulating new technologies. You know, we still haven't regulated social media which is 25 years old at this point. So I think we just need to like get our minds around the pace of change here and really like start preparing for continued change. And then I think the thing that I don't know we collectively are missing is that the sort of old adage that like today's AI is the worst it will ever be. Like I think we have just started to get our minds around collectively the progress that has been made in the last couple of years. But there's, to me, there's no obvious reason that these things will plateau at anywhere like at sort of around human level on a bunch of different tasks. And so I think the thing that is underappreciated is like what if the weirdos in the bubble in San Francisco are right and this thing just keeps scaling forever and ever and ever.
That for me is perhaps the scare as well as the underappreciated one. But I think that notion of taking the weirdo seriously is an interesting provocation and maybe an interesting provocation in a room where I think you might have more skepticism than that. I'll be really curious to see where the questions end up going as far as sort of an enthusiasm or skepticism. We're going to start circulating a mic in the audience. I just, because we are at an academic institution, I just want to offer a brief reminder about questions. Questions are interrogatives. They often end with a slight uptick in tone indicating that you are interested in what Kevin might say in response to this. They are not in fact short speeches. If you find yourself saying, I guess that's not really a question, it's more of a statement, that is not in fact a question. And so if you're planning on heading that direction, please don't. I was actually going to circulate the microphone because I have a tendency to loom over people when they start giving speeches instead of questions. But I will stay seated and hope that my speech will do it instead. I would love a show of hands of people who have questions that Kevin is capable of answering and we can go down here.
That was very good agent prompting. You're very good at this. I maybe I will actually have the ability to, yeah, audience management is very much a form of agentic prompting.
And please introduce yourself and tell us how you fit into the UMass or the Amherst community.
Um, hello, my name is Sam. I'm a junior undergraduate. I study resource economics and will potentially graduate with like a data analysis certificate. So I'm sure you've seen the recent Allbirds move from, you know, a shoe company to like a pure sort of lubrious AI holding whatever. And I was thinking about what you said about Starbucks and about coffee and how, you know, there's always going to be a human component, but Starbucks in the past few years has been hemorrhaging money on the coffee side and is sort of in the process of becoming a bank through their, you know, like Starbucks points. I guess without, you know, going too far into detail, what prevents companies that are operating under like rational assumptions and economics from turning into lugubrious AI holding firms or banks or, you know, abstract financial firm operators? What actually holds those companies to continuing the production consumption loops? Why would you do that if your return is just lower through having to actually produce and consume things and being constrained by supply chains?
I love the word lubrious. So, thank you for introducing that into the conversation. And the Allbirds thing that he's referring to. So, there was this company named Allbirds that made these hideous shoes that everyone in technology wore for several years and then they fell on hard times and got acquired at a massive discount and now are pivoting to becoming a cloud AI provider. There's going to be a lot more stupidity like that, by the way, because it's sort of like you remember a couple years ago when like companies would just say blockchain at the end of their name and then their stock price would go up 400%. Like that is happening now in the economy. I don't have high hopes for the Allbirds computing cluster. But I think you're right that there will be companies that sort of just decide that their model no longer makes sense, whether it's too labor dependent, whether it doesn't have the kind of margins that their competitors do. I kind of think that this is going to come more from, like, less from the sort of established companies, which I think are having a very hard time sort of getting their heads around AI and what it means for them. But I think it's like true that in every sector there will be like a company with like three to five employees that tries to take on the big incumbent in whatever category. So there'll be like a three-person insurance company that wants to take on State Farm and they'll like go off with a bunch of agents and do that. There'll probably be like a three-person newspaper that tries to like take on the New York Times and they'll, you know, make a run at that. Like I think there will be businesses in all these verticals that just attempt to provide the same service with a much different ratio of human labor to AI labor. And for some that may work and that may mean that you actually have to like pay a premium for if you want the human therapist, maybe that cost goes up in a world with cheap and abundant AI therapy because that's sort of the premium experience that people want. And so that's one of the things I worry about is a new kind of inequality, not necessarily of wealth will track to that, but like of access to human labor. And access, you know, it may be that the students from disadvantaged backgrounds go to schools where all the instruction is done by AI and the students who can afford to pay for higher price educations will get the contact with the brilliant human professors that you all enjoy here, and that will become sort of a luxury good. So that's one of my sort of fears about this is we're going to see a stratification.
