Good morning, everybody. Good morning. My name is Peter Chin and I'm a professor of engineering here at Dartmouth College, co-chair of the faculty AI leadership group, and a faculty chair of this conference. I want to thank every one of you, every beautiful human being out there, for coming and joining us at AI at 70: What Must Remain Human.
When President Beilock shared Huckland with us to have this event about a year ago to mark the 70th anniversary of the 1956 Dartmouth summer workshop on AI, my first question was, will anyone show up? And I'm happy to report that there are more than a thousand of you attending this conference in person and a lot more online. There are industry leaders, government policymakers, researchers, faculty and students, AI critics, AI enthusiasts — all of you, you are all welcome.
I'm especially thankful that lots of alumni came back to Dartmouth on this beautiful autumn day to share what they have learned here and how they are leading in the field and beyond. And I'm especially excited to have many Dartmouth students in the audience and participating throughout this conference. And I once again want to welcome everyone joining us here in person and online via live streams.
When, uh, 1956, seven years ago, John McCarthy, um, uh, John McCarthy, Marvin Minsky, um, Claude Shannon, and Nathaniel Rochester got together at Dartmouth Hall, about a few several feet away, to think about what does it mean to mimic human intelligence. They were convinced that if there was anything that was worth imitating, it was the intelligence of a human being. And after 70 years, we are all here to understand and to deeply think about what that actually means.
And today we will see AI through particular fields — through the field of medicine, engineering, public policy, and the arts. Tomorrow we want to broaden this conversation, uh, and we want to think about what does AI mean for universities and especially young people who are graduating, who are studying here, graduating from colleges, and think about what the next seven years of their life will be. This question is an especially personal one for me because I'm here not only as a faculty member and teacher and a researcher, but also as a parent of a senior who's graduating, and they have to think about what this AI means for them, uh, as they start their own career.
Today, uh, you will find, uh, a lot of, um — oh, and I should also mention on Saturday we have an amazing plan of student hackathon where more than 140 students have signed up with 80-plus teams hacking out what, uh, what AI means for them, and we, we are so excited that so many people have joined us. You can find up-to-date program on program and website. Uh, as I mentioned, uh, all of these events will be livestreamed, and, um, I hope that you can continue to check us out as we go through the next couple of days. Lunch will be provided, uh, today at the top of the hub and tomorrow at Mafia Plaza. And we have receptions both today and tomorrow at 5:30 p.m. I hope you will join us. I hope you will continue to walk this journey with us as we, as we continue to converse.
As I mentioned a moment ago, we are participants here with a wide range of perspectives. We want your questions. We want your disagreements. We ask that you give others the opportunity to speak and to listen, and that you allow every session to proceed without any disruption. Question and answers will vary by session, but it will be mostly available. So I want you to engage with us, and I want to thank you once again for making this a very successful and productive event for everyone who is here.
Now it is my pleasure to introduce Greg Lmcow. He is a class of 1991 Dartmouth, and he's the chair of our board of trustees to formally open this conference. Greg, I want to thank you once again for your leadership, for being with us this morning.
Morning, everyone. Great to see everyone here, and thank you, Peter, and thank you to everyone who worked so hard to bring this conference together. It's a pleasure to welcome you all to Dartmouth. As an alum and chair of the board of trustees, I'm especially proud to see Dartmouth bring together so many different voices into a conversation that means so much at this time.
70 years ago on this beautiful campus, 20 researchers came together for what they called the Dartmouth Summer Research Project. Remember, this is a time, 1956, when computers could barely play tic-tac-toe. Yet led by math professor John McCarthy, this group pursued an audacious idea to build an intelligent machine that could imitate the human mind. They didn't quite get there over the course of that summer, but they did set an entire field in motion. And as the group tried to describe their ambitions, they looked for a phrase to help explain to the world what they were doing. The term artificial intelligence was born.
