The Future According to Jaron Lanier
Released on 12/07/2023
I am here to introduce our next speaker.
So he has been working in tech for decades.
In fact, Wired magazine profiled him
our first year of existence in 1993.
He led the team that developed
the first multi-person virtual world and avatars
and he has been a vocal critic of the tech industry
and social media,
even writing a book in 2018 called Ten Arguments
for Deleting Your Social Media Accounts Right Now.
[audience cheers] [audience applauds]
I agree.
Here to talk about the disruptive rise of AI,
about virtual realty,
and about whatever else we want to talk about
is the fantastic Jaron Lanier.
Welcome.
[audience applauds]
[exciting music]
Hey.
Hello.
How's it going?
It's going, how are you?
Delirious.
Having a good day?
Yeah, yeah.
Wonderful.
I think we all are too.
Well, let me start.
I am gonna start by backing up a little bit.
I was getting ready for this conversation,
and I read an interview that you did actually
at another Wired event five years ago in 2018
with Nick Thompson,
who was the editor and chief back then.
And you said something about optimism
that really resonates with me that I loved.
I'm gonna read it back to you.
I'm so sorry.
Okay.
You said, To me, criticism
and optimism are the same thing.
When you criticize things,
it's because you think they can be improved.
It's the complacent person or the fanatic
who's the true pessimist
because they feel they already have the answer.
It's the people who think that things are open ended
that things can still be changed through thought,
through creativity,
those are the true optimists.
And I loved that and I think that that really speaks
to something very important and fundamental
about what we do at Wired
and what Wired at its best should aspire to.
I thought you summed that up really well.
And I'm curious, thinking about optimism, big picture,
when you think about Silicon Valley,
when you think about the tech industry,
I historically think of that industry
as a place of almost unbridled optimism, right?
The pursuit of advancement
without maybe adequate critical thinking.
How have you seen the industry change in that regard
in the decades that you have been working in it.
Mmm.
Ooh.
Yeah, that's a hard question
'cause there's been so many changes.
I've been doing this for over 40 years,
which is kind of freaky to say.
When I had my first startups in the '80s,
Silicon Valley was much less structured.
It really...
Everybody who did anything had to just make up how to do it.
It wasn't really like there was this established set
of venture capital firms or practices
or the usual way things go
or the usual people you have to impress.
And so the biggest change that I've seen is
it's become more of a who you know kind of a place,
like France. [laughs]
Like, Oh, I know these people on Sand Hill Road,
or, I know these people.
And it wasn't like that before.
That's not entirely a bad transformation.
The level of chaos and just complete confusion
in that period was actually dysfunctional.
So I'm not saying it's...
But it's definitely a major change.
And so now it's much more structured.
Now if you wanna come in and have a startup,
you know who you need to know,
how you need to do it, what it looks like.
I remember when I had startups,
there was this question,
what's an exit strategy,
and there wasn't really a clear idea.
These days you know what it is.
Yeah.
Typically, getting bought by one of the titans
at the end of the day or something like that.
In those days, it was a lot weirder.
Like why are we doing this?
How is this gonna work?
A virtual reality startup in 1983
is an unlikely IPO candidate, let's say. [laughs]
Yeah.
Hey, can I tell you a story about that?
Yes.
Recently, some of my friends
and former students who work at Apple,
which is about to introduce a headset,
you might have heard,
they got in touch and reminded me of this cool little thing
that they have in their archives,
which is back in the early '80s,
some of the people who worked
on the first version of the Macintosh
also worked on the first version of virtual reality
at my startup.
Andy Hertzfeld, people like that.
Is Andy here, I wonder?
No, okay, anyway.
So at the time, I had a conversation with Steve Jobs
about predicting what year Apple would sell
its first virtual reality headset.
And we even made an Apple virtual world,
which I have a picture of,
using the absolute cutting-edge graphics of the time,
which allowed us to have 20 polygons to make an apple.
And the year we came up with was 2011,
which I think, given the time involved,
is actually not that inaccurate.
So virtual reality has always felt like,
as long as I've been working in tech journalism,
the technology that's always just about to be mainstream.
It's always just about to be there.
There are these ebbs and flows.
Where are we in that right now?
Well, yeah.
