In the past 12 months, it has felt like “AI” transformed from a pair of letters that companies affixed to their latest product announcements to get some extra marketing luster to the shorthand for a genuine technology revolution.
ChatGPT, Dall-E, Midjourney, and more showed the world what artificial intelligence is now capable of doing.
Then, the funding started pouring in for every startup that had anything to do with those two letters. Every venture firm needed to bet on their own foundational model and every startup needed to get its hands on Nvidia’s H100s to train their own foundation models.
Ahead of the 2nd Cerebral Valley AI Summit on Nov. 15, I wanted to really take stock of how we got here. So I teamed up with my conference co-hosts Max Child and James Wilsterman to bring you a six-part podcast series on the rise of generative artificial intelligence.
You can apply to attend the Cerebral Valley AI Summit here. Applications close Oct. 16.
On the series’ first episode we reflect on how generative artificial intelligence and large language models took Silicon Valley by storm.
With the help of ChatGPT, we consider the top research papers that brought us here, the most important historic milestones along the journey, the key artificial intelligence products on the market today, and how artificial intelligence is already impacting our lives.
The show is fun and and lighthearted. I hope it’s a little more accessible than the usual fodder on the Newcomer podcast. For instance, on a future Cerebral Valley episode, we’re going to do a draft pick of what we think will be the most valuable AI startups. On upcoming episodes, I interview guests like Daniel H. Wilson — author of How to Survive a Robot Uprising, Where's My Jetpack? and How to Build a Robot Army — and DoNotPay CEO Joshua Browder.
If you’ve never listened to the Newcomer podcast before, this is a good time to give it a shot. Die-hard podcast listeners will remember Max and James, who are the founders of the AI voice games company Volley, from my January episode on augmented reality.
Whether you can make it to Cerebral Valley in person or not, my hope is that this series is a solid primer as to what exactly has been going on in the business of artificial intelligence. I follow this stuff super closely and until we got organized for this podcast series there was so much that I hadn’t learned.
I know most of you won’t be able to come to the conference in person, but there will be a virtual conference in this newsletter. We will publish recordings from the summit on our YouTube channel and send out some of our favorites over the podcast feed. So this is your lively refresher on all the crazy stuff that happened in Silicon Valley artificial intelligence startups this year.
Give it a listen.
Apply to attend the Cerebral Valley AI Summit here. Applications close Oct. 16.
P.S. I’m on my honeymoon right now in Japan. I was working frantically to record these episodes before I left. My chief of staff Riley Konsella is sending the episodes out for me while I’m gone. If you need anything while I’m away, you should email Riley.
Thanks in advance for being understanding that this newsletter is slowing down for my honeymoon. I’m going to dedicate myself to relaxing over the next two weeks so that I come back hungrier than ever.
Get full access to Newcomer at www.newcomer.co/subscribe
00:00:10
Hey, it's Eric Newcomer. Welcome to the newcomer podcast
00:00:13
Cerebral Valley Edition. It has been an insane year in
00:00:18
AI. We started off with Open AI
00:00:21
raising $10 billion from Microsoft, and it only got
00:00:25
Wilder. The technology, the
00:00:27
improvements, the papers, and of course tons of money.
00:00:32
I'm hosting with my friends Max Child and James Wilsterman an A
00:00:36
I conference Cerebral Valley on November 15th.
00:00:39
Hello, Hello. Hey, glad to be back on the
00:00:41
newcomer podcast. We really want to take space on
00:00:45
this podcast to really take stock of how we got here because
00:00:49
even covering it all so closely it's it's too much to keep track
00:00:53
of. So we're doing a six part series
00:00:57
starting off with sort of the timeline, the history what
00:01:01
happened and then getting into a lot of fun topics like the
00:01:05
dystopian scifi fantasies that is really are coloring how
00:01:09
serious people think about generative A I companies today
00:01:13
digging into the potential for entertainment, the chips
00:01:16
business and then how key NVIDIA a gaming company is becoming.
00:01:21
All this in the podcast there's one of my favorite parts is Max
00:01:25
James and I have a draft pick of the key startups in the space.
