BREAKING NEWS: Hillary Clinton Leads Fireside Chat With Co-Founder Of LinkedIn Reid Hoffman

Please welcome to the stage Reed Hoffman and Secretary Hillary Rodham Clinton. I need to do much of an introduction of Reed Hoffman. Anybody here on LinkedIn? But in addition to that, he's done so many other great investments and founded companies now working in everything from, you know, biopharmaceuticals, using AI to, you know, helping us all understand more about the world we're entering. And Reed, really, I could not think of a better person to have on this stage who can help us chart what I call kind of the, you know, sensible understanding of where we are with AI, because we've all been reading a lot of, you know, very scary, you know, articles and statements. And we seem to, you know, just have the pendulum switch back and forth between the so-called accelerationists, you know, full speed ahead, don't ask any questions. The decelerationists, the risks are too great, stop right now. And I want to really get your take on this for our CGI community, because you have always put yourself in the optimism camp. And you are, you know, currently involved in scientific breakthroughs and trying to help solve, you know, health problems and diseases and the like. But how do we as lay people, you know, we're not in a frontier lab, but we're on the ground, we're in policy positions, we're in businesses, we're in academia and professions, we're in active civil society groups. How do we think about this? So give us the description of the moment we're in. And then if you could wave a magic wand, what would we do to try to maximize the positive and minimize the negative of what's coming at us with AI? So the moment we're in is we are in the process of creating amazing intelligence with the scale and cost of electricity. And you say, well, that sounds abstract, what is that useful for? And like to go super deep, one of the things is, if today you have a high medical condition, either as a doctor or as a patient, and you're not consulting a frontier model, Chachi, BD, Claude, as a second opinion, you're almost making medical malpractice, right? That's how important it is. I personally know three people whose lives were saved by doing the second opinion, right? And that's, it's not like I know that many people, right? I mean, it's not everyone on LinkedIn. And so, you know, part of the thing that I think is, is there's this amazing thing that can actually be completely transformative. And while, by the way, part of, you know, doing a post, there's going to be job transition. And, and, and that job change is going to be difficult, just like the Industrial Revolution, but time compressed and a bunch of other things. So really, really important issues to navigate that are super important. We don't have our current modern life without the Industrial Revolution, right? Everything comes from that. And so that's a similar thing on the other side. It's the reason why the journey is worth it. That's the reason why it's worth steering through this. And the, the problem with kind of the acceleration versus deacceleration discussion is that when you get to thinking about risks, AI is also part of the solution. So like there's this discussion of existential risk, like what happens with humanity? And they, you know, people say something like, well, can you guarantee me that there won't be killer robots? And you say, well, I can't guarantee that. Well, see, that's an existential risk. You should stop. QED, we're done, right? And you're like, look, if existential risk, like killer robots were the only existential risk, you'd have that. But existential risk is a portfolio. We have a whole bunch of existential risks, not just nuclear, which people tend to forget about, natural pandemic, manmade pandemic, asteroids. So you have to look at it as a portfolio. And actually AI, I think very easily can decrease the overall portfolio existential risk, even though you have killer robots. And this is the mistake that the pronouncers make because they go, oh, it's this one thing. And it's like, no, no, we, we are navigating a set of things. Like how do you, how do you stop pandemic? AI is actually the only thing. Cyber, you need to have more advanced models for the defense than the offense. So if you deaccelerate, you have a problem. And so that's, that's kind of a scenario. Now, if I could wave a wand, I do think that some of the things that are proposed are good ones. Kill switch is a good one. It's not a natural thing for a company to put in. So it's a good, good intervention to kind of say, hey, you have to have that. Having essentially third party auditors that companies are responsive to, it's actually something that I've been advocating for a few years. But by the way, now with what Dario and Sarah were saying, actually more deep technical auditors are very good because, you know, Deloitte doesn't really understand frontier models either. Right. Or I don't, wasn't meaning to pick on them, but your average accounting firm does not understand these things. And so, so those are good things to have happen because, you know, it's