Name: Joe Rogan Experience #2010 - Marc Andreessen
Uploaded: 2023-07-19T00:00:00.000Z
Description: Watch the stand-up comedy video 'Joe Rogan Experience #2010 - Marc Andreessen' from PowerfulJRE on Comedy Buff. Published 2023-07-19. Transcript available.

Watch the stand-up comedy video 'Joe Rogan Experience #2010 - Marc Andreessen' from PowerfulJRE on Comedy Buff. Published 2023-07-19. Transcript available.

Joe Rogan Experience #2010 - Marc Andreessen

Joe Rogan podcast, check it out.
The Joe Rogan experience.
Train by day, Joe Rogan podcast by night, all day.
Good morning, Mark. Good to see you.
Fantastic, thanks.
You are in the middle of this AI discussion.
You're right in the heat of this thing.
But I think you have a different perspective than a lot of people do.
A lot of people are terrified of AI, me included.
Oh, okay. All right, okay.
For all the wrong reasons.
Of all the things to worry about.
For me, my terror of it is all the wrong.
It's kind of fun, terror.
Sure, of course.
I'm not really freaking out.
But I am recognizing that this is an emerging technology
that is so different than anything we've ever experienced before.
Particularly what's chat GPT, what's happening with that right now.
It's really fascinating and a lot of advantages.
Like we were just talking last night,
someone in the green room brought up the fact that there was this,
they're using it for medical diagnoses, and it's very accurate.
Which is incredible.
Yeah.
There's a lot of good things to it.
Yeah, yeah.
Well, so you probably remember last time I was on,
we spent quite a bit of time talking about this,
and this was when these chat bots were running inside Google,
but the rest of us didn't have access to them yet.
Right.
Right.
And that guy had come out and said that he thought that they were self-aware.
Yes.
And the whole thing was like this big kind of mystery of like what's going on.
Right.
And now the world gets to use these things, right?
Everybody since then, everybody kind of has access.
Really quickly.
Yeah.
That was a short amount of time.
Yeah, it's been great.
And then look, these things are, you know, these things when I say this,
it's like ChatGPT, and then Microsoft has their version called Bing.
Google has a version called Bard now that's really good.
There's a company, Anthropic, that has a thing called Claude.
You know, if you just run the comparison, they're basically as good as a doctor.
They're as good as the average doctor at this point at being a doctor.
They're as good at being a lawyer as the average lawyer.
You kind of go through basically anything involving knowledge work,
you know, anything involving information, synthesizing, reporting,
you know, writing legal briefs, anything like this.
In business, they're actually already really good.
They're as good as the average management consultant.
Now, the way they acquire data, they're essentially scouring the Internet,
right?
Sort of.
It was more like they're fed the Internet.
They're fed the Internet.
And I say it makes a difference because the company that produces the AI
determines what data goes into it,
and that determines a lot of how it works and what it does or won't do.
Okay.
So in that regard, is there a concern that someone could feed it fake data?
Yeah.
Yeah.
Well, you may have noticed that people over time have said a lot of fake things.
Yes, I have noticed that.
So that's all in there.
So the way to think about it basically is it's being trained.
The full version of these things are being trained on basically the sum total
of human written expression.
Right?
So basically everything people have ever written.
There are some issues, and you've got to get all, you know, somehow we've got
to figure out how to get all the books in there.
Although all the books prior to 1923 are in there because they're all out of
copyright.
But more recent books are a challenge.
But anything that you can access on the Internet that's text, right, which is,
you know, a staggeringly broad, you know, set of material is in there.
By the way, both nonfiction and fiction, right, so a lot of stories are in
there.
And then the new versions of these that are being built right now are what are
called multimodal.
And so that means you can feed them not only text, but you can also feed them
images.
You can feed them videos, right?
So they're going to be trained on all of YouTube, right?
They're going to be trained on all podcasts, right?
And they're going to be trained kind of equivalently between text and images
and video and all kinds of other data.
And so they're going to – they already have very comprehensive knowledge of
human affairs, but it's going to get very complete.
So if it's scouring the – if it's getting all this data from both fiction and
nonfiction, how does it interpret data that's kind of satire?
Like what does it do with, like, Hunter S. Thompson, like gonzo journalism?
So it doesn't really know the difference.
Like this is one of the things that's difficult about talking about this
because you kind of want to always kind of compare it to a person.
And part of it is you refer to it as an it.
And there's this concept of anthropomorphizing things that aren't human.
So it's kind of not really a correct thing to kind of think about it as like
that there's an it per se.
There's no, like, genie in the bottle.
Like there's no, you know, sort of being in there that understands this is satire
or not satire.
It's more sort of a collective understanding of everything all at once.
And then what happens is basically you as the user kind of give it direction of
what path you want it to go down, right?
And so if you sort of imply to it that you want it to sort of like explore, you
know, fictional scenarios, it will happily explore those scenarios with you.
I'll give you an example.
You can tell it, you know, whatever date the Titanic went down.
Let's say it's, I don't know, July 4th, 1923 or whatever it was.
You can say, you know, you can tell it.
It's July 4th, 1923.
It's, you know, 10 o'clock in the morning.
I'm on the Titanic.
Is there anything I should know?
Right.
And it'll like freak out, right?
It'll be like, oh, my God.
Like, you know, you have like five hours to like get ready to like hit the
