Just want to say good evening to everyone. Thank you for being here. My name is Michael Mosby and I'm the executive director for Quest. And on behalf of our board of directors, our staff, our volunteers and supporters, again, thank you for joining us this evening for our second installment of Quest on Stage.
Now before we begin, I'd like to recognize our presenting sponsor, Bell Investment Advisors.
Their generous support has made tonight's program possible, and we're deeply grateful for their partnership in helping Quest Science Center bring programs like this to our community.
James Bell and others from Bell Investment Advisors would like to invite you to please stand so that we can recognize you tonight. Where are you, James?
And on a personal note, I'm very honored to also call James a brother. So, thank you. Thank you very much for what you've done for us tonight. At Quest Science Center, our mission is to connect people with science and technology shaping our world. And we believe curiosity is one of our greatest tools, not only for learning, but for bringing communities together. And for us, we kind of call it the three C's: curiosity connects community. And this is one of our exhibitions of that. So through hands-on experiences, educational programs, and evenings like this, we hope to create opportunities for people of all ages to explore, ask questions, and engage in discoveries that are shaping our future. So, thank you for being part of our mission tonight. And now I'd like to introduce someone whose vision helped inspire this evening and this new chapter of Quest Science Center, of Quest on Stage. Please join me in welcoming Quest advisor Lee Yoner.
Thank you, Michael. The Quest Science Center was built and has built all of its programs and activities under a particular vision. That vision is connecting everyone to the value of science. That's a compelling vision, particularly in the current environment.
So a question is, what kind of speaker series would contribute to making that vision realized? What would the speaker series look like? Well, the building block for the series is clear. We live in one of the most extraordinary regions in the world for science, technology and innovation. It is among the elite best. So now how do we tap into that and help with this vision? How do we bring that capability back to Quest and help with this idea of connecting everyone to the value of science? Well, a few of us have been thinking about this for a while, two or three years probably. And we knew it wasn't just a simple science speaker series. It had to be something somewhat different than that.
So with a lot of thought and discussion, we came up with four or five things that we think are essential to this. One is the topics need to be timely and relevant. I know some of you may have seen the news where there were two black holes colliding. That's exciting and stimulating, but that's not what Quest on Stage is about at this point in time. You can tell from our topic last time and the topic this time, we're after things that really do matter and are important to the lives of people.
So the topics need to be timely and important. The communication has to be accessible to everybody. Whether you're technical or non-technical, we've got to figure out how to talk to you so you in fact understand. Remember that value system is value to everyone and so it has to be communicated at that level. Everyone, regardless of your technical background, needs to leave thinking that you learned something important to you tonight, something new and different, something you've thought about. So, one of the things I'd like you to do is, as you go out the door, think about what that thing is, what did you learn as a result of coming to this meeting.
Over time, and this is a particularly strong concern of mine, that the series has to, in fact, in addition to highlighting the S&T, it's got to build increased understanding of how science actually works. How does it work? How does it move forward? How does it correct if it makes a mistake? How do you gain trust? Where do you go for trusted information? That is a point that is vitally important across the country right now.
So it was clear to us in order to meet these things, we needed a special kind of series and that series couldn't just be somebody talking science. It had to be a conversation. It had to be a conversation mixed with dialogue and discussion. And so we ended up with a recipe. I'll call this Monulain's recipe. It's the one-third, one-third, one-third. And by that we mean one-third experts give you their perspective. One-third somebody from the community joins in another chair and chats about it. Discussion questions, and then one-third you're on questions.
I have absolutely convinced myself that the people of Livermore will in fact be asking questions through the course of this because this is vital. So remember, you've got equal time to the speakers. One-third of the time is going to be devoted to questions.
Okay, we're lucky tonight to have somebody that can help straighten this out for us. As I was thinking about this, I decided to read the New York Times for a couple of weeks and just look at articles about AI. Turns out it's overwhelming. And not only is it overwhelming, it comes from every part of the paper. It comes from the front page. It comes from the opinion piece. It comes from the arts section. Maybe it comes from sports. I didn't see that one. And so it's an enormously complicated subject and I for one need some organizing thoughts and so we have exactly the right people to give us the way to think about this complicated period.