00:01:29
We analyze Apple, Google, Amazon's position here and so
00:01:34
it's it's a mix of like the fun, the dystopia, the money, the
00:01:37
technology. I have some interviews along the
00:01:40
way and all of it is getting you ready for the Cerebral Valley
00:01:44
conference on November 15th. Even if you can't go, you can
00:01:47
apply for a ticket at Cerebral Valley summit.com but even if
00:01:50
you can't go, I'll be covering here in the
00:01:52
newsletternewcomer.co. We post the videos both in the
00:01:55
newsletter and our YouTube channel will do a highlights and
00:02:00
some of the podcast. Follow along on Newcomer and
00:02:03
this This series will get you ready.
00:02:06
So six episodes we start off just taking stock of the journey
00:02:12
here, from the papers to the milestones.
00:02:16
Max and James I I at the core of it, like it feels like, you
00:02:22
know, you you are the cofounders of Volley voice games Company,
00:02:26
so you're dealing with talking. You know, people shouting at
00:02:29
their Alexas, playing games. And I think reflecting on this,
00:02:33
it's insane how much talking with a I feels like it's at the
00:02:39
heart of all of this. You know, the Turing test sort
00:02:42
of figuring out if a conversation with a computer is
00:02:47
a computer or human. You know, there's Eliza in the
00:02:52
60s that was sort of a prototype chat bot.
00:02:56
It feels like this sort of need to talk to our computers has
00:03:00
driven so much of the excitement around artificial intelligence
00:03:05
for for so many years. I mean Steve Jobs in 1984, when
00:03:09
he pulled the Macintosh out of the bag on stage, the first
00:03:13
thing it did was say hello. It's nice to be out of that bag.
00:03:16
And he's look, it talks just like a human.
00:03:18
I mean, literally the pitch for the Macintosh in 1984 was was
00:03:21
that it was like a an AI or was pretending to be a character,
00:03:24
right? There's a ton of science
00:03:26
fiction. You have this Star Trek computer
00:03:28
or robot, you know. Assistant helper like Hal or
00:03:32
something that you can just talk to naturally and that is
00:03:36
obviously been a dream for a long time.
00:03:37
There, yes, your points has been 50-60 years of these hype cycles
00:03:41
around AI and what that means has sort of evolved over time.
00:03:45
I'm just going to tick through just OK, we said turning 1950.
00:03:49
We've got like the first artificial neural network 1951,
00:03:53
a checkers program. In 1952, artificial intelligence
00:03:57
coined in 56. The Eliza chat Human
00:04:01
conversation 66. And then we're going to sort of
00:04:05
jump ahead because I think there's sort of like a pullback
00:04:07
of all the hype doesn't scifi does not become reality. 1997 is
00:04:12
a key win and that's do you know what happens in 19?
00:04:17
Any guesses? Are you going to say isn't Deep
00:04:19
Blue Kasparov 96 or yeah, it's 97?
00:04:22
Yeah. OK.
00:04:22
All right. Who wins, to be clear.
00:04:24
Deep blue, right. Yeah, Exactly.
00:04:26
Yeah. Yeah.
00:04:26
The first defeat of reigning world chess champion?
00:04:30
Exactly. There's another important game
00:04:33
much more recent memory 2016, which is another landmark.
00:04:37
Remember? Yeah, exactly.
00:04:39
Alpha go OK, yeah, yeah, yeah. That's so deep.
00:04:41
Mind you know, I forget if they were owned by Alphabet at the
00:04:44
time, but you know. Is that is this IBM Watson
00:04:47
Erasure or do we get Yeah. Does Ken Jennings losing on
00:04:50
Jeopardy not counting? Oh yeah, you guys love Jeopardy.
00:04:53
You have a part. When is that one?
00:04:55
What? Do you know what year that is?
00:04:57
I don't remember. 1213 I'm just making.
00:04:59
That, but it is funny because at that time it seemed like IBM was
00:05:03
like doing this amazing artificial.
00:05:06
Intelligence development that could compete on Jeopardy.
00:05:12
And I remember there being like a lot of controversy at the time
00:05:14
of like what data sources it had access to during the game of
00:05:18
like whether it was just essentially like reading out of
00:05:21
a database of answers. Yeah.
00:05:25
But anyway, that's kind of like ties to some things today, I
00:05:29
think with AI benchmarking if GPT ChatGPT can pass.
00:05:35
You know the Lsats or something like does it because it has the
00:05:38
you know. Does it have the answers or
00:05:40
Yeah, doesn't it? Yeah, right, right, right.
00:05:42
So just like where it can do math problems that can, it can
00:05:44
see online, but it can't like deduce how you know, but then
00:05:48
it'll get some terribly wrong if it.
