I personally know three people whose lives have been saved by ChatGPT.
It's amazing that this is an industry where we have to remind ourselves that this is for human beings.
Yes. You should never connect AI or any automated system to the nuclear stuff, and you should keep the nuclear stuff disconnected from the internet. Not just because of AI, but because of all the human beings. You're like, I'm a middle-of-the-road kind of guy.
Yes. I get hit by cars going both directions. I like weaving in the middle of the road. Reid Hoffman, welcome to the Newcomer Podcast. Thrilled to have you in person. You are fresh off a Substack post laying out a sort of middle ground on artificial intelligence, and obviously there's lots to talk about there. I want to just give you a second to give the high-level view. I feel like you resist — you're not claiming one of the, you know, monikers like I'm an EA, I'm an accelerationist, I'm a D. Do you have a brand for this middle-of-the-road take, or what is it? At least give us sort of the argument behind it.
Well, maybe a smart accelerationist, a pro-steering versus accelerating or de-accelerating. And you know, generally speaking, I think I tend to think of myself as a techno-humanist, which is how do you have technology enhance humanity, which is I think where we mostly orient towards most technology going, and we just need to keep steering in that direction. It's amazing that this is an industry where we have to remind ourselves that this is for human beings.
Feels like that gets lost sometimes, and then there has to be sort of a corrective of, oh, who are we building this for? Not the machines themselves. What is the main thing you're calling for here? Because there's sort of like the data center piece of it. There's like access to AI.
Yep. So the primary thing is kind of two things that need to go together. One, which is like why is it important that societies — and let's talk about the US, America needs to win at AI because it is the next industrial revolution, and industrial revolution is the thing that determines future prosperity, just like the previous industrial revolution did. So if you go, we want to opt out of the industrial revolution, that's what crushed a whole number of societies for centuries, you know, whether it's India, you know, China, a bunch of others. Like no, no, that you actually have to do. So succeeding at it really matters. But then simultaneously and at the same time, more than 60% of Americans are having a very difficult time, right?
And you know, there's questions around like, okay, the enormous gas price rises, the cost of beef, you know, kind of like how many jobs you need to work, and it's the responsibility of the entire American society to try to make that, because in a sense our ethos as a society is a middle-class ethos to make that work better. And so I think AI can — that doesn't mean there won't be, and this is part of the thing that was the backdrop of addressing in the essay, a lot of significant issues around work change, and people are legitimately concerned about like, well, what can you tell me about that? When is it going to happen? How is it going to be? And by the way, no one actually in fact really knows. Neither the people who go, oh my god, it's going to be huge in a year or two, or the people who are like, oh no problem, nothing this year. Actually nobody really knows, and most people in the nobody-really-knows get nervous about that, although that's I think the kind of the nobody-really-knows part of it means that it's like going down a road. It's like if you're planning a highway trip, do you know there's an accident on the road already? Do you know that there's not? Usually, you know, you have the capability to navigate it. You set yourself up. This could come, this could come, I'll do that. But you don't just say, I need to know everything ahead of me or I will not drive anywhere.
Exactly. So it's kind of like — so my thing is get on the road and drive and keep driving. And by the way, sometimes you hit a curve or traffic, you slow down a little bit for that, but it's a forward-motion orientation.
Some people think we're at the curve now though. I guess to ask that directly, do you worry like, should we be pacing? Should we be halting? Or you think in terms of what you've seen,
Keep driving.
So definitely keep driving. But by the way, this is the other nuance in the pacing and so forth is like, for example, there's a pacing of building and there's a pacing of deploying. And I actually think when I look at like Dario's call and you know, Sam and I go, well, I think you probably still need to be doing the building because by the way, in order — the problems they're worried about is you have a recursive self-improvement, and they've built up over the years a whole bunch of different techniques for, you know, I presume most of the people on your podcast know what alignment is, but I'll explain it very quickly, which is how do you have these things align to positive human outcomes so they don't tell you how to build weapons. They, you know, generally speaking, provide you useful things in your life and not things that help you mess up other people's lives or your own. And the challenge in recursive self-improvement is you're essentially saying, okay, improve yourself. How do you keep that alignment as you improve yourself? And it's like, well, we don't really know how to do that. That's like — we have some ideas, but if you said guarantee me you can keep your alignment, they're like we can't guarantee it because we don't know what that pattern is. But the problem is you won't know — we won't know what the pattern of recursive self-improvement is until we start doing it. So that's why development and deploy. So like I tend to say no, no, no, keep developing, but maybe slow down deploy, you know, have much higher monitoring on access to the internet and some other things. And by the way, be simultaneously building AIs for measuring what the states are of these complicated RSI models because we need to be understanding them in order to say, oh, these patterns of alignment work, or by the way, if we're discovering no patterns of alignment working well, then at that time a real slowdown can make a difference.
