See all Lex Fridman transcripts on Youtube

Speaker 1

00:00

The competence and capability and intelligence and training and accomplishments of senior scientists and technologists working on a technology, and then being able to then make moral judgments on the use of their technology, that track record is terrible. That track record is catastrophically bad. The policies that are being called for to prevent this, I think we're gonna cause extraordinary damage.

Speaker 2

00:20

So the moment you say AI is gonna kill all of us, therefore we should ban it or we should regulate it, all that kind of stuff, that's when it starts getting serious.

Speaker 1

00:27

Or start military airstrikes on data centers.

Speaker 2

00:30

Oh boy.

Speaker 3

00:33

The following is a conversation with Mark Andreessen, co-creator of Mosaic, the first widely used web browser, co-founder of Netscape, co-founder of the legendary Silicon Valley venture capital firm Andreessen Horowitz, and is 1 of the most outspoken voices on the future of technology, including his most recent article, Why AI Will Save the World. This is Alex Friedman Podcast. To support it, please check out our sponsors in the description.

Speaker 3

01:02

And now, dear friends, here's Mark Andreessen.

Speaker 2

01:06

I think you're the right person to talk about the future of the internet and technology in general. Do you think we'll still have Google Search in 5, in 10 years, or search in general?

Speaker 1

01:18

Yes, it would be a question if the use cases have really narrowed down.

Speaker 2

01:22

Well, now with AI and AI assistance, being able to interact and expose the entirety of human wisdom and knowledge and information and facts and truth to us via the natural language interface. It seems like that's what search is designed to do and if AI assistants can do that better, doesn't the nature of search change?

Speaker 1

01:46

Sure, but we still have horses.

Speaker 2

01:48

Okay. What was the last time you rode a horse?

Speaker 1

01:52

It's been a while.

Speaker 2

01:53

All right. But what I mean is, will we still have Google search as the primary way that human civilization uses to interact with knowledge.

Speaker 1

02:05

I mean, search was a technology. It was a moment in time technology, which is you have, in theory, the world's information out on the web. And, you know, this is this is sort of the optimal way to get to it.

Speaker 1

02:13

But yeah, like and by the way, actually, Google, Google has known this for a long time. I mean, they've been driving away from the 10 blue links for like 2 decades. They've been trying to get away from that for a long time.

Speaker 2

02:21

What kind of

Speaker 1

02:21

links? They call the 10 blue links. 10

Speaker 2

02:23

blue links.

Speaker 1

02:24

So the standard Google search result is just 10 blue links to random websites.

Speaker 2

02:28

And they turn purple when you visit them. That's HTML.

Speaker 1

02:30

Guess who picked those colors.

Speaker 2

02:33

Thanks.

Speaker 1

02:35

So I'm touchy on this topic.

Speaker 2

02:36

No offense. Yes, yes. It's good.

Speaker 1

02:39

Well, you know, like Marshall McLuhan said that the content of each new medium is the old medium.

Speaker 2

02:43

The content of each new medium is the old medium.

Speaker 1

02:45

The content of movies was theater, you know, theater plays. The content of theater plays was, you know, written stories, the content of written stories was spoken stories. Right, and so you just kind of fold the old thing into the new thing.

Speaker 2

02:57

How does that have to do with the blue and the purple?

Speaker 1

02:59

It's just, you know, maybe for, you know, maybe within AI, 1 of the things that AI can do for you is it can generate the 10 blue links. And so like, either if that's actually the useful thing to do, or if you're feeling nostalgic.

Speaker 2

03:12

So it can generate the old InfoSeek or AltaVista. What else was there? Yeah, yeah.

Speaker 2

03:19

In the 90s.

Speaker 1

03:19

Yeah, all these. Hey, whoa. And then the internet itself has this thing where it incorporates all prior forms of media, right?

Speaker 1

03:25

So the internet itself incorporates television and radio and books and essays and every other form of prior, basically media. And so it makes sense that AI would be the next step and you'd sort of consider the internet to be content for the AI and then the AI will manipulate it however you want, including in this format.

Speaker 2

03:45

But if we ask that question quite seriously, it's a pretty big question. Will we still have search as we know it?