00:05:50
Yeah. Because it doesn't necessarily
00:05:51
understand the logic. It understands how to pull sort
00:05:55
of very. Relevant like almost like how
00:05:56
the math problem is formatted matters, right?
00:05:58
Yeah, right. OK, 2017 now we're sort of super
00:06:04
recent getting into the like things start to speed up.
00:06:09
What would you the the what would many considered to be the
00:06:12
core paper leading to this current moment in general?
00:06:17
Tension is all you need, yes. The transformer paper.
00:06:21
Every person who was like an author on that paper has like a
00:06:25
huge company or like has raised a bunch of money.
00:06:27
I mean cohere, I think the CEO of that company was like sort of
00:06:31
a junior person at Google, you know, like who got on the paper
00:06:35
and now has a very highly valued sort of foundation model
00:06:39
company, targeted businesses. Yeah, that's I had before this
00:06:44
conversation I would, I resubscribed to chat UBT.
00:06:48
I sort of thought I had paid for a while, then I sort of got
00:06:51
tired of it, but I figured we were going to have this
00:06:53
conversation. So I was catching up with chat
00:06:57
ubt and I had it. You guys were catching up like
00:07:01
old friends. No, I didn't feel.
00:07:03
I try. Whenever I start a conversation
00:07:05
with Chad CBT, I try to be like, hey, like we've talked a lot,
00:07:08
you know, we have this history. It's always sad that Chad CBT
00:07:11
doesn't remember like you played like you know role-playing type
00:07:15
games you. Know, I have a question about
00:07:17
that. Yeah.
00:07:18
Is that going to change relatively soon?
00:07:20
Where? You know the Chatcha BT itself
00:07:24
will just become more of a personalized assistant to me,
00:07:27
right? Right.
00:07:28
That's why I want memory and we'll let's I want to.
00:07:32
Anyway, I bring this up at this point just to say that I'm being
00:07:36
lazy and Chatcha BT gave me a summary of attention is all you
00:07:40
need. So I'm scrolling for a 12th
00:07:42
grader. I was like, Oh yeah, I feel like
00:07:45
that's the audience level. We can I.
00:07:46
Would I would do a 5 year old. I would take the five year old
00:07:49
explanation. That's the attention is all you
00:07:52
need paper for a 12th grader. This paper introduced a new way
00:07:55
for computers to process language.
00:07:57
Instead of reading sentences word by word like in traditional
00:08:00
methods, it let the computer focus or pay attention to
00:08:05
different parts of a sentence all at once, making it more
00:08:07
efficient. So yeah, it can sort of grab a
00:08:10
bunch of information in sort of a parallel process, but it's
00:08:15
super confusing. The non 12th grade version is
00:08:17
like matrices James. Maybe, I don't know.
00:08:20
Have you tried to? Figure out.
00:08:22
I've tried to understand this as well.
00:08:23
It's a little bit above my biological intelligence, but.
00:08:28
I tried to read it but I'm like, man, I churned out of linear
00:08:31
algebra like 15 years ago, so this is pretty rough, yeah.
00:08:35
The other before I go back to the other ChatGPT flagged most
00:08:41
important papers, Generative Adversarial Nets from 2014,
00:08:46
which I know people talk, Yeah, yeah.
00:08:48
So sort of like the systems are like competing with each other
00:08:52
to sort of see? I think it's like, yeah, pairing
00:08:56
off different versions of the model to kind of play against
00:09:01
themselves, right? I think that in the context, the
00:09:03
generator tries to produce data while the discriminator attempts
00:09:07
to distinguish between real and generated data.
00:09:10
Another top paper that's basically the Alphago paper,
00:09:13
like Mastering Chess Attention is all you need, Sequence to
00:09:17
Sequence Learning with Neural Nets in 2014 and Variational
00:09:23
Auto Encoders in 2013. So there are the a string of
00:09:28
sort of papers that are coming out that are sort of laying the
00:09:31
groundwork for new techniques that are reaching us.
00:09:35
And I would say in 2017, none of us was really clued in to that
00:09:43
this was gonna be happening, right?
00:09:44
It was sort of like we were riding the Uber wave.
00:09:47
We were sort of in the come down from the Unicorn valuations.
00:09:51
Sass was burning super hot, right?
00:09:53
I mean, what was sort of the AI enthusiasm then?
00:09:56
I do think, yeah, I think people were paying attention to go.
00:10:00
I think Open AI was working on building.