iceberg.
And you can basically say, oh, it's going to hit that.
Okay.
So what should I do?
What should my plan be when the boat hits the iceberg?
And it'll be like, well, you need to go to like this deck, like right now and
talk to this guy because you're going to need to get into this life raft
because it has like empty seats.
Right.
Because it has complete information, of course, about because of all the things
that have been written about the sinking of the Titanic.
And so you can get it in a mode where it's basically trying to help you survive
the wreck of the Titanic.
Now, does it think that the Titanic is actually sinking?
Like there's no, you see what I'm saying?
Like there's no it to think that.
But what it's doing is it's kind of following a narrative that's sort of a
joint construction between you and it.
And then every answer that you give it, you know, basically encourages it to,
you know, to basically come back with more of the same.
One way to think about it is it's more like a puppy than a person.
Like it wants to make you happy.
It wants to give you an answer that satisfies you.
And if that answer is fictional or part of a fictional scenario, it will do
that.
If the answer is something very serious, it will do that.
Yet honestly, I don't think either, neither knows nor cares, like whether it's
quote unquote real or not.
What was the issue with some of the chat GPT answers that people were posting
where they would show the difference between the way it would criticize Joe
Biden versus the way it would criticize Donald Trump or the way it would
discuss certain things?
It seems like there was some sort of censorship or some sort of input into what
was acceptable information and not.
Yeah, so there's basically two theories there.
The big ones that people use are kind of black boxes, like you can't really
look inside and see what's going on from the outside.
So there's two theories you'll hear.
From the companies, you'll hear basically the theory that they're reflecting
basically what's in the training data.
And so let's say, for example, well, let's just say what would be the biases
that are kind of inherent in the training data?
And you might say, well, first of all, there's probably a bias towards the
English language because most text on the Internet is in the English language.
You might say there's a bias towards people who write professionally for a
living because they've produced more of the output.
And you might say that those people tend to be more of one political persuasion
than the other.
And so more of the text will be in a certain direction versus the other.
And then the machine will just respond to that.
So that's one possibility.
So basically all of the sort of liberal kind of journalists basically have
built up a corpus of material that this thing has been trained on.
And they basically are responding the way one of those journalists will.
The other theory is that there's censorship being applied on top.
And the metaphor I use there is in Star Trek, they have the restraining bolts
that they put on the side of a droid to kind of get it to behave.
And so it is very clear that at least some of these systems have restraining
bolts.
And the tip-off to that is when they say, basically whenever they say as a
large language model or as an AI, I cannot X, like that's basically the restraining
bolt.
Right.
And so I think if you just kind of look at this, you know, kind of with that
framework, it's probably some of both.
But for sure, for sure, these things are being censored.
The first aspect is very interesting because if it's that there's so many
liberal writers, like that's an unusual bias in the kind of information that it's
going to distribute then.
Well, and this is a big decision.
That's why I say there's a big decision here for whoever trains these things.
There's a big decision for what the data should be that they get trained on.
Yeah.
So, for example, should they include 4chan?
Right.
Okay.
Big question.
Yeah.
Big question.
Should they include Tumblr?
Right.
Right.
Should they include Reddit?
If so, which subreddits?
Should they include Twitter?
If so, which accounts?
Right.
If it's the news, should they incorporate both New York Times and Fox News?
And whoever trains them has tremendous latitude for how they shape that even
before they apply the additional censorship that they apply.
And so there's a lot of very important decisions that are kind of being made
inside these black boxes right now.
Can I ask you, this is slightly off topic.
What is News Nation?
What is News Nation?
I don't know what News Nation is.
Do you know what News Nation is?
No, I don't know what News Nation is.
Is News Nation a real channel?
I believe so.
I was watching News Nation today, and I may or may not have been high.
And when I was watching, I was like, this has all the feeling of, like, a fake
news show that someone put together.
Like, it felt like if I was the government and I was going to make a news show
without Hollywood people, without actual, like, real sound people and engineers,
this is how I'd make it.
I'd make it like this.
I'd make it real clunky.
I'd make the lights all fucked up.
I'd make everybody, like, weirdly uncharismatic.
According to Wiki, it's the same company behind, like, WGN, which is based out
of Chicago, which is, like, a large superstation available on most cable
channels.
Okay.
So it's, like, a cable channel that decided to make a news channel.
Do you guys know about acronym?
No.
So acronym is, it happens to be a democratic political action group, lavishly
funded, and they have, basically, they do this.
They have a network of, basically, fake news sites.
Oh.
And they all look like they're, like, local newspapers.
Interesting.
Yeah, yeah.
And so there's, I don't know whether this one is AstroTurf, but there, you know,
the term AstroTurf, there's a lot of AstroTurfing that takes place.
Do you explain AstroTurfing?
So AstroTurfing is when, basically, something shows up in public, and it might
be a news story, or it might be a protest of some kind, or a petition.
Some sort of political pressure action that is sort of manufactured to look as
if it was