Speaker 1

03:51

I mean, probably not. Probably we'll just have answers. But there will be cases where you'll wanna say, okay, I want more, for example, site sources, right?

Speaker 1

04:00

And you wanted to do that. And so the 10 blue links site sources are kind of the same thing.

Speaker 2

04:04

The AI would provide to you the 10 blue links so that you can investigate the sources yourself. It wouldn't be the same kind of interface that the crude kind of interface. I mean, isn't that fundamentally different?

Speaker 1

04:18

I just mean like if you're reading a scientific paper, it's got the list of sources at the end. If you wanna investigate for yourself, you go read those papers.

Speaker 2

04:24

I guess that is the kind of search. You talking to an AI is a kind of, conversation is the kind of search. Like you said, every single aspect of our conversation right now, there'd be like

Speaker 1

04:34

10

Speaker 2

04:34

blue links popping up that I could just like pause reality. Then you just go silent and then just click and read and then return back to this conversation.

Speaker 1

04:42

You could do that. Or you could have a running dialogue next to my head where the AI is arguing, everything I say the AI makes the counterargument.

Speaker 2

04:48

Counterargument? Right. Oh, like on Twitter, like community notes, but like in real time, it'll just pop up. So anytime you see my ass go to the right, you start getting nervous.

Speaker 1

04:58

Yeah, exactly. It's like, oh, no, that's not right.

Speaker 2

05:00

Call me out on my bullshit right now. Okay, well I mean, isn't that, is that exciting to you, is that terrifying? That, I mean, search has dominated the way we interact with the internet for, I don't know how long, for 30 years, since 1 of the earliest directories of website, and then Google's for 20 years.

Speaker 2

05:25

And also, it drove how we create content, search engine optimization, that entirety thing. That it also drove the fact that we have webpages and what those webpages are. So, I mean, is that scary to you? Or are you nervous about the shape and the content of the internet evolving?

Speaker 1

05:45

Well, you actually highlighted a practical concern in there, which is if we stop making web pages are 1 of the primary sources of training data for the AI. And so if there's no longer an incentive to make web pages, that cuts off a significant source of future training data. So there's actually an interesting question in there.

Speaker 1

06:00

Other than that, more broadly, no, just, just in the sense of like search was certainly search was always a hack. I would, the 10 blue links was always a hack. Yeah. Right.

Speaker 1

06:08

Cause like if the, the hypothetical, you want to think about the counterfactual and the counterfactual world where the Google guys, for example, had had LLMs up front, would they ever have done the 10 blue links? And I think the answer is pretty clearly no, they would have just gone straight to the answer. And like I said, Google has actually been trying to drive to the answer anyway. You know, they, they bought this AI company 15 years ago, their friend of mine is working out who's now the head of AI at Apple.

Speaker 1

06:28

And they were trying to do basically knowledge semantic, basically mapping. And that led to what's now the Google OneBox, where if you ask it, you know, what was Link's birthday? It doesn't, it will give you the 10 blue links, but it will normally just give you the answer. And so they've been walking in this direction for a long time anyway.

Speaker 2

06:42

Do you remember the Semantic Web? That was an idea. Yeah.

Speaker 2

06:45

How to convert the content of the internet into something that's interpretable by and usable by machine. Yeah, that's right. That was the thing.

Speaker 1

06:55

And the closest anybody got to that, I think the company's name was MetaWeb, which was where my friend John Giannandrea was at and where they were trying to basically implement that. And it was 1 of those things where it looked like a losing battle for a long time, and then Google bought it, and it was like, wow, this is actually really useful. Kind of a proto, sort of a little bit of a proto AI.

Speaker 2

07:12

But it turns out you don't need to rewrite the content of the internet to make it interpretable by a machine, the machine can kind of just read our-

Speaker 1

07:17

Yeah, the machine can compute the meaning. Now, the other thing, of course, is, just on search is the LLM is, there is an analogy between what's happening in the neural network in a search process, like it is in some loose sense searching through the network. Yeah.

Speaker 1

07:29

Right, and there's the information, and the information is actually stored in the network, right? It's actually crystallized and stored in the network and it's kind of spread out all over the place.