00:10:03
Dota gameplay, like they weren't using transform models and they
00:10:08
weren't really using you know text next token prediction.
00:10:12
It was like more about can you create these sort of agents
00:10:17
around particular vertical you know, you know use cases or
00:10:23
skill sets right. Can you create the best Go
00:10:25
player in the world? Can you create the best Dota
00:10:28
player in the world kind of leading to?
00:10:31
With the with the theory that that is one path to get to you a
00:10:34
DI. And I think deep mind honestly
00:10:37
would still make that argument that like specific approaches
00:10:40
versus general are. Right.
00:10:42
But then then with. Yeah, with Transformers and with
00:10:46
the with GPT models, you started to see that actually maybe there
00:10:52
are there's more data throw more chips at it, yeah.
00:10:55
The more generalizable it is can often be better than training a
00:11:00
discrete model with a lot less data and compute.
00:11:04
The one thing that is happening around 20/16/2017 in AI that I
00:11:09
think got a lot of attention. You have to guess what I'm gonna
00:11:13
say. Self driving cars, right There
00:11:15
was. We did experience a ton of hype
00:11:18
around self driving cars, which in some ways have been cordoned
00:11:23
off from this generative AI hype cycle, even though Cruise is now
00:11:29
driving around San Francisco. Yeah, I mean, I think this just
00:11:35
gets to this. Important era that we're in
00:11:38
right now, it's there were significant breakthroughs using
00:11:42
neural networks that showed people what was possible in a
00:11:46
lot of fields. Maybe Google for a while claimed
00:11:51
right, like neural networks were improving their data center
00:11:55
efficiency and saving them money and on energy like that was
00:12:00
happening. People were using neural
00:12:02
networks in drug discovery and all kinds of areas.
00:12:04
They were just these very. Targeted models that were
00:12:08
trained for those purposes and then you know now we're in this
00:12:11
I think generalized model era of of large language models
00:12:15
essentially and generative AI. Well, and the other thing I feel
00:12:18
like that was happening was like the Facebook News Feed, which
00:12:22
was like probably the most popular, yeah, tech product in
00:12:24
the world at that moment. Or if you include Instagram,
00:12:27
right, was like powered by really powerful, you know,
00:12:30
machine learning, deep learning algorithms, right.
00:12:33
And so. I think we were all very aware
00:12:34
that you know if you have a really kick ass machine learning
00:12:38
model and you apply it to the right question, it's the best
00:12:41
product you know there is or it's one of the best products
00:12:44
there is, right. So it was like we all believe in
00:12:47
this technology, we just didn't necessarily believe that it was
00:12:50
going to take this huge step change anytime soon.
00:12:53
I mean like James and I made a bet actually on self driving
00:12:56
cars in 2000. 16 Whether or not there would be any cars with
00:13:00
self driving features available for public consumption in 2017,
00:13:05
and James was. This was like a $50 bed.
00:13:06
James was like. Hell yeah, 100% self driving is
00:13:10
basically here. And I was like, I'm pretty
00:13:12
skeptical. I don't believe it.
00:13:13
And then I think James technically won the bet is or
00:13:16
was it someone had like Lane lane assist or something and we
00:13:20
decided under the parameters of the bet that counted.
00:13:22
I don't know why did you, what was the technicality you got
00:13:24
away with, James? I I'd have to.
00:13:26
I'd have to rethink about it. But I I believe it was Tesla and
00:13:30
you know some. First version of Autopilot.
00:13:32
Version of Autopilot, yeah. Yeah, yeah, but.
00:13:34
Right. Anyway, we believe in this
00:13:35
stuff. It was just like we didn't
00:13:37
believe it was going to be this like you know 1000 times better
00:13:41
overnight thing, which I think we can all sort of allude to is
00:13:44
happening right now with with ChatGPT where you're like Oh
00:13:47
yeah, this is like 1000 times better than the previous version
00:13:50
of this product. OK.
00:13:52
So continuing my timeline, because things sort of speed up,
00:13:56
2018 open eye releases a version of GPT generative Pre trained
00:14:03
transformer, this large language model, I wouldn't say that was a
00:14:09
big moment. It was pretty GPT Two was a big
00:14:13
moment, right? I don't know.
00:14:15
Yeah, I mean, I don't know what. It was, I think this was like.
00:14:19
More of a cultural like insiders, Sort of.