Today, AI is everywhere. It's solving decades-old math problems, finding new ways to detect cancer, and helping scientists discover new medicines. It's also raising difficult questions about the way we work, the decisions we leave to machines, and who gets to make the rules. There's no better place to examine where AI is headed than the place where AI was born.
Over the next few days, we have some fascinating discussions for you. We'll hear from representatives of OpenAI, Google DeepMind, and Anthropic. And right alongside them, we'll have critics and researchers challenging them on what AI can and should be doing. I'm also delighted to see so many Dartmouth alumni contributing, along with Mira Murati. We'll hear from Senator Kirsten Gillibrand, Peter Diamandis, and Alan Trefler, amongst others. And on Saturday, as Peter mentioned with a breaking update, it's now more than 80 student teams that will present their work as our hackathon concludes. I hope you'll all make time to see what they've built.
Here at Dartmouth, AI has raised big questions about how we teach, how our students learn to think for themselves, and how we pursue new discoveries. We're asking ourselves every day the same question we will ask all of you at this gathering: In the age of AI, what must remain human?
Before the 1956 gathering, one researcher, Marvin Minsky, wrote that he hoped participants would, quote-unquote, reach unprecedented agreement on philosophical and language matters. I won't make the same bold prediction for this conference. But I do hope we'll challenge one another, we'll be open about our disagreements, and we'll be open to changing our own perspectives. That's what a place like Dartmouth is for. That's what we're all about. And I hope we'll see plenty of it over the next few days.
We have a wonderful conversation to begin the morning: The Future Worth Building Is Human, with Mira Murati and Dartmouth President Sian Leah Beilock. Sian has been asking many of these questions with our faculty and students as Dartmouth considers how AI should fit into the education we provide. She'll open the conversation and then welcome Mira to the stage. Please join me in welcoming President Beilock.
Thank you, Greg and Peter. Thank you to you and your colleagues for all the thought and work you put into these next few days. I'm so pleased to welcome everyone to Dartmouth. Greg alluded to this, but one thing I hope that comes out over these next few days is that we're going to be honest that we are still figuring this out because there's no playbook yet. We've been grappling with where AI can augment learning without replacing it, where it should and shouldn't be used, what our principles should be around disclosure. These are big questions.
Leading is messy. We're going to make mistakes, not get it all right the first time. But I firmly believe that we have an obligation to set these norms for ourselves as an institution of higher learning. Our goal is to have the hard conversations and drive a direction rather than having industry or government do it for us.
To that end, yesterday our provost here at Dartmouth, Santiago Schnell, announced his resignation following an independent and faculty-led review of his AI use. It was certainly a difficult situation, but one that raises exactly the kinds of issues and challenges we want to explore over the next few days. We want to lead in an environment that is fluid and fast-changing. That means first and foremost having leaders who are modeling the transparency and human judgment we expect from our students and our scholars. And it means making space for dialogue on the most difficult issues where folks don't always agree, just as we're doing today and tomorrow.
Our ambition for Dartmouth is to offer the best undergraduate education within a world-class university. That is especially relevant when it comes to AI. As you'll hear over the next few days, we have faculty doing incredible work with these tools, developing new treatments for diseases, building platforms that help address mental health challenges and isolation. We want our students, especially our undergrads, to have the opportunity to be partners in this work. Dartmouth's size makes that possible. Our faculty know their colleagues in other departments and what they're working on. It leads to collaboration that you might not see at a larger university.
That is why I believe so strongly in a liberal arts education. Anyone can get a convincing answer from AI in seconds. The hard skill now is judgment. Is that answer right and complete? What are the ethical and moral considerations? Do I understand why? And can I explain and defend a viewpoint in a classroom, a boardroom, a research lab? That is what we want our students to be able to do.
That means building a shared foundation in AI literacy beginning from orientation and continuing through their four years. And when I say AI literacy, I don't just mean teaching students to write better prompts. What we want to help them understand is how to think about the technology, how to be uncomfortable with pushing and learning without the technology, how to know whether an answer is right or wrong, how to know when to pick up the technology and when to put it down.