Is recursive self-improvement — sorry, development or deployment? Like you want to keep using recursive self-improvement
Yes, but in deployment, like, okay, let's pace that.
I mean that gets to one of the issues that comes up in your essay, which is the more we say the answer to this is labs hold back what they have, you know, maybe, you know, there are a lot fewer people to get angry about it, nobody's seeing it, but it's fundamentally denying sort of those capabilities to the American public. Like part of what you're talking about is like make sure they get actual applications that they can use. They can see the benefits of AI. I agree with that. But then there's also just the piece of a world where they're too tightly guarded is a world where corporations and maybe governments have this immense power that everybody else doesn't have access to. How do you balance those things?
Well, so first in my Substack essay, one of the things that I think is critical for anyone who — and I think it's the right thing to have to try to help people have a positive understanding of AI. I think government should be actually helping with that. I think obviously companies desperately need to. And so it's like have a medical assistant, like make a 24/7 medical assistant that's only available to the super wealthy right now, but could be available to everybody. Legal assistant. Most people can't hire a lawyer to say, I'm signing this rental agreement. What are — is this okay? Is it standard? So have a legal assistant on a whole wide variety of things like, oh, what are my legal responsibilities if I'm driving the car? A tutor, right? So tutor for your kids, tutor for yourself. Obviously that has some play to what could have — all of that stuff should be deployed aggressively. And to really deploy it well, government will have to change some liability law, like on medical. One of the things —
I desperately want this. I think we could solve a — not to — there's so many intertwined problems, but we live in this like surveillance state potentially that AI is going to get better at creating. Like you could have like an AI sort of bill of rights where it's you have legal protections, you're allowed to use it, you're — yes, there are cameras everywhere, but this is where they're limited. Like there's a lot of issues that could be bundled into a productive conversation.
Exactly. And by the way, the use of surveillance is frequently a sloganeering word. You know, like a great title called Surveillance Capitalism, and you're like, wow, but it's not just bad. It's like, well, surveillance medicine is awesome. We do it all the time.
Exactly. And when you know, when there is like a serial killer or a manhunt, usually people are like, how can they not find him right away? But then you know, other times they hate sort of the dragnet. I mean, I don't think we like that we're being sort of lied to about maybe the extent of the dragnet. And also, by the way, it should be a highly understood environment that is legally bound to be on your side.
Right. Right. And this is one of the things, like for example, most people don't realize that London has an enormous amount of surveillance, and the Londoners are actually quite happy with it because it's like, oh, when someone falls down with a heart attack, the ambulance arrives sooner. That's good. Or when there is a crime, the police respond quicker. We like that because it's for us, not against us. And that's I think one of the really key things here. And so I think continuing to deploy carefully — that I actually, and my slowdown in deployment, it wasn't any of these apps. I actually even think kind of the work productivity and all the rest, I think that's part of the industrial revolution, you need to be doing it. It's more the question of, okay, so we're making a much more intelligent AI through RSI. Let's make sure we understand what that intelligence is. Right? So like don't willy-nilly kind of just go, great, we've made something super smart and now we're just throwing it out. You're like, make sure we have a good understanding that it's kind of aligned with human interest. And then — but for —
There's also this risk, you put it out, it's so easy to distill the models. Then someone's going to rip it off, make it open source, which we've seen a bunch of.
And by the way, so like one of the things on the open-weight side, which is another frequent discussion here. Look, I was on the board of Mozilla for 11 years. I'm a massively pro-open-source on all kinds of things. The open-weight models, when it gets to AI, you really do have to have good answers to like cybersecurity and hacking. You have to have good answers to like, you know, creation of new biological materials because of terrorism. Like that's not a, oh yeah, it's fine, or oh, you can discover those documents on Google, because discovering the documents on Google is not the same thing as having a PhD collaborator that will do what you ask it to do. And it's like, okay, we should be a little bit more careful about that. Again, open is generally good, but it's an area to be careful about.
There's so many pieces of this that I don't want to drift from earlier things, but I'm also interested in that piece of it. I mean, like the lawyer, the medical sort of app, just to put it directly, like do you think people should — you know, if I talk to ChatGPT or a legal-specific app, I should be able to treat that as privileged like a lawyer? Like we should —
Unequivocally, right. And by the way, we're just fuzzy on it right now because by the way, you know, your email that's on the, you know, Google or Microsoft servers is privileged, and it's still in that thing, it's being handled by software, etc. Like that's already there. There's no reason that an interaction with an agent should be treated any differently, right, from a privilege standpoint, whether it's legal, medical, etc. And it's really just people's —
But if I write down — I guess there's one, it's like I'm using it as a tool to talk to lawyers. But there's another where it's like obviously if I wrote a Word document with like my confession and they found on my computer, they'd use that against me. But if I'm like in ChatGPT and I'm asking it like, I think I was speeding, like what — I don't know, like do you think we should have protections there that it's like a therapist, or as long as I'm not like actively saying I'm murdering that person?