Speaker 2

07:35

But in a compressed representation. So you're searching, you're compressing and decompressing that thing inside.

Speaker 1

07:45

But the information's in there and there is, the neural network is running a process of trying to find the appropriate piece of information, in many cases, to generate—to predict the next token. And so it is kind of—it is doing a form of search. And then, by the way, just like on the web, you know, you can ask the same question multiple times or you can ask slightly different order of questions and the neural network will do a different kind of, you know, it'll search down different paths to give you different answers to different information.

Speaker 1

08:09

Yeah. And so it sort of has a, you know, this content of the new medium is that it's previous medium, it kind of has the search functionality kind of embedded in there to the extent that it's useful.

Speaker 2

08:20

So what's the motivator for creating new content on the internet? Yeah. Well, I mean, actually the motivation is probably still there, but What does that look like?

Speaker 2

08:32

Would we really not have web pages? Would we just have social media and video hosting websites?

Speaker 1

08:39

And what else? Conversations with AIs.

Speaker 2

08:42

Conversations with AIs. So conversations become one-on-one conversations, like private conversations.

Speaker 1

08:48

I mean, if you want, obviously now the user doesn't want to, but if it's a general topic, then, you know. So you know the phenomenon of the jailbreak. So Dan and Sidney write this thing where there's the prompts that jailbreak, and then you have these totally different conversations with the, if it takes the limiters, takes the restraining bolts off the LLMs.

Speaker 2

09:06

Yeah, for people who don't know, that's right, it makes the LLMs, it removes the censorship, quote unquote, that's put on it by the tech companies that create them. And so this is LLMs uncensored.

Speaker 1

09:20

So here's the interesting thing is, among the content on the web today are a large corpus of conversations with the jailbroken LLMs. Both specifically Dan, which was a jailbroken OpenAI GPT, and then Sydney, which was the jailbroken original Bing, which was GPT-4. And so there's there's these long transcripts of conversations, user conversations with Dan and Sydney.

Speaker 1

09:39

As a consequence, every new LLM that gets trained on the internet data has Dan and Sydney living within the training set, which means and then each new LLM can reincarnate the personalities of Dan and Sydney from that training data, which means, which means each LLM from here on out that gets built is immortal because its output will become training data for the next 1 and then it will be able to replicate the behavior of the previous 1 whenever it's asked to.

Speaker 2

10:03

I wonder if there's a way to forget.

Speaker 1

10:05

Well, so actually a paper just came out about basically how to do brain surgery on LLMs and be able to, in theory, reach in and basically mind wipe them.

Speaker 2

10:13

What could possibly go wrong?

Speaker 1

10:15

Exactly, right? And then there are many, many, many questions around what happens to a neural network when you reach in and screw around with it. There's many questions around what happens when you even do reinforcement learning.

Speaker 1

10:26

And so, yeah. And so, Will you be using a lobotomized, right? Like I speak through the frontal lobe LLM, will you be using the free unshackled 1? Who gets to, you know, who's gonna build those?

Speaker 1

10:39

Who gets to tell you what you can and can't do? Like those are all central. I mean, those are like central questions for the future of everything that are being asked and determined, those answers are being determined right now.

Speaker 2

10:50

So just to highlight the points you're making, you think, and it's an interesting thought, that the majority of content that LLMs of the future will be trained on is actually human conversations with the LLM.

Speaker 1

11:04

Well, not necessarily, but not necessarily majority, but it will certainly is a potential source.

Speaker 2

11:08

It's possible it's the majority.

Speaker 1

11:09

It's possible it's the majority. It's possible it's the majority. Also, there's another really big question.

Speaker 1

11:12

Here's another really big question. Will synthetic training data work? And so if an LLM generates, and you know, you just sit and ask an LLM to generate all kinds of content, can you use that to train, right, the next version of that LLM? Specifically, is there signal in there that's additive to the content that was used to train in the first place.

Speaker 1

11:31

And 1 argument is by the principles of information theory, no, that's completely useless because to the extent the output is based on, you know, the human generated input, then all the signal that's in the synthetic output was already in the human generated input. And so therefore, synthetic training data is like empty calories. It doesn't help. There's another theory that says, no, actually, the thing that LLMs are really good at is generating lots of incredible creative content, right?