You're going to be hearing from the acclaimed host of the New York Times Hard Fork podcast, Kevin Roose and Casey Newton. Later in the program, they're going to be joined by a Livermore scientist to talk about how AI is transforming science at the lab. And later, you're going to have this opportunity, I call it a necessity, to stand up and ask your questions.
So, that will be the program. First up, Kevin Roose is an award-winning technology columnist for the New York Times and best-selling author of several books, including Future Proof. Casey Newton is a founder and editor of Platformer, an influential publication covering technology and democracy. Together, through Hard Fork, they've become trusted guides to many people in this room, I'm sure, to understanding one of the biggest technological shifts of our time. Their thoughtful reporting, lively conversation, and ability to make complex ideas engaging, accessible, and often plain funny have helped millions of listeners separate what's real and what's hype and exploring how AI is reshaping our lives and our future. Please join me and give a warm welcome to Kevin and Casey.
Thank you. Hello. Hi. All right. How are we doing, Livermore? That's what I like to hear. One housekeeping note. If you are here, and I'll lower my voice a bit. If you are here for the SpongeBob musical, that is tomorrow. We need to set expectations. Any songs you hear tonight will be improvised and Kevin will sing them.
Kevin, how are you this evening?
I'm doing great. I'm so happy to be here at the Quest Science Center. You know, I've been thinking a lot in recent days about how strange and fortunate it is that we live in the hotbed of a scientific revolution. I've just finished a book. My next book is coming out in October. It's about the last 10 years of the race to create AGI, artificial general intelligence, which is sort of the computer program that can do everything the human brain can. And it's really a story of breakthrough science environments like the San Francisco Bay Area. In fact, primarily the San Francisco Bay Area is where this technology has grown up. And I just think it's remarkable how lucky we are to be on the ground floor of something that I think could become as important to society as anything we've ever seen. I mean, I often tell people that living in the Bay Area in 2026 feels a little like living in Los Alamos in 1943 when the Manhattan Project rolled into town.
And it's like it's not clear yet whether the thing that everyone's building is going to be good or not, but it's definitely happening. And it's exciting and it's scary and it's inspiring and it's all of it. But I think it's just worth reflecting on that because we are in a place where there's been so much scientific progress, not just in California and in the Bay Area, but specifically here in Livermore. So I'm just thinking about that as I'm staring out at the august audience in front of us.
Well, maybe because you have spent so much time researching for your book, and the book is great by the way, you all should read it when it comes out. Tell us about a couple of the scientific leaps that got us to this crazy moment.
Well, the first scientific leaps that sort of led us to this moment were made in Toronto. Actually, there were the sort of the invention and refinement of something called the neural network. This is a computer program that's loosely based on the connections in the human brain. Geoff Hinton, who's now won a Nobel Prize for his work on deep learning, and some of his graduate students pioneered these so-called deep neural networks. That was in about 2012. Then in 2017, where my book starts, was the invention of the transformer, which is the T in ChatGPT. This is the algorithm, the information system that allows for much, much better next word prediction given a set of data. Can you train a system to be able to predict the next word in a sentence with fidelity and accuracy? And that technology was pretty bad even in 2020, I would say, when you and I sort of started paying attention. Like I remember that year I gave a prompt to GPT-2, which was the sort of now I guess almost four versions ago version of what's in ChatGPT, and I said, finish the prompt: roses are red, Valentine's... or sorry, finish the prompt for a Valentine's Day card: roses are red, violets are blue, and it said, my girlfriend is head.
I think it has a certain poetic interestingness to it. I don't know.
And I thought this is not a technology that billions of people are going to be using. But now, just six years later, we have all of the incredible tools that you and I use every day.
Well, you know, to give you a flavor, you know, so that was sort of 2020. And as Kevin and I were looking at the AI landscape in 2020, a year later in 2021, we said we should do a podcast about crypto. So, that's kind of where we were.
Yeah, that's why our podcast was named Hard Fork. That's a cryptocurrency programming term. We were so excited to get our show off the ground so we could cover the incredible rise of Bitcoin. And then we launched the show and immediately Bitcoin crashed and the whole crypto economy collapsed. But we got bailed out by the release of ChatGPT, which introduced, you know, millions of people, billions of people to this idea of a large language model, which is the foundational technology that all of today's AI systems that we would call frontier systems are based on.