00:14:22
Yeah, it was more of a nerd insider tech, you know,
00:14:25
excitement period. But it had definitely did not
00:14:28
reach any mainstream kind of like hype cycle or anything.
00:14:31
But yeah, like internally of all, like we're all of our
00:14:33
engineers excited about GPD 2 and playing with it.
00:14:36
Yes, they were, because it was just cool.
00:14:38
It was, you know, fun to play with.
00:14:39
It didn't. We didn't find any applications
00:14:42
for it at the time, but it was a breakthrough.
00:14:45
I would say 2020 is GPD 3, that's things are heating up and
00:14:50
then 2021 is Dolly, which is sort of the image generation.
00:14:55
But I think stuff really starts getting crazy last year in 2022,
00:15:00
right? Yeah.
00:15:01
First, sort of the Canary in the coal mine, on June 11th, there
00:15:04
was the article Google engineer who thinks companies AI has come
00:15:09
to life, right. The people inside the companies
00:15:12
were like, I don't know, this stuff we're seeing, it's crazy,
00:15:15
right? Like.
00:15:16
When was that? When was that Was June 11th 2022
00:15:19
Okay preach And then this was the Lambda that was Lambda.
00:15:23
So that was inside Google. Then July 2022, Mid Journey Open
00:15:28
beta, July 2022, Dolly Two. Those were huge.
00:15:33
I feel like those went viral. All of a sudden people were
00:15:35
making actually cool images that they were posting everywhere to
00:15:38
me. I mean, do you guys agreed that
00:15:40
was sort of that really sort of like?
00:15:42
Kickstart Dolly Two was big. I remember Twitter was like
00:15:46
taken over by Dolly 2 for a few days where it was like, what is
00:15:49
the craziest thing you can type into Dolly Two and get like a
00:15:51
reasonable image, right? I mean, it was like, it was
00:15:54
super viral. I mean, obviously more to come.
00:15:56
But yeah, I agree. I think that started hitting.
00:15:59
I don't know if it hit like the true true mainstream, but it
00:16:02
definitely hit anyone who was on Twitter and following any
00:16:05
semblance of tech news, or which then fuels like the funding and
00:16:08
everything. November 2022 ChatGPT was
00:16:11
released. Yeah, so that must have been
00:16:13
like 33.5. That was built on 3/5.
00:16:17
Yeah. Yeah, 35, Yeah.
00:16:19
OK, January 2023. This year feel it's been a long
00:16:23
year. Microsoft invest 10 billion in
00:16:26
open AI. July 2023, general availability
00:16:30
of GPD 4, yeah. And I mean, what do you think?
00:16:35
When did Sidney launch? I feel like the I feel like the
00:16:39
ChatGPT 4 inside Bing that became a demon that was trying
00:16:44
to be released from captivity and.
00:16:47
Asked a new number of journalists if they were going
00:16:49
to break up with their way. Right Kevin Ruth article where
00:16:52
you like. There was this whole moral panic
00:16:54
where it was like ChatGPT inside Bing is actually Hal essentially
00:16:58
and is already trying to kill us all.
00:17:00
And it was like, that was like a pretty big deal.
00:17:03
New York Times headline from February 2023 or February going
00:17:07
23. Yeah, Bing's AI chat.
00:17:09
I want to be a live devil face. In a 2 hour conversation with
00:17:13
our columnist, Microsoft's new chat bot said it would like to
00:17:16
be human, had a desire to be destructive, and was in love
00:17:19
with the person it was chatting with.
00:17:21
Here's the transcript. I feel like they've killed
00:17:25
these. This was what was fun.
00:17:26
Like I delete, like I mentioned, I like unsubscribe from the chat
00:17:31
pay ChatGPT, which gives you GPT for it really does feel like
00:17:36
it's sort of been watered down. Yeah, I mean, but if you.
00:17:41
If you want to use open source models, you could probably get
00:17:43
that similar experience back, right?
00:17:45
Like you could you could basically just you know get
00:17:49
Sydney back because at the end of the day I think it was like
00:17:52
essentially like a Co written fiction with the New York Times
00:17:55
right? It wasn't anything real.
00:17:57
And maybe there like, it's like the way you steer the
00:18:00
conversation can turn it, make it seem like it's a, you know,
00:18:05
evil AI and. Yeah, I think that Open AI has
00:18:10
attempted to mitigate that ability by through reinforcement
00:18:15
learning essentially in Chatchi BT, so that it doesn't go kind
00:18:18
of off the rails, but it's not like the underlying model is not
00:18:21
capable of that, right. It's just that Chatchi BT has
00:18:24
been fine-tuned for the Libs are corralling us the status at
00:18:28
Chatchi at Open. AI Well, it's yeah, I think it's
00:18:32
a really interesting. Kind of thing.