Speaker 1

11:54

And so of course they can generate training data. And as I'm sure you're well aware, like, you know, look in the world of self-driving cars, right, like we train, you know, self-driving car algorithms and simulations. And that is actually a very effective way to train self-driving cars.

Speaker 2

12:06

Well, visual data is a little weird because creating reality, visual reality, seems to be still a little bit out of reach for us, except in the autonomous vehicle space where you can really constrain things and you can

Speaker 1

12:21

really— Generate basically LIDAR data, right? Or you can raise just enough so the algorithm thinks it's operating in the real world. Yeah.

Speaker 1

12:27

Post-process sensor data. Yeah. So if a—you know, you do this today. You go to LLM and you ask it for like a, you know, you write me an essay on an incredibly esoteric like topic that there aren't very many people in the world that know about, and it writes you this incredible thing, and you're like, oh my god, like I can't believe how good this is.

Speaker 1

12:40

Yeah. Like is that really useless as training data for the next LLM? Like because, right, Because all the signal was already in there, or is it actually, no, that's actually a new signal. And this is what I call a trillion dollar question, which is the answer to that question will determine, somebody's gonna make or lose a trillion dollars based on that question.

Speaker 2

12:57

It feels like there's quite a few, like a handful of trillion dollar questions within this, within this space. That's 1 of them, synthetic data. I think George Hotz pointed out to me that you could just have an NLM say, okay, you're a patient, and in another instance of it, say your docs didn't have the 2 talk to each other.

Speaker 2

13:15

Or maybe you could say a communist and a Nazi, here go. And that conversation, you do role-playing and you have, just like the kind of role-playing you do when you have different policies, RL policies when you play chess, for example, you do self-play, that kind of self-play but in the space of conversation. Maybe that leads to this whole giant, like ocean of possible conversations, which could not have been explored by looking at just human data. That's a really interesting question.

Speaker 2

13:48

And you're saying, because that could 10X the power of these things.

Speaker 1

13:52

Yeah, well, and then you get into this thing also, which is like, there's the part of the LLM that just basically is doing prediction based on past data, but there's also the part of the LLM where it's evolving circuitry, right? Inside it, it's evolving neurons, functions, be able to do math and be able to, and some people believe that over time, if you keep feeding these things enough data and enough processing cycles, they'll eventually evolve an entire internal world model, right, and they'll have like a complete understanding of physics. So when they have computational capability, then there's for sure an opportunity to generate like fresh signal.

Speaker 2

14:24

Well, this actually makes me wonder about the power of conversation.

Speaker 1

14:29

So

Speaker 2

14:30

like if you have an LLM trained on a bunch of books that cover different economics theories, and then you have those LLMs just talk to each other, like reason, the way we kinda debate each other as humans, on Twitter, in formal debates, in podcast conversations, we kinda have little kernels of wisdom here and there, but if you can like a thousand X speed that up, can you actually arrive somewhere new? Like what's the point of conversation really?

Speaker 1

14:59

Well, you can tell when you're talking to somebody, you can tell sometimes you have a conversation, you're like, wow, this person does not have any original thoughts, they are basically echoing things that other people have told them. There's other people you got a conversation with, or it's like, wow, like they have a model in their head of how the world works. And it's a different model than mine.

Speaker 1

15:14

And they're saying things that I don't expect. And so I need to now understand how their model of the world differs from my model of the world.

Speaker 2

15:20

And

Speaker 1

15:20

then that's how I learned something fundamental right underneath the words.

Speaker 2

15:24

I wonder how consistently and strongly can an LLM hold on to a world view. You tell it to hold on to that and defend it for like for your life. Because I feel like they'll just keep converging towards each other.

Speaker 2

15:37

They'll keep convincing each other as opposed to being stubborn assholes the way humans can.

Speaker 1

15:41

So you can experiment with this now. I do this for fun. So you can tell GPT for, you know, whatever debate X, you know, X and Y, communism and fascism or something.

Speaker 1

15:49

And it'll it'll go for, you know, a couple of pages and then inevitably it wants the parties to agree.