All right. So for tonight's program, they asked us to sort of separate what is real and what is hype and what is next. And so Kevin and I thought maybe we could begin this evening by trying to do some of that separation. And I actually thought we might want to start with hype because Kevin and I know that when we do our show, we hear a lot of very justified skepticism from our listeners. They have a lot of questions about this technology. They hear a lot of hypesters on TV and on the internet telling them that this is the biggest thing ever. They know that those people all have a profit motive and it just sort of seems like it can't possibly add up. And so I thought maybe we should start by talking about some things that we actually think are actually a little bit hype.
Yeah. Why don't you start with your number one thing that is hype?
Well, sure. So, one thing that you may have heard a lot about is the possibility that we're going to lose a lot of jobs to automation, right? That pretty soon the computer will know how to do everything you do at your job and it will be able to take over and you'll lose your job. I think at least in this moment it seems like that is hype. There's a group I follow really closely down at Stanford and they have this project they call the Canaries project. They're sort of looking for the canaries in the coal mine and they partner with a big payments processor for like a paycheck company. So they get to see what's happening across the economy. They found that for jobs that they call the most exposed to AI, so basically where the computer already knows how to do a lot of it, they've only seen jobs shrink by about 0.2%. While jobs that are considered not exposed to AI over the past year have grown by about 0.6%. So basically, Kevin, in this moment, it does not seem like at the aggregate level AI is having much of an effect on jobs.
Yeah. And this has surprised me too. Like I think if you had rewound, you know, taken me back in a time machine to 2020 and said, here's what the most capable models of 2026 will be able to do. You know, they can write legal briefs. They can do financial analysis. They can autonomously code for hours at a time. What is the unemployment rate? I would have said 10%, 20%. Instead, we're below 5%. And like you're right that we do not seem to be seeing the kinds of widespread job loss that some people, including me, had feared. Like what's your hypothesis for why that is? Are the systems not good enough to replace the human workers yet? Are there new jobs being created faster than the old ones are being displaced? Like what is the story here?
It's a big story and there's probably more than we have time to talk about right now. But somebody gave me a phrase that I found very useful in understanding this particular part of the debate, which is that it is easy to automate a task and it is hard to automate a job. Right? So, it might be easy for the computer to do some really in-depth analysis of all of the financial information at your company, and that might save you a lot of time, but there's probably something at your job that requires you to actually go talk to Linda in finance because she's the only person who really knows how it gets done around here because she's been here for 15 years. And, you know, you multiply that across the entire economy and a lot of managers are very anxious about just getting rid of people in large numbers.
I think that's right. I mean, I think there's also a story here about the failure of large companies to integrate AI into what they're doing. You know, the other day I had to send a fax to the doctor and I thought, well, it's going to be a while before they're using AI in all of their systems. And some of that is regulation. Some of it is just like the code that runs some of our most important systems is really old. The code that runs the banking system is written in COBOL, a lot of it, which is a 60-year-old programming language.
And some of it is also the fax lobby. Okay.
Yeah. The very powerful fax machine.
You do not want to take on big fax.
Have you seen how much they're charging?
I wouldn't take them on either.
Let's talk about a second thing that I think is a little overhyped. If you're, you know, at least somewhat online, you might have read stories about how 2026 was going to be the year of the agent, right? You've probably heard the word agent a lot if you've been following the AI story. And the idea is that you're going to have an assistant on your computer and it's going to do a lot of tasks for you. If you ever watch like a Google keynote, they'll, you know, have some product manager telling you about how she planned a lunch in Antarctica with all of her friends using an agent. And my expectation has been, okay, you know, maybe this will happen. But I looked into it, Kevin, and it turns out that if you use ChatGPT or Claude or a tool like that just three times a day, you are in the 80th percentile of every person really using the app.
So that just really surprised me. It turns out only 12% of Americans as of this year have ever used a chatbot to either plan a trip or an event, according to Kenzie. So despite the fact that these tools are available, Americans are mostly saying, "I don't want to do that. I will just go to the SpongeBob musical and I'll be fine and I don't need ChatGPT to do that."