In my opinion, which is informed on this matter.
I would say, yeah.
Fair enough.
Most people have misinterpreted
what the biggest challenge is.
So the usual interpretation about the biggest challenge
is that it's hardware.
It's that the damn headsets are too awkward or something
or the images aren't quite good enough or something.
That's all true.
I'm not denying that.
Although I gotta say, for myself,
I think this idea of making the headset smaller and smaller
until they look like glasses is probably a mistake.
I feel like if you're in virtual reality, you should own it.
Just go big. It's just this giant thing
with feathers and weird lightning bolts.
No, because you don't see somebody on a cool motorcycle
trying to pretend it's a bicycle, right?
Sure. [laughs]
Own it.
Just really live it up and make the thing incredible.
So I have a different attitude
about headset design than most people,
but, but, but that's not the main issue anyway.
The main issue is every time you make a virtual world,
it's too perishable.
So I'll give you an example.
The first time I saw a really cool virtual reality
experiential lesson in general relativity was '92.
That thing worked for like a half a year,
and it stopped working
'cause there wasn't any standard to keep to,
and the people who made it graduated and got jobs,
and it just went away.
Okay.
I don't even think it would be possible
to reconstruct it now with an infinite budget.
It's just gone.
Gone, gone, gone.
All right, so I've seen that happen
with relativity experiential lessons
like a dozen times since then,
over and over and over again.
Roughly concurrent with the lifespan of Wired magazine.
[Jaron laughs]
But you can still look at a back issue of Wired.
Old VR things are just gone,
and the reason why is that it's just too hard
to keep up with the platform
that has too many different layers that can all change.
So what I'm hoping is that
this new era of prompt engineering will give us a way
to make VR content that's malleable enough
to keep up with the platforms as they keep changing.
Nobody's pulled that off yet,
but I think now we finally have a fighting chance.
And if you can have persistent adaptable longer term rewards
for doing VR development,
I think that's actually the most important thing.
So we'll see.
Maybe that's too optimistic, but that's my current take.
I like that.
And let me ask you.
So it was about a year ago, ChatGPT was released.
The world went crazy, right?
It felt like generative AI was just,
all of a sudden it existed,
and it was all anybody could talk about.
You have a knack for seeing around corners.
Did you see that coming,
where it felt like a lot of the world
was really taken by surprise?
Yeah, so for those of you who don't know,
I'm also called the prime unifying scientist at Microsoft,
so I'm in the middle of all that stuff too.
So here's an interesting thing.
The level of GPT performance that was available
in the ChatGPT of a year ago had been previously available,
but with a different user experience.
But that chat user experience somehow,
that experience mattered to people in a new way.
And it suddenly,
there was suddenly this very strong response to that shift
in the experience of that level of capability.
I think it's very...
I could convince you that I know why. [chuckles]
I'm absolutely convinced that I could convince you,
but I'd be lying.
I think there's something... [laughs]
I'm not lying about that.
I'm a better journalist than that.
But I do think there is ultimately
a little bit of a mystery
in how people respond to different things.
I don't think...
I think anyone who claims that they can always predict it
is probably fooling themselves.
For whatever reason,
it turns out that this chat type of experience
just really reaches people in this profound way.
And part of it is just the familiarity of the mode
because chat is casual communication.
Part of it is
in order for people to communicate with one another,
we have to be very forgiving of one another.
And so there's a natural way
that the user becomes a little more forgiving in that mode.
Well, how do you talk about generative AI?
I think there's a lot of hysteria.
There's a lot of fuzziness and murkiness.
How do you think it's helpful for people to think about it
and talk about it and conceive of it?
Well, if you are willing to let me talk for a second,
I'm actually- I am.
I'm play testing a new way of describing
how big model generative AI works,
and I'll try it out on you.
I've done it I think three or four times now for audiences.
This is a couple weeks old, and I think it's working.
So when we go about our lives in modernity,
we all have these little cartoon images
of how technologies work that we're not expert in.
Like I have a little cartoon model in my head
of how vaccines work,
and I have a little cartoon model in my head
of how rockets work,
and I know a little bit about both things,
but not enough to reproduce them myself certainly.