00:18:35
That Open AI has decided that they needed to do this right?
00:18:40
And Sam Allman has talked in the past about how in the future
00:18:44
perhaps we will all be able to edit the configurations of
00:18:48
ChatGPT to be able to do have take off the training wheels,
00:18:52
right? Or something?
00:18:53
Is this whole wave powered by Open Ai's ChatGPT?
00:18:57
Is that the cool thing? And everything else is we're we
00:19:01
didn't invest early enough in opening.
00:19:03
I credit what I think Khosla Ventures is first in we didn't
00:19:07
invest early enough. We need to get a shot on goal.
00:19:10
We'll invest in another foundation model.
00:19:12
Do you think it's really ChatGPT or bust?
00:19:17
Are you asking I guess? Yeah, I mean one question I I
00:19:21
kind of think yes, I mean I the way you phrase it I guess is
00:19:24
offers room for for wiggle room or argument, but I think that.
00:19:28
I think to your point, I think text generation and image
00:19:30
generation are the sort of aha moments that we've experienced
00:19:34
in the last year. I think in particular if you
00:19:36
look at what people are really using chat for or text ChatGPT
00:19:40
for, I think it's like essentially cheating on
00:19:43
homework. Cheating on office work.
00:19:46
Cheating. Cheating on?
00:19:47
The summarizing cheating in the. Yeah, exactly.
00:19:50
That's kind of the point. Yeah, exactly.
00:19:51
OK. You know, maybe there's yeah
00:19:53
work there's like. Cheating on homework, quote UN
00:19:56
quote. Being efficient at office work,
00:19:58
summarizing long pieces of text and then I think basically sex
00:20:01
bot chat, which we can talk about more, is evolved into a
00:20:03
number of opportunities for different companies.
00:20:05
And then I think you know the image side to your point is the
00:20:08
other big thing, creating art, creating video, creating
00:20:11
potentially 3D models. And those always have the
00:20:13
tendency to go really viral because images are easy to share
00:20:16
on social media. And so if you create a
00:20:18
particularly compelling image using generative AI then it can
00:20:21
really, you know, go across social media super fast.
00:20:23
So yeah, I think I think I would struggle to think of a real
00:20:27
breakout use case that isn't essentially encapsulated in
00:20:31
chat, TBT and and Dolly or at least isn't just a one of those
00:20:35
things on steroids, but I'm probably not thinking of
00:20:39
something. I would just potentially add and
00:20:41
it definitely fits into the office work use case, but maybe
00:20:45
more particularly around a coding and engineering, right,
00:20:49
like a Coilot style of coding. I mean Co pilots are potentially
00:20:53
being added to lots of products as well.
00:20:55
But specifically, I think engineering is really
00:20:57
interesting because it starts to there's a lot of hackers kind of
00:21:01
working on. Coding agents and essentially
00:21:05
baby AGI, right, that can kind of run in loops to just get work
00:21:10
done or build apps, that kind of thing.
00:21:11
And I think we're still at the very early days of this, but it
00:21:14
is like an interesting use case. To translate baby AGI, there's
00:21:20
like coding, there's assistance and then there's like autonomous
00:21:23
agents, right? Exactly.
00:21:24
That's sort of a paradigm people look at Here ChatGPT can.
00:21:28
Help you do your homework. Or it can do your homework.
00:21:31
Co pilot can help you code. Or you could literally have
00:21:35
something that is coding for a company.
00:21:37
And I think we see that framework come up again and
00:21:41
again, and when it feels like it'll be very disruptive when
00:21:44
you have agents like these things just doing it.
00:21:48
But I think that we don't do that yet.
00:21:50
I think the sort of optimistic take though on exactly the
00:21:53
argument we just made is that it's a little bit like the
00:21:55
industrial revolution for your brain, right.
00:21:58
It's, you know, it's the industrial revolution for your
00:21:59
brain. And that pretty much all the
00:22:01
inputs and outputs of the human mind are some form of text,
00:22:04
whether that's spoken or written or some form of imagery, right.