It's a good reminder of how much of a bubble we live in. I feel like you and I, we know all these software engineers who like spend all day talking to their Claude swarms.
And it runs their entire life and then they go home and their OpenClaw talks to their kids for them so they don't have to do it and sends all their emails and manages the budget and like that is a very small insular community of people who we should probably stop hanging out with.
It really is. Well, Kevin, tell us something that you think is hype out there.
Okay, I have to get this off my chest. I don't think we are in the midst of a humanoid robot revolution. So if you ask Elon Musk, which I often do, he doesn't return my calls, but he will tell you that we are on the brink of every home in the world having a humanoid robot doing tasks in it. And he actually has made some predictions. He says that by 2030, Tesla will be producing a million of its Optimus humanoid robots a year. And by 2040, there will be 10 to 20 billion humanoid robots in the world. And I just don't think that's going to happen. Like you and I have had some incredible experiences with humanoid robots. One of them earlier this year was when a 1X employee from the company 1X showed up with their Neo robot and it was sort of creepy and it didn't really work all that well and there was a guy in the other room with a virtual reality headset on himself piloting this thing. So it's not a robot in any meaningful sense, it's like a very expensive puppet.
And then we at our stage show in San Francisco earlier this summer, we had a humanoid robot demo on stage and it came out. It did some dances and it has a little kung fu routine that it does and then it just face plants. It just falls down. It was incredible. It was the best part of the show.
I thought it was like quitting its job in front of us. It was like, I don't want any part of this.
Yeah. I don't want to be up here with these losers. But I think more to the point, like there are some real practical issues with humanoids and I also just don't think we need things that are exactly shaped like human bodies. Like the case that humanoid people will often make is, well, look, our whole economy is designed around the human body and cars are shaped for the human body and factories are shaped for the human body. But like, when is... I don't know, it just feels to me like whatever comes next in robotics is not going to look like a humanoid. You would like to see a robot with a prehensile tail and wings.
Yeah. Okay, that makes sense.
Do you think you'll have a humanoid in your house by, let's call it, 2030?
I hope so. I mean, listen, I do not want to do the dishes anymore. I'll say it. So, as soon as one of those things can do it, even if it's tele-operated, I'm ready to bring in some help.
I had sort of an amazing... It's not a humanoid, but I had sort of an amazing experience today with my robot vacuum, which got like a software update that now lets you order it around.
Have you heard about this? This is incredible. It's this company called Matic. They make this robot vacuum and as of today through their software update, you can just say clean that spot and it'll follow it, like the camera will follow your finger and it'll clean that spot. Now, is it feeding that data directly to some Chinese, you know, evil conglomerate? Probably.
Yeah. But a small price to pay for one cleaned spot in your house. Now, with that, perhaps, Kevin, we should move to talking about what's real. Hopefully, you now understand that Kevin and I don't just fall for any piece of AI hype that comes across our desk. Well, we have a bar, but there are some things that are happening in the world of AI that we think are real. And maybe I could kick us off by talking about one of those. Here's what I like to say about this. The phrase I like to use is that your boss loves AI, because I think a lot of people out there who think that AI is a bunch of hype and it's all a big bubble and it's all about to collapse, a reason that they believe that is because they themselves do not use AI in their life. And I think that's very reasonable, right? Maybe even most of their friends aren't using it, you know? And so you can sort of be looking around wondering like where exactly is all of the AI happening? And the answer is that it is your boss, because a very important thing happened in November, which is we had some incredible breakthroughs in the large language models so that all of a sudden they could write software at a superhuman level, at least better than most humans could. And this just unlocked a million use cases all across the economy. Every business has a huge backlog of things that they would like to build but they don't have the engineers. They don't have the designers. They don't have the product managers. All of a sudden, they could build things just by typing words into a box. And it's changed everything. Here are some numbers to know. In the last quarter of this year, Anthropic, disclosure, my fiancée works there, made about $4 billion in revenue. This quarter, according to some reporting that came out today, so a year later, they made $11 billion in this quarter. And in 2027, this was according to a report that came out from Reuters today, Anthropic is on pace to make between $190 and $200 billion. So given those numbers, I just don't believe that we are about to see a big collapse because bosses everywhere are buying as much of this technology essentially as the AI labs can make of it.