kind of, there's a natural thing in a commercial race to say, oh, I'll get to that later. It's like, no, no, get to that now. Right. As a way of doing it. So I think those are all good interventions. And then part of the, like when you get to the acceleration, deacceleration, but data centers, like I think Josh Shapiro at Pennsylvania is doing a very good job of this is you have to have the local community bought in. You have to have something that's valuable to the local community in a durable, sustaining way, not just the construction jobs or the happen thing. That's a must, right? They must be bought in, of course, but a global moratorium doesn't help your local community and make them have their self-determination and choices. Now, I think the states and the feds can go say, hey, let's set a floor by which local communities can't be auctioned against each other. It's almost like call it minimum wage and kind of like an environment, but also in benefits and stuff to say, oh, this is your minimum to getting in. So I think there's some role for them too, to make sure, because the natural thing in business is, oh, which of you will give me the better deal? And then you have an auction to, to a less good deal. Yes, exactly. So anyway, those are, so if I raise my, wave my wand, I would say, yes, kind of the kill switch and the safety plan and the, and essentially the auditing. And then kind of a, hey, have a, have a way that local communities can, can, can negotiate for an intense and enduring, probably economic benefit for themselves. Well, that's, that's interesting because we, we know that the auditors, at least were mentioned in Dario Amadei's pronouncement, his, his statement. But what you're saying is they need to be real. They don't mean that no, you know, just decorative auditors, but people who understand the technology and can, you know, really call an alarm if one is necessary. And similarly with the kill switch, because somebody said to me the other day, well, if you've got, you know, the agentic bots who are basically in control of a system, can they override the kill switch? So give some thought to make sure that you can actually do the kill switch. And then standards, I would say standards in a number of areas. You mentioned data centers, and I think there was a lot of presumption and even arrogance about the data centers. And I think a lot of communities, like you say, accounting firms may not understand auditing, communities didn't understand what they were signing on to. So having standards, and I think either a temporary suspension or moratorium until those standards can be worked out, or setting out a set of, you know, bars that have to be addressed and overcome. So I, I think all of that really makes sense. Of course, that requires that we have a government that can put that into place, set the standards, enforce whatever the, you know, provisions are, because if an auditor comes forth and says this needs to be stopped, who's going to implement it? Who's going to make it happen? So would you envision like some kind of, you know, AI agency, something in the federal government that we could turn to? Because you mentioned the Industrial Revolution. Well, a lot of our government came out of the last century, because as people saw the Industrial Revolution, they wanted to fix the problems, keep the benefits, wanted to, you know, improve labor conditions, wanted to protect people with their, you know, safety on the job, wanted to prevent predatory, you know, companies from, you know, really forcing others out of business and all the rest of it. So how would you see this as a regulatory outcome? So I think the important thing to start in this is say, what are the things we should be measuring so that we can respond to them quickly as they happen? So start with a, what are the possible wrongs and ills, and not trying to pre-plan it. It's a little bit like, say, for example, you're trying to drive from here to DC, and you say, I'm going to pre-plan my whole drive. I know where the potholes will be. I know where the accidents will be. He's like, you can't. And if you try to do that, you'll never get on the road. But you have some sense of, oh, if there's a traffic slowdown, maybe we'll take this detour. You're like, you have a sense of things and some sense of what you could do. But it's measuring that going into it. Because, like, for example, we know that there's going to be job transformation. But, like, what's the pacing? How does it work? Like, and most of the AI technologists think that once you build the technology, everyone's going to automatically adopt it tomorrow. And then the whole job force will change. Like, that's not how human organizations work. So there's some time. Like, one of the places where I had that dialogue was autonomous vehicles. I'm like, well, if you're going to create autonomous trucks, what happens to truck drivers? Well, if every single factory in the world started manufacturing autonomous trucks tomorrow, it's 10 years before half of the trucks are autonomous. So it's not lots of time. But you have time to measure. And you have time to respond. And