We also know that no two classes will be alike. That is the power of a Dartmouth education. There are going to be classes with AI, without AI, and about AI. A computer science professor might teach a class about AI. An economics professor might ask students to use AI to help write code. A history professor might have students write tests in blue books. All three are totally valid. In every case, it is our faculty who are empowered to make these decisions.
And through places like our Dolly Lab, our Magnus Center for Entrepreneurship, our design initiative at Dartmouth in the school, our goal is to ensure that every student, in addition to their class experience, has a build experience with these technologies as well. Whether it's developing an app to help manage mental health on campus or using AI to turn a business idea into a prototype, these are skills that students will use no matter where they go out into the world.
And it makes these skills and our idea of having students learn them particularly relevant when we see alums who are excelling in this space. And with that, it is my pleasure to begin a conversation with Mira Murati. I'd like to invite Mira out on stage.
So Mira earned her engineering degree at Fyer in 2012, and today she is co-founder and CEO of Thinking Machines Lab, where she's building AI designed to work collaboratively with people. In preparing for this conversation, I asked our faculty, students, and staff what they wanted to hear most from Mira, and their questions helped shape the discussion we'll have today. We'll talk about what she's building, what she's learned from seeing people use AI, and what she would tell students here about preparing for their own futures.
And I will just say that I have run office hours every week, um, in Collis to hear from students. And I had a student last week come up to me and say his bucket list was hearing a talk in person from Mira Murati and Yann LeCun. And I thought that was a pretty good group to be in. So we're so thrilled to have you here today.
So Mira, for those who are less familiar with Thinking Machines Lab, what are you building and what do you hope it will make possible in people's lives?
Very happy to be back here, first of all. And, uh, yeah, I think the questions that we're talking about today are so timely because 70 years ago we had, you know, Claude Shannon and Marvin Minsky here thinking about whether we will actually build machines that can think, whether that's possible at all. And obviously today we have machines that can do some thinking, uh, and the question is what they should think about and who decides. Uh, and those are the questions that we're exploring here today. But also, uh, they're the questions that sort of inspired the beginning of Thinking Machines.
And Thinking Machines is focused on building strong AI systems that can advance human will and human judgment. And by will I mean, um, the intent of deciding to do something, and in judgment, deciding what to do, what to work on, and also being able to, uh, decide whether the outcome is good, bad, you know, having a standard for what's actually good.
And if you look at, um, the current capabilities of the models that we use, they're incredible, right? They're extraordinary. And if you're a researcher, um, you know, in a lab or in a university, let's say, you know, you're running an experiment and the experiment fails. Of course you can use these models and have access immediately to all publications that you're interested in. Uh, but these models don't have access to the specific experiment that you ran in the lab, and they don't know why the experiment failed.
However, of course, the researcher has this knowledge, and in some cases it's articulated and it's clear, and in other cases it's sort of like it just sits in their minds, right? Or like it sits among the group of researchers. And so knowledge has this quality that is dispersed. It's tacit. It's not perfectly articulated and written everywhere. A lot of it is, and a lot of it isn't.
And so in order to make AI systems useful, we need to have — they need to know what you're working on. They need to have similar level of context and knowledge as the people doing the work. And they also need to know the style of collaboration. And so these two things are what's driving a lot of the work at Thinking Machines.
So we build models, strong models that people can continue to train, adapt, and customize. And we just released, um, this summer two open-weight, uh, models called Inkling, and people have been using them already to customize them and adapt them. And we also have a, um, a platform that customers and researchers have been using to customize these models. It's called Tinker. And this combination is a system that allows people to truly own the intelligence versus renting it, because they can continue to train, uh, the model on their data and their use cases.