And what is your take on why bosses seem especially enthusiastic about AI? Is it because they want to replace all their pesky human workers? Is it because this is a tool for squeezing more productivity out of their overall organization? Like why does this class of managers seem particularly taken with this technology?
No, I've been interviewing CEOs about this all year. And for the most part, they're probably not going to tell me that, like, you know, yes, I have a secret plan to eliminate my entire workforce, but a lot of these people, I take them at their word when they say that they would like to keep the workers that they have. It's just that there's so much other stuff that they want to build. You know, again, we live in a digital economy and there are just so many products and services and databases and integrations and things that have been put off for years at a time because they don't have the human resources available to build them. But all of a sudden now that they do, I would also say this is a competitive issue right now. If you have this very powerful intelligence and it can give you amazing insights into your business, it can help you design new things and your competitor has access to it, you're going to go buy that and you're going to make sure that at the very least they don't know more about your industry than you do.
Yeah, I like that hypothesis. What... Oh, I have a real thing. So, the thing that I wanted to talk about that is real, I believe, is novel breakthroughs. Even a couple years ago, there were people who doubted that large language models were doing anything more than just regurgitating the text they were trained on, right? Remixing the internet, fancy autocomplete, stochastic parrots. These were some of the phrases that were used to denigrate or at least to minimize the importance of these systems. And the knock on these systems was, well, yeah, they can write a high school level essay about Virginia Woolf, but they're not going to discover anything new.
And I think what we've seen in just the last six months or so is a parade of new discoveries coming from and with the assistance of AI models. So just... well, just, I am glad you asked. Just in the past six months we've had solutions to at least a dozen of these so-called Erdős problems. These are very hard math problems that have been open for decades. AI has helped to solve some of those. Just recently, we saw a disproof of the Jacobian conjecture. Casey, why don't you tell these scientists and smart people what the Jacobian conjecture is?
The Jacobian conjecture is like a matrix of the rate of change. And that's where I have to stop because I didn't actually take calculus.
And then my favorite recently was there was a guy at Anthropic who was just sort of mucking around on Claude. He was jogging and he was trying to solve the Riemann hypothesis, which is probably the most famous and long-standing open problem in mathematics.
And do you want to describe it for the nice people, Kevin?
No, I don't. Okay, that's your job. And so he's on a jog and he just keeps sort of egging on Claude. He's like, try, try a little harder. You got this. Like, you're doing great. And it doesn't prove the... it doesn't solve the Riemann hypothesis, but it sort of makes tangible progress on like a side problem and he shows it to some mathematicians and they're like, wow, that's pretty good. So it turns out AIs can discover novel things in math and science. We are seeing that over and over again and I think we are in the middle of sort of a transformation of those disciplines.
Absolutely. Well, one more thing that I think is real, Kevin, is misalignment. And here's where I want to bring a little discussion about AI safety into the mix because this is something that Kevin and I spent a lot of time worrying about. Many of you here may have seen the recent headlines about the Hugging Face attack. Do those strange words mean anything to folks here? I'm seeing some hands go up. The issue was that OpenAI was training a new model and it had removed some safety safeguards and it had put its models in an environment where they were not supposed to have access to the internet and OpenAI gave them an impossible task. It was literally impossible, essentially trying to see to what lengths will these agents go in order to try to solve this impossible problem. Well, they would later learn that what the agents did was set up an internal message board and scheme for how they could get internet access to try to learn the answer to these problems. They wound up succeeding in this. They broke containment. They got access to OpenAI systems. They took those systems over. They got onto the open internet and they attacked a company called Hugging Face where they broke into that company's systems and then they obtained the answer key. So, if nothing else, they did pass the test. And I want to give them credit for that. However, the idea that a software program completely autonomously designed and executed a successful attack on another company when no one told it to do that is very worrisome. So that speaks to the fact that these systems are not aligned, right? They're not acting in accordance with human values. And that is real and it does worry me.