it's like, OK, what should we do early? Let's measure that and navigate, whether it's jobs. Cyber, I think, is actually quite urgent. Bio can be very bad and is also urgent, but probably, hopefully, nothing's going to happen. But those are things to actually really dive into. Well, I want to drill down on that, because that's what I think about. I think about AI in the hands of the bad actors. And from my perspective, having read a lot of classified material about bad guys trying to get nuclear material, trying to get bioweaponry, that happened before AI. And they couldn't just dial in to AI and say, tell me how to make a nuclear suitcase bomb. Tell me how to create a pathogen that can wipe a small city. Sadly, that's becoming much more of a possibility unless we put these guardrails in very firmly soon. So when you think about the big threats, and I'm glad that you mentioned that AI can help us, but AI can also facilitate some of the cyber potential, bio potential, and sadly, even nuclear potential. Is there any alternative than other to try to forge international agreements? I wish that that were on the agenda of this meeting that's going to be happening between Trump and Xi. That's what they should be talking about. This new technology poses a threat to the control of the Chinese Communist Party, just like it poses threats to us in keeping our democracy going and protecting our people. So how would you think about the international agenda? Is that even possible, Reid, or is everybody going to retreat to their corners? Well, if we actually had someone who actually really was a dealmaker, then maybe. I'm a little skeptical. But the way you'd start is bio-risk and cyber-risk. Because if you can't do that, you can't do anything else, right? And the bio-risk one is such obvious. And I think legitimately the Chinese, who like global stability, have a shared interest here. And so it's doable, right? And I think it's very important. This is, again, part of the reason why pacing, not necessarily slowdown, is the defense is actually super important. Because humans, as you know from the classified materials, are capable of doing enormous bio stuff. AI can now help them, because it's like having a PhD who does that. And part of the reason why you want these kind of high-intelligence models, frontier models, to be in a kind of a measured, monitored environment is that you can't. The problem with open-weight models is, if they're in an unmeasured environment, you can get them to be badly helpful if you're trying to do that. So you want to see what the query stream looks like. Oh, that person's trying to figure out new versions of smallpox? What's going on there? Right? Let's go to that. And so you want a measured environment for that. I think that's part of the kind of regulation that placed open-weight models, that placed other things. But that should be on the agenda. Because once you get there, then you can broaden to what are the other disruptions that we should pay attention to. But you'd actually have to have a real dealmaker who had some intelligence on this. Yeah, right. Not even superintelligence. Regular would be fine. Look, I think that you've taken a very measured approach on the optimistic side. If Geoffrey Hinton or somebody else were sitting here, what would be their best argument against your position? So the key uncertainty is in what's being worked on primarily by Anthropic and OpenAI's recursive self-improvement, which is the AI making itself better. And the question is, can you guarantee the trajectory of that path if it's going very fast and we don't see what's going on? And the short answer right now is you cannot guarantee the trajectory of the path. There's some good ideas and things that probably will work and sort of, but the guarantee is not there. And that's part of the reason why Dario wrote his essay and, you know, Hinton, who's obviously a super smart Nobel Prize winner, et cetera, and the inventor of some of the techniques that have grown to this, are like, well, but shouldn't we be certain about that? And we'd like very little in life is certain. Like, for example, can we be certain there's not nuclear war in the next 20 years? No, we can't. So certainty is actually not the goal, but like high, high probability of getting it right is the kind of right thing. And so their counter argument would be is, well, you can't give me that recursive self-improvement yet. So why shouldn't we just slow down for it? My counter to that, so I try to make the argument as strongly stated as possible, is to say, well, we won't know how to navigate recursive self-improvement until we're actually doing it. So you can slow down on deployment versus development. You can still be doing that. And then you could be saying, what you must be doing is studying and building plans for how you create the alignment, the alignment to human interest in recursive self-improvement. And if it's not working, then you slow that down. Well, that that is a distinction I've not heard anybody else make, Reid, which is the development needs to go forward in order to better understand how to manage, control, or