And we also work on other dimensions of, uh, intelligence, uh, and specifically collaboration. I think intelligence is multi-dimensional, and we've already been working on a lot of those dimensions and have reached amazing capabilities. But I think when it comes to collaboration and having, uh, systems that allow us to collaborate seamlessly with them, this is a very understudied area and very underdeveloped.
Uh, and we've done some work including with, uh, something we've called interaction models that allow people to work with machines in a way that's more similar to how we work with each other. And this is important because it gives us a lot more agency. Until this point, we've been the ones to try to understand how to use the computer, how to make it work for us. But now we have these incredible capabilities, and the computer for the first time can come towards us and understand us. Uh, and this is very much possible.
So these are kind of some of the areas that we're exploring and working on with the intention of creating this alternative future where people can own and customize intelligence. Um, and I think besides these technical capabilities, it's also very important that we think about the role — how we're shaping the ecosystem, society around us. This is something that I deeply care about.
Um, and I think we're going to have to have a view on, um, you know, what does it mean to have human worth in this time of machines being able to supersede us in some domains. And in some ways, we already have this view, right? Like we learn a lot of stuff that other people know way better, uh, and we master it, uh, and just the process and the pursuit of seeking this knowledge changes our relationship with the world. Uh, but I think this part of the conversation needs to come more into the public consciousness and build the ecosystem.
And a university is a great place to have that conversation, and I know we're going to be having it over the next few days. You mentioned something I think really important — that you're building Thinking Machines at a time when other places are doing it differently. A small number of frontier labs shaping the direction of AI. How does Thinking Machines fit into that landscape?
There is certainly a concentration, uh, because it takes enormous capital and compute to be able to build these technologies, and so the economics of it, um, favor scale and concentration. And so there's definitely risks with concentration, right? Uh, you can build incredible, incredible capabilities, but there are risks. You know, you end up concentrating a lot of intelligence, decision-making, um, economic value, uh, human values.
And so when we were building Thinking Machines, we thought about an alternative path — like what is an alternative future that we can bring about and nobody seems to be focusing on. And the path for us was figuring out a technical roadmap, a product strategy, a business strategy that can deliver this future where instead of just one or two or three, uh, super-intelligent models that everybody rents and consumes, we have many intelligent models that amplify, uh, already the intelligence that we have in the world through ownership and customization of these models.
Um, and, you know, people talk about concentration. Uh, I think little work has been done around it, uh, so far, but people have been talking about the concentration of power and decision-making and economic benefits for a long time. Uh, I think something that has been less talked about is — and since we're here — uh, the sort of loss of intellectual diversity as you build, you know, a couple of very powerful models that everybody consumes.
Uh, everyone sort of ends up inheriting the same strengths and the same blind spots. And, you know, as we use these systems for very important decision-making and to advance important things in the world, um, this creates this risk of sameness. And instead of this very strange, weird world with so many different ideas and people disagreeing everywhere, you end up with a convergence, um, of thinking. And so that's another risk to having this concentration of, uh, capabilities.
But it is so — by the nature of what we do, you know, it takes enormous capital, compute, talent, and so it takes real effort and action to counter this momentum of concentration.
Yeah. We talk a lot about the negative effects of groupthink, and it's so interesting to hear you talk about a small number of, uh, frontier labs sort of converging on the same ideas and that groupthink having really possibly problematic effects for where we're going. Um, you've spent so much of your career building and bringing AI systems into the world. I know our students often ask, in looking at alums and others, um, about different moments in their lives that had a big impact, and so I want to ask you: was there a moment that changed how you understood the technology's promise as well as risks?
I think, yeah, there have been a few meaningful moments. The first time that we brought deep learning into a product form was, um, creating an API for LMS — for GPT-3 when I was at OpenAI, and that was truly the first time that we were bringing deep learning into the economy. And there was so much discovery in the process of creating this new paradigm, this product, figuring out how to make this technology useful for the world. And I don't think it's something that we could have done in silo.