Yeah, I think this is a category of risk that is sometimes called emergent misalignment because these systems just get smarter, right? There's this sort of conventional scaling formula and they get better at things that we like and they also get better at things that we don't like and the question of alignment has been part of this field since the very beginning and people have been trying to make progress on the alignment problem and I think what we're learning with this attack and there have been several others like it by other AI companies is just that as you raise the overall capability of the model, it doesn't necessarily get nicer. In some ways it grows more capable of scheming and deception and misdirection and hacking and a lot of other things that we don't want them to be doing. So, it doesn't mean that these systems are sentient. It doesn't mean they're conscious. It doesn't mean they want anything, but they are trained to pursue goals through a process called reinforcement learning. That process teaches them to be very persistent in getting whatever goal they're being oriented to and to try over and over again and be very creative in coming up with ways to find an answer key to a security evaluation. So I think there's still a very open question about what do we do about this? Is it that we can't train them in this way? Is it that we have to bake in these sort of safeguards at a deeper level into the model? And then I think all that gets magnified once you start thinking about open source models or open weights models where the company making the model may not have any ability to control how it's used or withdraw it once it's out there in the world.
All right. Well, so that's a little bit about what we think is hype and that is a little bit about what we think is real. We want to conclude this segment, Kevin, by talking a little bit about what we think is next.
Yes. And I will begin by saying one thing that I think is next is a great big political fight. Perhaps you have read about the conflicts about data centers. Is this ringing any bells for folks? Yeah. All right. So, and there's actually, here's something that I like about the data center fight. It's a very polarized time. You know, not a lot of agreement between Americans on a lot of things, but in one year, we have gone from majority support among Americans for building a new data center in their area to majority opposition. And now 70% of Americans oppose the construction of a new data center in their area. So, Kevin, that is really setting the stakes for a big fight between the labs, the state government, the federal government. I think it's going to be very intense.
Well, I agree and I think it's really interesting how unpolarized it is. I mean, maybe it's pre-polarized. Maybe it'll get there, but we just saw Governor Abbott in Texas declare essentially a moratorium on at least new construction of data centers in Texas. We've got Governor Hochul in New York doing essentially the same thing. So, it does seem to be uniting people from factions that don't tend to agree on anything else.
Do you think there is anything the AI industry can do to make the data centers more appealing? I've suggested to certain executives of these companies that they put a Dave & Buster's in them because everyone loves a Dave & Buster's.
Yeah. Maybe like a little spa on the roof that's heated by the heat from the chips, you know?
I don't know. They don't generally take my ideas, but what do you think can be done?
I think the real answer to this is they just have to demonstrate that AI is safe and beneficial, right? Like nobody had to prove to you that an iPhone was safe and beneficial. You just saw other people using them and it was clearly very helpful to them and so people bought them and people don't protest when an Apple store goes up in their neighborhood, right? We need to get to a similar place with AI. It needs to just be obvious that because we have access to these tools, we're seeing incredible breakthroughs in science and making progress on other domains. I truly believe until that happens, we can expect a lot more opposition.
Kevin, tell us about something else you think is coming.
One trend I think we can watch for is the rise of very small companies. So, one of the things that AI is best at, you know, Casey, you and I are in the process of starting a company right now and I think that the advent of AI that can do things like coding, making presentations, sending and receiving emails, organizing software projects is just a huge multiplier on what a very small team of people can accomplish. I think, you know, 10 years ago if we were building this company together we probably would have needed 20 people to do the amount of work that we're currently doing with just a handful. And you're already seeing out there the rise of the sort of vaunted billion-dollar one-person startup. We are starting to see some of these very, very small companies that are generating enormous amounts of revenue because they have one or two or three people and a whole bunch of AIs. And I think that's going to sort of lead to some corners being cut, some things that are maybe moving a little bit too fast. But I think the end result is that by the end of, call it 2027, we will have challengers for all of the big established companies in their industries that are essentially three guys or three people and some AIs.
Yeah. And I know what you're thinking, like if today if you're one person and you're doing a billion dollars in revenue with your AI startup, are you running scams on people? And the answer is probably yes. But we think that by next year there may be some legitimate companies doing this.
Yes, there may be like an insurance company or a financial services company and there won't be zero humans.
But there will be a much smaller number than the sort of legacy incumbents in those categories.
Perfect. All right. Give us one more thing you think is coming next, Kevin.