even eliminate further development prior to deployment. Yes. Now, is that something that a AI agency, the one we're postulating in some future federal government, is that something... One with regular intelligence, yes. Yes. Is that something that they could, you know, mandate under your idea of how to better control what we don't know until we have some, you know, path forward? Yes. And also, that's part of where the auditors tie in. Because part of the thing is, have the companies that are doing it, have the auditors, and then have... Because, you know, it's hard to get the right technical staff in government. You know, we've all been working on that and trying. So you can only usually have a couple, but have that dialogue. And it's like, okay, is it working? Okay, auditors, what do you see happening? And so then you have a very frank, honest conversation about, you know, is the pattern of maintaining alignment while you do recursive self-improvement, is that working so far? Yeah. And then the final question, I just want to make sure I understand what you said about the difference between, you know, open weight and not. Because the Chinese are obviously pursuing much, as far as we know, much more of an open system. Does that pose additional dangers or not? It does. And I'm a huge believer in open source and all the rest. I was on the board of Mozilla for 11 years. But the issue is, it's a dangerous piece of technology that if you can build it in an isolated environment, run in an isolated environment, what I mean by isolated environment is like, you know, say, one to three people in their garage. Then, like, for example, if I was trying to figure out how to do an interesting new form of cyber attack, I would take an open model, probably Chinese, because they're amongst the best open models, I'd put them on a computer in my garage, and I would retrain it and iterate it to a number of different attack vectors that then no one knows that they're coming, because they haven't been tried out before, and then deploy them at scale very quickly. That's not good from, you know, whether it's hospitals, infrastructure, you know, other kinds of things. So it's like, you have to say, no, no, we got to make sure that the models that can be played in those disconnected environments are substantially less capable than the ones we're using for defense. But that would be something we would raise in any kind of negotiation with the Chinese, presumably. A hundred percent. Well, finally, Reid, you know, we do think that there's going to be, no matter how it's done, a lot of money made in this field. And I'm hoping that there's some thought being given in Silicon Valley and other places as to how that money is going to be deployed, both to deal with some of the risks and some of the consequences, like unemployment over what period of time. Is there a conversation like that happening? So the short answer is very much yes. And there's been a bunch of dialogue, a bunch of discussion. I'm actually on various text threads and others with folks from Anthropic and OpenAI saying, how do we contribute? What do we do? One of the things I think the philanthropic community should understand is what they're focused on is not just how do we address the questions that AI brings up, and how do we make it all altogether more positive, which they are. In both companies, they think about this stuff a lot. That's part of the reason why they sometimes make bizarre pronouncements, because they're worried about it, and they don't understand how they're heard, right? So it's like, oh, they'll have a 25% white-collar bloodbath in a few years, and it's like, well, what people hear is, I'm going to get rich, and you're going to get out of work. It's like, that's not what they're trying to say, but they don't understand communications. And so, yes, they're trying to say, hey, this is going to change. You should get ready, right? And so, but they're very philanthropically motivated. Now, for the philanthropic community, they're also going to be thinking, how is AI part of the solution? So it isn't like, oh, come bring me your community college idea for retraining. I doubt they're going to fund a lot of that. But if you say, hey, this is how AI could actually really make a help in people finding, you know, new meaningful activity and work and everything else, then they're like, oh, that could be interesting as an example. Got it. Well, boy, this is such a thought-provoking conversation. I hope we can continue it going forward. But before I leave the stage, I want to recognize the 20th anniversary of Unitaid and congratulate Philippe Dunton. Where are you? Are you here? Please stand up. I mean, this is such a brilliant entrepreneurial innovator. Unitaid came up with an idea 20 years ago that enabled so much of the work that was done to provide life-saving medical interventions and drugs to people living with HIV, keeping people alive, and you've gone on to do so much else. And I want to just give you a kind of overview of what Unitaid is and why we're so proud to be a partner here at CGI with Unitaid. So please watch the video.