And it required, um, actually talking with people, working with specific businesses to figure out how to make the technology useful, because there is so much interaction and shaping to it. And, uh, but at the time, you know, we were still not sure whether, like, the product was the right thing or were the capabilities not advanced enough. There were a lot of open questions, and it wasn't until ChatGPT that we realized the capabilities had been good for a while, but it was this combination of the technology and the interface that made ChatGPT so easy to use and such a magical experience where people were discovering how to use it and they were interacting with it, and in that interaction shaping both the technology and the product.
Uh, so we learned a lot from the deployment of ChatGPT, and I think every time now that we're releasing something completely new in the world — that's a new paradigm — that's when we learn the most, because you see concrete cases. Um, like, say, for example, you know, we had known for a while and studied the issues of hallucinations or misuse. Um, there were research teams working on these ideas and studies, and we had published some of them as well.
Um, but it wasn't until we saw these cases in the wild, the concrete cases where we saw the reality of how this looks like, uh, that we were able to advance and, uh, deal with building more robust guardrails and figuring out how to handle a lot of the misuse that previously we didn't know how to deal with. And so it's extremely valuable to do research with real-world, uh, case studies. It's very hard to simulate the reality of friction and usefulness.
It's like the power of getting out of the lab. And it's one of the things I think is so special at Dartmouth, whether it's the SURF program you did where you got to experience different institutions, or our students, our undergrads with the D-Plan who actually spend two quarters off campus in their four years, either working in internships or studying abroad. There's something special about mixing what happens on a campus and outside, and it changes your thinking in important ways.
Absolutely. Yeah. My senior project was actually building a hybrid race car, and that took multiple months. And obviously, you know, we'd do something in the lab, and actually it would not work on the road. Uh, so that was a very, very meaningful experience.
And it's why we're really pushing to make sure that all of our students, regardless of major, get a build experience, because we know that you can see when it fails. We like students to fail. It's important to learn and get up and try again, and, um, that just the act of engaging in that is so important.
So, Thinking Machines emphasizes AI that works collaboratively with people. I've heard you talk about that before. Help us understand in more detail what that collaboration between people and increasingly capable AI looks like. I think we're all kind of grappling with that right now, with how we think about agents, how we think about that use together. Tell us a little more so we're better informed.
So, uh, it's not that we have figured it out. Uh, we've shared some of the initial ideas and research aspirations and where we'd like to take things, but I think it's actually a very hard problem. And the way that I like to think about collaboration between humans and AI is, uh, much like collaboration within a team.
Think of a very fruitful collaboration with a colleague or a teammate. You're able to work in parallel and then come back together. Um, and a lot of collaboration happens in real time, and there is a lot of information that is shared in that real-time interaction. Some of it is like even when someone is silent, that is information, right? Um, and like the interruptions and being able to advance someone's thought. Uh, so that's very different from the way that we're currently collaborating with AI systems, which is through prompt boxes.
But being able to expand to all the modalities and having real-time interaction provides a lot of information about how we work, how we think, how we perceive information, uh, that can make AI a far better collaborator to us than it is today. Today's systems are very optimized for autonomy, and autonomy is good. We need it, uh, especially for tasks where AI systems get very strong and reliable. We will definitely want them to continue to do things autonomously.
But we also need to develop and measure, uh, collaboration between AI systems and ourselves, and measuring the outcomes, having some sort of benchmark, and making this something that we care about as a field. Um, and so part of the collaboration is autonomy, but it's also proactivity, right? And so we definitely want AI systems that understand what we want and can be proactive, much like our teammates or colleagues can be.
Uh, but again, that requires a level of trust, and, uh, it requires understanding, you know, when to delegate things. It requires more reliability and so forth. Uh, but I do think productivity is something — productivity and more reliability is something that we will be seeing in the AI systems soon, and I think this will enable far richer and more fruitful collaboration.