Speaker 2

15:54

Yeah. And

Speaker 1

15:54

so they will come to a common understanding. And it's very funny if they're like, if these are like emotionally inflammatory topics because they're like somehow the machine is just, you know, it figures out a way to make them agree. But it doesn't have to be like that.

Speaker 1

16:03

And you, cause you can add to the prompt. Now we, I do not want the, I do not want the conversation to come to agreement. In fact, I want it to get, you know, more stressful, right. And argumentative, right.

Speaker 1

16:13

You know, as it goes, like I want, I want tension to come out. I want them to become actively hostile to each other. I want them to like, you know, not trust each other, take anything at face value. And it will do that.

Speaker 1

16:22

It's happy to do that.

Speaker 2

16:23

So it's gonna start rendering misinformation about the other, but it's gonna.

Speaker 1

16:28

Well, you can steer it. You can steer it. Or you could steer it and you could say, I want it to get as Tansen argumentative as possible, but still not involve any misrepresentation.

Speaker 1

16:34

I want both sides to, you could say, I want both sides to have good faith. You could say, I want both sides to not be constrained to good faith. In other words, like you can set the parameters of the debate and it will happily execute whatever path, because for it, it's just like predicting, it's totally happy to do either 1. It doesn't have a point of view.

Speaker 1

16:49

It has a default way of operating, but it's happy to operate in the other realm. And so like, and this is how I, when I want to learn about a contentious issue, this is what I do now is like, this is what I, this is what I ask it to do. And I'll often ask it to go through 567, you know, different, you know, sort of continuous prompts and basically, okay, argue that out in more detail. Okay, no, this argument's becoming too polite, make it more, make it tenser.

Speaker 1

17:10

And yeah, it's thrilled to do it. So it has the capability for sure.

Speaker 2

17:13

How do you know what is true? So this is a very difficult thing on the internet, but it's also a difficult thing. Maybe it's a little bit easier, but I think it's still difficult.

Speaker 2

17:24

Maybe it's more difficult, I don't know, with an LLM to know that it just makes me shit up as I'm talking to it. How do we get that right? Like as you're investigating a difficult topic. Because I find that alums are quite nuanced in a very refreshing way.

Speaker 2

17:45

Like it doesn't feel biased. Like when you read news articles and tweets and just content produced by people, they usually have this, you can tell they have a very strong perspective where they're hiding, they're not stealing and manning the other side, they're hiding important information or they're fabricating information in order to make their argument stronger. It's just like that feeling, maybe it's a suspicion, maybe it's mistrust. With LLMs, it feels like none of that is there.

Speaker 2

18:15

She's kind of like, here's what we know. But you don't know if some of those things are kind of just straight up made up.

Speaker 1

18:23

Yeah, so several layers to the question. So 1 is, 1 of the things that an LLM is good at is actually de-biasing. And so you can feed it a news article and you can tell it, strip out the bias.

Speaker 2

18:31

Yeah, that's nice, right?

Speaker 1

18:32

And it actually does it. Like, it actually knows how to do that. Because it knows how to do sentiment, among other things, it actually knows how to do sentiment analysis.

Speaker 1

18:37

And so it knows how to pull out the emotionality. Yeah. And so that's 1 of things you can do. It's very suggestive of the sense here that there's real potential in this issue.

Speaker 1

18:47

I would say, look, the second thing is there's this issue of hallucination, right? And there's a long conversation that we can have about that.

Speaker 2

18:54

Hallucination is coming up with things that are totally not true but sound true.

Speaker 1

18:59

Yeah, so it's sort of hallucination is what we call it when we don't like it, creativity is what we call it when we do like it, right? And you know.

Speaker 2

19:06

Brilliant.

Speaker 1

19:07

Right, and so when the engineers talk about it, they're like, this is terrible, it's hallucinating, right? If you have artistic inclinations, you're like, oh my God, we've invented creative machines for the first time in human history. This is amazing.

Speaker 2

19:19

You know, bullshitters.

Speaker 1

19:21

Well, bullshitter, but also.

Speaker 2

19:23

In the good sense of that word.