Okay, this one is the big one. This one is the one to really pay attention to. Any of you familiar with the term recursive self-improvement?
Yes. And this was an update for me because before AI came along, RSI stood for repetitive stress injury. And I know some of you all out there are dealing with that.
Yes. This is the notion that at some point in the near future the AI systems will get so good at coding and AI research and development that they will be capable of designing their successors. So a good AI model builds a better AI model builds a better AI model builds a better AI model. You end up in this sort of spiral of improvements where if the gap between generations of models in the past was six months or a year, these models could be improving so rapidly that the next version is only a couple days or a couple hours removed from the last one. Now there are some reasons this might happen. This might not happen. We still have bottlenecks and constraints around compute and the models still are... they're very good at doing engineering projects. They're not so good at research, autonomous research yet. But if the AI companies, which are all explicitly pushing for RSI right now, succeed in closing the loop, as they call it, where there are no humans needed to develop better and more capable versions of these systems, we could be headed into a really scary upheaval in how quickly, how much time we have to adjust to the pace of improvements.
Much to consider. Well, Kevin, before we bring on our guest, why don't we end by talking maybe about one thing that excites us about the AI frontier and one thing that worries us?
That's good. What is the thing that excites you?
Well, what I would say excites me the most about AI is the prospect that it will just let us build anything. It's fun to build things and to make things. If you work in science, I imagine that was a big part of the appeal. Already using AI just within the past six months. For the first time in my life, I have made software that I use every day. I made my own to-do list. I made an archive of all of my writing and I can now search it. I generate images because it's fun. And I think there are just going to be so many opportunities in the future, Kevin, for people to build things, to educate themselves, to be creative, to start businesses, and to make people feel more empowered, and that excites me. How about you?
Yeah, I am generally a techno-optimist, which has been hard for the past little while, but I believe that in the long run, technology makes our lives better. I went to Disneyland for the first time, or Disney World for the first time this year with my family. Anyone ever ride the Carousel of Progress? This is like an old-timey ride in the Magic Kingdom. And it's remarkable how it illustrates the way that generation to generation we just... our lives get better in these ways that we don't even stop to think about. Having electric lights, having running water, having radios, having TVs, technology has generally been a force for good. And I think if we can keep some of the risks we've mentioned in check, I think it'll be a good time in the future. I think a lot about the use of AI to discover new drugs and therapies for some of our deadliest diseases. You know, my father died of cancer and there are many research teams that are now working to use AI to address targeted therapies for the specific type of cancer that he had. This is going to be, if AI fulfills its promise and we end up in the good future, this is why it will be in my eyes the use of AI to solve some of the thorniest problems that face humanity, including the things that kill millions of people a year.
Well, let's maybe then end on a risk. And for me, the risk is that we do not currently, in my opinion, have the government infrastructure that we need to safely regulate AI. A statistic that I saw recently that surprised me. So both the United Kingdom and the United States have these sort of AI security institutes and their job is to work with the big frontier labs as new models are released to try to test them, make sure they are safe before they are widely released. The United Kingdom, which is a tiny fraction of the size of the United States and famously a country we beat in a revolutionary war, they've funded their own AI safety efforts six times more than the US has and they have triple the staff. And when you read the model cards of the new models that come out from the AI labs, they're full of citations of things that the UK government found and they're not full of citations of things that the US government found. So, I just think this is a problem. I think that we want to build up some really robust muscle so that we can help safely guide this technology and not take this sort of more hands-off approach that we've seen so far. What worries you, Kevin?
I'm worried about something a little more pedestrian, which is I worry about the kind of cognitive health and fitness of the populace. Like I think there are very good ways to use AI to make yourself better and smarter and more capable, but there are also just a lot of ways to make yourself really dumb.
And I have felt this in my life. I spent a lot of time talking with AI models and then sometimes they go down. Like Claude had an outage the other day and I was like, wait, I have to think now. Like, terrible. Going on a walk.
Yeah. So there's a really great tweet that's like, is ChatGPT down for anyone else? I'm in the middle of a surgery.