But the thing is, the cartoon models,
while they're inevitably inaccurate,
they generate enough reasonably good intuitions
that you can have a conversation
with people who do about them.
And that's I think absolutely essential
to participate in modernity,
and not just be a recipient,
but to be a citizen of modernity, right?
But we haven't really done that for AI.
So here's my try.
Are you ready?
We're ready.
Are you alive, audience?
Are you?
Come on.
Are you ready? Okay.
We're gonna do this in three steps.
Probably a lot of people in the audience
will know this stuff already,
but suffer, if you're hearing stuff you already know.
We're gonna start with step one,
which is how do you tell a cat from a dog.
All right, so this was a big thing like a dozen years ago.
Cats and dogs are similar.
Can a program tell a picture of one
from a picture of another, right?
So if you take just a few measurements,
like can you find the eyes or something,
that doesn't work 'cause they both have eyes,
they both have snouts.
That kind of thing just doesn't get you anywhere.
If you take a whole bunch of little statistical measurements
like how blue is the corner
and how many parallel lines are there in the center,
whatever it is, just some really simple stuff,
that doesn't tell you.
But let's say you take all those little measurements.
We're gonna give them a ridiculous name.
We're gonna call each measurement a neuron, okay?
I'm saying ridiculous 'cause we don't even know
how biological neurons really, really work for sure,
but we're gonna call these things neurons.
All right, now, that doesn't work.
But then you put up another layer,
and that one does measurements of the previous layer.
And then you put up another one,
and you make a big pile of them like a skyscraper of layers
of measurements of measurements of measurements.
It turns out if you make a big pile like that,
it can start to work.
That's called deep.
So when you hear deep learning,
that means there's a bunch of layers, right?
Doesn't work at first.
You have to train it.
What training means is that you keep on running it
and telling how well it did.
And whenever it does better,
you add a little, what's called a weight,
a little indicator to some of the neurons
that they're doing a good job.
And then you can pull back some of the other ones.
And once you do enough of that,
you have a trained model, okay?
And that works.
Now, it might feel weird that that should work.
Two things about that.
One is we don't really understand why it works in detail.
We don't...
We're just starting to have in some special cases,
some ways of understanding why one weight matters
versus another and what it means.
That's all very new.
It's kind of a craft in a way.
It's a technique that works.
But the thing I wanna suggest is that
it would be even weirder if it didn't work.
Wouldn't it be weird if a bunch of statistical measurements
couldn't tell a dog from a cat?
It should work at some point 'cause statistics is real.
All you need to do is find the right configuration
to get you there, and this is it,
so far as we know.
So that's step one.
Is that all clear?
Step two, do that, but for everything.
[audience laughs]
So what you do is you take in the whole internet,
and every other piece of data you can possibly grab,
and then you think that adjacency
is gonna have some correlation with meaning.
So if there's some text next to an image,
maybe the text and the image
have something to do with one another.
So in the OpenAI and Microsoft world,
the thing that does all that is called CLIP.
We take in literally billions of these examples,
and we try to do exactly the same thing I said
about cats and dogs
for everything everywhere.
And that's called the large model, okay?
So now instead of just one tower of these things
that can tell cats from dogs,
you have this huge mush of towers
that are there in the waiting.
And when you need a tower, it might be there.
It probably will be there if there were enough examples.
So one of the towers might distinguish
whether a language is pirate talk or not,
and another one might distinguish
whether it's a flying saucer or not, whatever.
And we don't even really know what all's in there
because it's too giant.
It's this huge, huge, huge big thing.
Super expensive to calculate, but that's what we do.
That's the large model.
Okay.
Ready for step three?
[Audience] Yeah.
Step three, you run the thing in reverse.
And I'm gonna do this with images,
with something analogous happens,
with text and music and computer code and everything.
So with images, we're gonna start
with a field of snow of randomness, right?
And now what you're gonna do is
you're gonna add some more randomness to the randomness.
Now that sounds useless, doesn't it?
Yes, it does.
It does, but it isn't.
And the reason why [chuckles] is you add some randomness,
then you're gonna run it by one of these towers.
Let's say the one that can find cats.
And if the newly,
the different randomness is a little bit better
at finding a cat,
it triggers the cat detector.
We call this a classifier.
A cattifier.