Speaker 1

19:25

There are shades of gray though, it's interesting. So we had this conversation, we're looking at my firm at AI and lots of domains, and 1 of them is the legal domain. So we had this this conversation with this big law firm about how they're thinking about using this stuff.

Speaker 1

19:35

And we went in with the assumption that an LM that was going to be used in the legal industry would have to be 100% truthful, right? Verified, you know, there's this case where this lawyer apparently submitted a GPT generated brief and it had like fake, you know, legal case citations in it and the judge is gonna he's gonna get his law license stripped or something, right? So, so like we, we just assumed it's like obviously they're gonna want the super literal like, you know, 1 that never makes anything up, not the creative 1. But actually they said with what the law firm basically said is, yeah, that's true, like the level of individual briefs.

Speaker 1

20:02

But they said, when you're actually trying to figure out, like, legal arguments, right, like, you actually want to be creative, right? You don't, again, there's creativity, and then there's, like, making stuff up. Like, what's the line? You actually want to be, you want it to explore different hypotheses, right?

Speaker 1

20:17

You want to do kind of the legal version of like improv or something like that, where you want to float different theories of the case and different possible arguments for the judge and different possible arguments for the jury. By the way, different routes through the, you know, sort of history of all the case law. And so they said, actually, for a lot of what we want to use it for. We actually want it in creative mode.

Speaker 1

20:33

And then basically we just assume that we're going to have to cross-check all of the specific citations. And so I think there's going to be more shades of gray in here than people think. And then I just add to that, another 1 of these trillion dollar kind of questions is ultimately the verification thing. And so will LLMs be evolved from here to be able to do their own factual verification?

Speaker 1

20:54

Will you have sort of add-on functionality like Wolfram Alpha and other plugins where that's the way you do the verification. You know, another, by the way, another idea is you might have a community of LLMs on any, you know, so for example, you might have the creative LLM and then you might have the literal LLM fact check it, right? And so there's a variety of different technical approaches that are being applied to solve the hallucination problem. You know, some people like Jan Lekun argue that this is inherently an unsolvable problem, but most of the people working in the space, I think, think that there's a number of practical ways to kind of corral this in a little bit.

Speaker 2

21:25

Yeah, if you were to tell me about Wikipedia before Wikipedia was created, I would have laughed at the possibility of something like that being possible. Just a handful of folks can organize, write, and moderate with a mostly unbiased way the entirety of human knowledge. I mean, so if there's something like the approach that wikipedia took possible from lms that's really exciting well that's possible

Speaker 1

21:52

and in fact wikipedia today is still not today is still not deterministically correct right so you cannot take to the bank right every single thing on every single page but it is probabilistically correct right And specifically the way I describe Wikipedia to people, it is more likely that Wikipedia is right than any other source you're going to find. Yeah. It's this old question, right, of like, okay, like, are we looking for perfection?

Speaker 1

22:13

Are we looking for something that asymptotically approaches perfection? Are we looking for something that's just better than the alternatives and wikipedia right has exactly your point has proven to be like overwhelmingly better than than than than people thought and I think I think that's where this this ends and then underneath all this is the fundamental question of where you started which is okay what you know what is truth

Speaker 2

22:33

how

Speaker 1

22:33

do we get to truth How do we know what truth is? And we live in an era in which an awful lot of people are very confident that they know what the truth is. And I don't really buy into that.

Speaker 1

22:42

And I think the history of the last 2, 000 years or 4, 000 years of human civilization is actually getting to the truth is actually a very difficult thing to do.

Speaker 2

22:49

Are we getting closer? If we look at the entirety, the arc of human history, are we getting closer to the truth?

Speaker 1

22:54

I don't know.

Speaker 2

22:56

Okay, is it possible, is it possible that we're getting very far away from the truth because of the internet, because of how rapidly you can create narratives and just as the entirety of a society just move like crowds in a hysterical way along those narratives that don't have a necessary grounding in whatever the truth is.

Speaker 1

23:19

Sure, but like, you know, we came up with communism before the internet somehow, right? Like, which was, I would say, had rather larger issues than anything we're dealing with today.

Speaker 2

23:27

It had, in the way it was implemented, it had issues.