Yeah. Look, I think broadly yes. Although also maybe with — like one of the things I think it kind of gets down to kind of how you do a terms of service. I would ideally I think have a terms of service that if someone's like speeding, etc., that should be within the personal — if it's like how do I make a new biological terror weapon, no, not so much, right?
So it's probably a question of where the boundary should be. But like for example, because if you think about this — we're in a far healthier society if everybody has the advice of some legal counsel, and it's very difficult to afford —
Having at least an ability to have an interaction for, can you proxy legal counsel for me —
— is I think a very good feature to have.
That's what I was essentially gesturing at earlier. Well, you know, to bring a lot — it's like I certainly don't support people attacking Flock cameras, but you can imagine a world today, it's like the government has all these tools to come after somebody with Flock cameras, and who's like destroying them, and somebody goes to ChatGPT and says, I've done this, like what — you know, maybe it needs to be a legal-specific app because the problem is lawyers can get in trouble if they give you advice of abetting a crime versus here's the legal advice I can give you around a potential crime or something. But yeah, do you think like there needs to be some framework where we have talking to a model for legal advice is as if it was talking to a lawyer?
I think that's essentially correct. And actually, by the way, you might do is you might say in that mode, like the model only talks legal, like so you can't go —
To switch into it, you're not —
Yes. And like for example, it's like, well, you know, I'd like to figure out how to break into this building, and maybe the model can help you. Well, a lawyer won't do that. So if you're talking on legal mode, it was like, no, my legal advice to you is don't break into the building.
Exactly, right. And on the medical front, do you think we need new policy?
In the same — at the minimum from legal liability. One of the things that I've been trying to advocate for a number of years, and it's part of the reason I've been doing it in to governments and a bunch of other folks, and now I'm just kind of think, okay, let me just start saying this.
Just shout about it loudly enough and hope the culture takes it on.
Yes. Exactly. Because if we shift — part of the reason why most people are not aware that you can actually get pretty good medical advice from ChatGPT and Claude, and the reason is they don't advertise it and they lightly steer away from it to try to avoid medical liability.
If they shout, they say we can do it, then they can get sued when they fail.
Exactly. And what you want to do is say, okay, government should say, look, if you follow the following things, then you have essentially a kind of a safe harbor from medical liability. So that — and by the way, so like the kind of things I think you would do is you'd say the app has to be free, the data has to be entirely the own individual's, you have to not sell the data to anybody, you have to fundamentally ask, do you have access to a doctor? Do you have access to, you know, kind of like medical facilities and so forth? And then when you're giving advice, you should say, look, it's still far better if you can consult a doctor, but if you can't, you know, like timing, availability, etc., here's what I would think, right? You know, kind of thing. And if you had that, at least three companies would go provide that. It would cost the government nothing. But — and they would advertise that they would have it because it would be like, look, this is a really useful feature. And by the way, it is a really useful feature. And here's how useful the feature is today. I personally know three people whose lives have been saved by ChatGPT by using it as a second opinion. Right? So like in a hospital, and they go, I'm not sure, and they go to another hospital, and that hospital — and it was like, well, if you arrived here about an hour and a half later, you'd be dead. Right? You know, that kind of thing. And that's — we want that for everyone, right? And that's easy enough to do. Now, of course, lawyers will argue, no, no, no, no. We need to have medical tort. It's the only way to keep it. And it's like, well, just describe what the guidelines are for where you have to stay. And if you think, oh, that's a little borderline, then bring in the guideline a little bit. It's like only within — like part of a medical assistant could be the, I'm sorry, I'm not allowed to give advice on that topic, right? You know, please find someone to consult with. If you want, I can pull up a list of public, you know, like potentially publicly available resources that you might be able to use.
It's a really interesting point on the why don't OpenAI, Anthropic, etc. do a better job of taking credit for the good things they can do. It's like here's a case where maybe they're afraid that they'll get sued if they brag about it. Interesting, right? And it seems — I mean, it's clearly a good that you can, you know, put your medical documents into it, get just get a different opinion, like if it disagrees with your doctor.
And also, by the way, like another thing is literally for people who have high-end concierge doctors, they put their blood work into ChatGPT, and then you can talk to the agent for two hours about like what does this mean and how many people have this and you know, does this mean that I have the following condition? What would be the probability? How else would I measure it? Should I take the blood test again? Doctors don't have time for that. AI has all the time in the world.