Anyway, if it starts pinging you, save that.
And if it gets worse, you throw it away.
And then you do that over and over and over again,
and out of the snow emerges a cat.
Okay, now, and that works,
and it'd be weird if it didn't work, right?
So now we're at the point of generation.
But we're not done.
The magic is that you can have it solve
for multiple classifiers at once.
And this is where we hit something new.
Now, you've probably seen some people,
like my friends Timnit Gebru or... [hand slaps]
[groans] Forgetting his name.
Another friend of mine wrote a piece in the New Yorker,
your sister mag,
talking about how all AI can do is regurgitate randomly.
Timnit called it stochastic parrots.
Is Timnit here by any chance?
Anyway, so this is true.
This is what it does.
But the trick, the thing that's really new is that
you can solve for multiple classifiers at once.
So you can say, Can you give me a cat
that's using a parachute and playing the banjo
and is dressed in Western wear,
and the whole thing's in watercolor, or something.
Just the kinds of crazy prompts
that people send to programs like DALL-E
or Midjourney or Stable.
You do that, and it's round robinning through all of them.
It has to address all of them at once
for the image to come out of the snow.
And here's what's amazing.
It has to stochastically explore ways of addressing
all the classifiers at the same time.
And in doing that, it solves problems
and it's a form of creativity.
If a cat's wearing a parachute,
how exactly does the parachute go on the cat?
What is the cat's posture?
How exactly does a cat hold a banjo?
All right.
No, these are actual...
If you were doing that,
you'd have to think through those things.
This happens just in order
to meet all the classifiers generally.
Sometimes it just screws up and gives you nonsense,
but generally it starts to find these things.
So what I propose here is that we can understand
what generative AI does,
and here's where this is gonna get into something new,
as filling in the space between the towers.
So what happens is you have one tower for cat,
one tower for parachutes, another tower for banjos.
Previously, there was a space between them,
and now, in order to generate,
it fills in that space with something
that meets all of them at once.
So what's nice about this intuition is
it both shows what's special about this,
but also its limits.
It goes up to the level of the towers,
but it's unlikely to go much higher.
Now, I know some people who are very pro-AI,
whatever that means, will say,
But do people go above the towers?
I don't know.
That's not my point here.
The point is to talk about what the technology can do.
And so what the technology can do is
it can fill in between the towers, and yet,
we shouldn't expect it to grow beyond the towers.
So the people are saying,
Oh, it's gonna become super intelligent
and solve global warming,
so we don't have to worry about that.
Let's just build AI.
No.
All it can do is reconcile towers
that we've already defined,
that we've already worked on enough that they're solid.
Now, that's an incredible value.
When you program, you're inevitably tediously
reconstructing something similar to things
that people have done millions of times before,
and it's painfully tedious.
Well, with generative AI,
you can leverage all those previous times
and get a version that matches your circumstance
and saves you a bunch of time.
That's filling in the space between the towers.
So AI does something,
filling in the space between the towers,
but not everything,
which is climbing way up above the existing towers.
And there you are.
I like it.
I like it very much, yes.
[audience applauds]
Let me ask you a little bit about that.
Using that framework, what about the data,
the inputs that actually make up the towers
and the idea of ownership of that material?
That's something that we've talked about
on other panels today.
I know you've talked a lot about the idea of data dignity.
Can you talk a little bit and offer your critique
on what's actually feeding these LLMs?
This is this other stream of work
that I've been working on for a long time, which is hard,
but I'm still really super hopeful about it.
So the idea here is that let's say you're concerned
about the quality of what generative models do,
and you might be concerned
'cause you're afraid they'll take over the planet
and kill all the people,
or you might just be concerned
'cause they'll discriminate against certain people,
or you might be concerned because you're just afraid
that they're not functional enough,
or there are all kinds of different flavors of concern.
Now, when we...
We spend so much time on that.
You wouldn't believe how much time.
Really, I think most people don't appreciate
what a priority the whole issues of safety and fairness
and so on have been.
It's a like a massive spend and a massive focus.
And I do wanna point out that so far,
the GPT world of AI has,
with the exception of telling journalists
to leave their wives occasionally,
or something like that,
has been pretty okay.
Nothing too terrible has happened so far.