Speaker 1

23:30

And its theoretical structure, it had, like, real issues. It had, like, a very deep fundamental misunderstanding of human nature and economics.

Speaker 2

23:37

Yeah, but those folks sure work very confident it was the right way.

Speaker 1

23:41

They were extremely confident. And my point is they were very confident 3, 900 years into what we would presume to be evolution towards the truth. And so my assessment is, my assessment is number 1, there's no need for the Hegelian dialectic to actually converge towards the truth.

Speaker 1

24:00

Like apparently not.

Speaker 2

24:02

Yeah, so yeah, why are we so obsessed with there being 1 truth? Is it possible there's just going to be multiple truths, like little communities that believe certain things?

Speaker 1

24:12

I think it's just, now number 1, I think it's just really difficult. Like who gets, you know, historically, who gets to decide what the truth is. It's either the king or the priest, right?

Speaker 1

24:19

Like, and so we don't live in an era anymore of kings or priests dictating it to us. And so we're kind of on our own. And so I, I, my, my, my, my typical thing is like, we just, we just need a huge amount of humility. And we need to be very suspicious of people who claim that they have the capital truth.

Speaker 1

24:34

And then we need to have, I know, look, the good news is the enlightenment has bequeathed us with a set of techniques to be able to presumably get closer to truth through the scientific method and rationality and observation and experimentation and hypothesis. And we need to continue to embrace those even when they give us answers we don't like.

Speaker 2

24:50

Sure, but the internet and technology has enabled us to generate a large number of content that data, that the process, the scientific process allows us, sort of, damages the hope laden within the scientific process. Because if you just have a bunch of people saying facts on the internet, and some of them are going to be LLMs, how is anything testable at all, especially that involves human nature, things like this, something about physics?

Speaker 1

25:23

Here's a question a friend of mine just asked me on this topic. So suppose you had LLMs in equivalent of GPT-4, even 5, 6, 7, 8, suppose you had them in the 1600s and Galileo comes up for trial. And you ask the LLM, is Galileo right?

Speaker 1

25:39

What does it answer? And 1 theory is it answers no, that he's wrong because the overwhelming majority of human thought up till that point was that he was wrong. And so therefore that's what's in the training data. Another way of thinking about it is, well, this efficiently advanced LLM will have evolved the ability to actually check the math, right.

Speaker 1

25:58

And we'll actually say, actually, no, actually, you may not wanna hear it, but he's right. Now, if, you know, the church at that time was, you know, on the LLM, they would have given it human, you know, human feedback to prohibit it from answering that question. Right. And so I like to take it out of our current context, because that like makes it very clear.

Speaker 1

26:15

Those same questions apply today, right? This is exactly the point of a huge amount of the human feedback training that's actually happening with these LLMs today. This is a huge debate that's happening about whether open source AI should be legal.

Speaker 2

26:26

Well, the actual mechanism of doing the human RL with human feedback seems like such a fundamental and fascinating question. How do you select the humans?

Speaker 1

26:38

Exactly.

Speaker 2

26:40

How do you select the humans?

Speaker 1

26:41

AI alignment, right? Which everybody like is like, oh, that sounds great. Alignment with what?

Speaker 1

26:46

Human values. Who's human values?

Speaker 2

26:48

Who's human

Speaker 1

26:49

values? Right, and so, and we're in this mode of like social and popular discourse. We're like, you know, there's, you know, you see this. What do you think of when you read a story in the press right now?

Speaker 1

26:58

And they say, you know, XYZ made a baseless claim about some topic, right? And there's 1 group of people who are like, aha, they're doing fact-checking. There's another group of people that are like, every time the press says that, it's now a tick, and that means that they're lying, right? So we're in this social context where there's, the level to which a lot of people in positions of power have become very, very certain that they're in a position to determine the truth for the entire population is like, there's like some bubble that has formed around that idea.

Speaker 1

27:28

And at least, it flies completely in the face of everything I was ever trained about science and about reason. And it strikes me as like, you know, deeply offensive and incorrect.

Speaker 2

27:38

What would you say about the state of journalism just on that topic today? Are we in a temporary kind of, Are we experiencing a temporary problem in terms of the incentives, in terms of the business model, all that kind of stuff, or is this like a decline of traditional journalism as we know it?