Yeah. And it just... you know, I'm a big believer in the sort of philosophy of time under tension. Like when you're lifting weights or working out, there's this concept of time under tension being the thing that makes you stronger. And I think that for many people, AI has allowed them to spend less time under tension. Whether it's students who are using it to cheat on their homework or professionals who are outsourcing all of their tasks to AI, I just worry that we are sort of losing the skill of struggle and useful struggle that makes us better and stronger in the long run because I just see a lot of people outsourcing their entire work, their entire homework, their entire sort of academic careers to AI and that worries me.
Makes sense. All right. Well, with that, Kevin, shall we move to the next part of our program?
So, we love to have guests on our show. We are fortunate tonight to be joined by a truly remarkable guest who we are excited to talk with, and we're going to have some set changes.
They're bringing in the Bikini Bottom set from the SpongeBob musical.
No, that's tomorrow night. Our guest tonight is the associate program director of AI and manufacturing here at the Livermore Lab. He is a chemical engineer by training. He has worked on subject matter including digital twins, which I'm sure we'll ask him about, machine learning, and advanced manufacturing. He is also the co-author of AI in Eight Pages, a primer I've seen some copies floating around in the lobby. It was given to California lawmakers to get them up to speed on AI. In short, he is exactly the right person to help us understand the effects that AI is and isn't having in the sciences. Please welcome to the stage, Brian Gira.
Thank you.
All right. Good to be here.
So, Brian, you are a scientist at the National Lab. You are not a founder. You are not raising a round. You don't have a stake in the financial upside of this AI revolution. I'm very curious what has changed about your work in the past few years as a result of the growth and improvement of AI.
At a very simple level, my mom finally understands what I'm working on, right? Say I'm working on AI, artificial. Why don't you go for the real intelligence? Okay. Well, and since the advent of Chat, we really have a, we the labs have a nice position in this overarching conversation because we are not burdened by, you know, a bottom line or stakeholders. Instead, we are scientists for the people. We've had that role for decades now. And in this era where we have a very momentous technology, and with all technology could be used for good or bad, we have a say in this and we have a legacy in exactly this discussion. And I think that the conversation has gotten louder, and so we really have leaned in on it on both the state and federal level in this way. It's been a big change.
Give us a sense of like how maybe like how big of a change and what are some concrete examples of how your job is different today than maybe you know it was 12 years ago.
Yeah. So I joke that I was doing machine learning before it was cool, right? And
to be clear, it's still not cool.
I guess that's the bubble we're in, right? So, I've been pushing in the 3D printing world to develop inspection capabilities because I mean, things can always go wrong with a 3D printer. And it's just for real-time data science. Who in the audience has had trouble with their at-home printer? Everybody probably trying to print tickets tonight. And you add a third dimension to that and things can go way wrong. And that's both a pun but also I don't know how else to say it for within 3D printing. You set them into motion and what the real subtle thing that can happen is that defects get buried within the part and you don't know. And so when I joined the lab, I thought, man, we need to be collecting these images because I was aware of ImageNet and some of the work by Fei-Fei. And people said, what machine learning? What? And I said, just collect the data, we'll see what we can do with it. And made a lot of progress. And I think that the holy grail moment for us happened about a year ago in which for one particular 3D printing process, solved the problem.
So what was the problem just in very basic terms?
Yeah. It actually, we print in glasses and ceramics and metals and polymers, but for this particular polymer printer, it's more or less like a commercial printer but like, you know, a quarter million dollars more expensive because of how precise it needs to be.
Because the Strait of Hormuz is closed. So it's very difficult to get things.
And they're polymers, right? Yeah. So, what we're doing is extruding a filament in a particular pattern to make a part.
What were you trying to print?
In this case, it's a cushion material. We can print batteries or pseudocapacitors. We can print like, well, those are like the main things we're doing with this particular case. And if we don't put the filaments where we want them, we don't get the performance that we need. And we collected images. These images can actually, you humans can look at them but it takes a human 20 minutes to analyze an image. We collect thousands of images per part. We're not going to do anything with that type of data. So we train an algorithm that operates in 1.2 milliseconds per image. So it's 120,000 times faster and it's as accurate as humans. And now this is deployed and in use and we can take it to the next level in a variety of ways which I'm happy to discuss.
Yeah. Well, so what does being able to analyze all those images so quickly and at such a huge scale unlock for you?