Now, some might disagree with me about that.
And there are a couple of cases of it being applied
in the courts and in law enforcement
that are very concerning, I have to say.
So let me correct myself there.
But that wasn't our stuff, but even so.
So right now, if you look at the pipeline for big model AI,
there's the training data.
The training data is generally intelligible,
readable by people 'cause naturally that's what it is,
then there's the model you calculate,
which is generally impenetrable
and not directly readable by people.
And then there's the output, which is readable by people.
So you have readable, not readable, readable.
Now the problem is
because there's this big thing in the middle
that's not readable,
when we try to improve it,
we tend to try to sensor it
or channel it on the output side,
or we try to,
we try to filter what the training data is.
But there's this other alternative,
which is what if, when you're training,
you could trace the,
each segment of training data
and how it ended up in the big mush.
And then when you do generations,
you could say, For this particular generated output,
which were the most important segments of training data?
Currently, we can't do that,
or at least we can't do it efficiently.
But if we could, if we could,
we'd have this new way of having legible interpretation
and management of AI.
We could say,
if some AI tells a journalist to leave their spouse,
we could say, Well, that's because it was relying a lot
on this fan faction
and these soap opera transcripts or whatever.
And if we just suppress those,
then what we'll get back is something maybe a little better.
And if we're worried about the AI
learning how to operate their mind
and building weapons without our knowledge or something,
if we trace back which documents it's looking at,
we would be able to tell if it was making weapons.
It would just make the whole thing
genuinely readable and open.
So I'm trying to do that thing.
It's hard because tracing and tracing efficiently
are actually difficult technical problems,
but I believe they're solvable.
And I can see very clearly why there would be,
why that would be a logical thing to do, right?
And you can anticipate the problems
that that would solve longterm
by getting that right at the outset.
Do you feel confident that the companies...
I know you cannot speak for Microsoft,
but when we're looking at the landscape
of the companies building these generative AI tools,
do you feel like there is adoption of that philosophy,
that that is an intuitive way
to go about building this technology?
If it makes it better and it doesn't cost too much,
relative to how much better it makes it, yeah, sure.
Most of the actors in this space,
despite our occasional displays of craziness,
are fundamentally rational.
So I feel pretty optimistic about it.
Well, why don't we end on a positive note,
an optimistic note?
Let me just ask you.
When you look 30 years into the future,
paint us a picture about what could go right
in the context of where we are today.
What does a better future look like to you?
Answer that however you want.
Give us some good vibes.
I want a future
where we aren't entered into a nostalgia cycle
where our problems are solved based on
what we did before exclusively within big models
or any other technical or media structure,
but instead, this real emphasis on creativity.
I want a future where people are motivated
and rewarded and loved in part
because of the new data they add to the models
to make things more creative.
I want a future where things
that we think of being rote become new art forms.
I want the topiary of the future
to occasionally be holographic, according to fashion.
I want the people of the future to,
to feel that they're full citizens
and not just recipients of technology.
I wanna create new creative classes,
instead of dependent classes whenever possible,
so that whatever basic income scheme might show up is only
for exceptional tough cases,
and not for everybody all the time.
I want a future of eternal creativity and reinvention,
but also one in which people believe people are real,
where people treat people as special enough
that it's possible to create, be creative,
and also avoid cruelty.
Well, I will check back in with you in 30 years
to see how we did.
Thank you so much for being here.
Oh, yeah absolutely. It's always an honor
to talk to you. Hey.
RE:WIRED 2021: What Should the Future of AI Look Like?
RE:WIRED 2021: World-Building and World-Fixing: A Conversation with Author Neal Stephenson
May 2014 Business Issue: Trust Me
Will AI Enhance or Hack Humanity? - Fei-Fei Li & Yuval Noah Harari in Conversation with Nicholas Thompson
WIRED25 2020 Q&A: Protecting Health Care Systems From Cyberhackers
Breaking the Fourth Wall: From Virtual Stages to AI-Driven Narratives
The Future of Technology: Technology Supporting Accessibility | WIRED Brand Lab
Live WIRED
WIRED25 2020 Q&A: The Canaries in Our Environmental Coal Mines
RE:WIRED 2021: Speaking Truth to Biased Algorithms