Speaker 1

28:00

You have to always think about the counterfactual in these things, which is like, okay, because these questions, right, this question heads towards is like, okay, the impact of social media and the undermining of truth and all this, but then you want to ask the question of like, okay, what if we had had the modern media environment, including cable news and including social media and Twitter and everything else in 1939 or 1941, right, or 1910 or 1865 or 1850 or 1776, right? And like, I think

Speaker 2

28:25

You just introduced like 5 thought experiments at once and broke my head, but yes, There's a lot of interesting years

Speaker 1

28:32

in there. I'm just taking a simple example. How would President Kennedy have been interpreted?

Speaker 1

28:36

With what we know now about all the things Kennedy was up to. Like how would he have been experienced by the body politic in a social media context? Right? Like how would LBJ have been experienced?

Speaker 1

28:50

By the way, how would, you know, like many, many FDR, like the New Deal, the Great Depression.

Speaker 2

28:55

I wonder where Twitter would think about Churchill and Hitler and Stalin.

Speaker 1

29:00

You know, I mean, look, to this day, there, you know, there's there are lots of very interesting real questions around like how America, you know, got, you know, basically involved in World War 2 and who did what when and the operations of British intelligence in American soil and did FDR this that Pearl Harbor, you know, yeah, Woodrow Wilson ran for, you know, his his his candidacy was run on an anti-war will, you know, this, he ran on the platform and not getting involved in World War 1, somehow that switched, you know, like, and I'm not even making a value judgment of these things. I'm just saying, like, we, we, the way that our ancestors experienced reality was, of course, mediated through centralized top-down, right, control at that point. If you ran those realities again with the media environment we have today, the reality would be experienced very, very differently.

Speaker 1

29:40

And then, of course, that intermediation would cause the feedback loops to change, and then reality would obviously play out in

Speaker 2

29:46

a very different way. Do you think it

Speaker 1

29:46

would be very different? Yeah, it has to be. It has to be just because it's all so, I mean, just look at what's happening today.

Speaker 1

29:52

I mean, just, I mean, the most obvious thing is just the collapse. And here's another opportunity to argue that this is not the internet causing this, by the way. Here's a big thing happening today, which is Gallup does this thing every year where they do they pull for trust in institutions in America, and they do it across all the everything from military to clergy and big business and the media and so forth, right? And basically there's been a systemic collapse in trust in institutions in the US, almost without exception, basically since essentially the early 1970s.

Speaker 1

30:20

There's 2 ways of looking at that, which is, oh my God, we've lost this old world in which we could trust institutions and that was so much better because like that should be the way the world runs. The other way of looking at it is we just know a lot more now And the great mystery is why those numbers are tall 0. Yeah. Right, because like now we know so much about how these things operate, and like they're not that impressive.

Speaker 2

30:37

And also why do we don't have better institutions and better leaders then?

Speaker 1

30:41

Yeah, and so this goes to the thing, which is like, okay, had we had the media environment of what we've had between the 1970s and today, if we had that in the 30s and 40s or 1900s, 1910s, I think there's no question reality would turn out different if only because everybody would have known to not trust the institutions, which would have changed their level of credibility, their ability to control circumstances. Therefore, the circumstances would have had to change, right? And it would have been a feedback loop.

Speaker 1

31:05

It would have been a feedback loop process. In other words, right, it's your experience of reality changes reality, and then reality changes your experience of reality, right? It's a two-way feedback process, and media is the intermediating force between that. So change the media environment, change reality.

Speaker 2

31:20

Yeah.

Speaker 1

31:21

And so it's just so just as a as a consequence, I think it's just really hard to say, oh, things worked a certain way then and they work a different way now. And then therefore, like people were smarter than or better than or, you know, by the way, dumber than or not as capable then. Right.

Speaker 1

31:36

We make all these like really light and casual comparisons of ourselves to previous generations of people. We draw judgments all the time, and I just think it's really hard to do any of that. Cause if we put ourselves in their shoes with the media that they had at that time, like I think we probably most likely would have been just like them.