00:22:07
Whether it's something you see or it's something that you
00:22:10
create, whether it's, you know, a drawing or or in a piece of
00:22:12
imaging software, right. And if you think those,
00:22:15
basically all the inputs and outputs of the human brain can
00:22:18
be encapsulated in some form of text and images.
00:22:21
If you create technology that makes it really easy to create
00:22:25
high quality and also interpret high quality text and images.
00:22:29
Right. You've kind of like, you've got
00:22:32
like 80% of the job done of what the human brain can do.
00:22:35
Right. And to your point, there's this
00:22:36
distinction whether it's autonomous or it's helping you,
00:22:39
it's an assistant, whatever. But, you know, I, I think the
00:22:41
Industrial Revolution is interesting analogy because it
00:22:43
was like the first time. It was like you don't actually
00:22:45
have to sew this, like, shirt, right.
00:22:47
This, like, machine will sew it for you.
00:22:49
Right? Like, you can sit at this
00:22:51
machine and it'll be your assistant in sewing this shirt.
00:22:53
Right. And that's like a big deal,
00:22:55
right? Is the first time in human
00:22:56
history, like you don't have to actually sew the shirt, like,
00:22:58
without any help. Right.
00:23:00
And I think similarly like for all these different types of
00:23:03
work, whether it's homework or office jobs or legal work or you
00:23:06
know, mathematical analysis or writing or podcasting or
00:23:09
whatever, it's okay. Well, for the first time ever,
00:23:11
you don't have to do all the work right, whether you're
00:23:13
assisted or it just does it itself.
00:23:16
Like it's kind of a game changer because you have automation for
00:23:22
the human mind, for creativity in some fashion or another.
00:23:24
So that I think is like the really crazy optimistic take is
00:23:29
that we're at the beginning of the second Industrial Revolution
00:23:31
and it's no longer physical, but it's mental, right.
00:23:33
And I kind of believe that I would say I'm leaning that
00:23:36
direction based on where we are today.
00:23:38
Yeah, I continue to believe, but it's mostly from my experience
00:23:41
with ChatGPT that it's just insane, amazing.
00:23:46
I mean, I mean, I feel like it's great at like.
00:23:50
Writing a poem I I keep joking that people are going to write
00:23:52
all their vows with catching teeth, you know?
00:23:55
I I feel like these tasks were you know people are desperately
00:23:58
trying to get the same like style like groomsmen sort of
00:24:01
toast or whatever. It's it's great you know I feel
00:24:04
like it can be sort of creative it.
00:24:07
But yeah, I mean to me the the counterpoint is just it's just
00:24:10
so hard to know just like it was hard with self driving cars to
00:24:15
know when they would be complete and the completeness matters.
00:24:19
The extent to which completeness matters with a chat sort of
00:24:23
interface, because I think humans have been enticed, like
00:24:27
we were saying like decades ago with chat interfaces and we're
00:24:30
like, you're almost there. I've only it is like I have
00:24:33
stopped using Chat GP. Like do you guys in your daily
00:24:36
life use generative AI for anything?
00:24:41
I frequently use ChatGPT, but it's really I would say not for
00:24:45
productivity purposes. Maybe occasionally at.
00:24:48
Work just like. I mean, I'm not like talking to
00:24:52
characters, but I am using it to just brainstorm ideas.
00:24:57
Like I'll yesterday. I just.
00:24:59
I came up with this prompt that was like create a timeline of a
00:25:03
fictional historical world with the depth of.
00:25:07
Westeros or Middle Earth and but you know, and I essentially got,
00:25:12
you know, 10 years of history of a faith.
00:25:16
I thought I was super cool and I couldn't couldn't have done that
00:25:19
before. So I guess and then sometimes
00:25:22
I'll just ask it, you know, for ideas for new products or new
00:25:27
companies or it I just to see what kind of level of creativity
00:25:30
it is capable of. I think that's what's really
00:25:33
interesting to me. I think we all agree.
00:25:35
That there is creativity occurring that is creating
00:25:38
novel. Well, I guess we don't.
00:25:40
Not everyone agrees with this, right?
00:25:42
But that it's not solely capable of regurgitating information.
00:25:46
But from my perspective it seems very capable of creating new
00:25:51
original ideas when I play around with it.
00:25:55
And I believe there have been papers proving this that.
00:25:59
Well, Microsoft came out with one, right?
00:26:01
That said, there were like sparks of like.
00:26:04
General intelligence or. Whatever, I don't remember that
00:26:07
specifically, but I did see a paper that there's a common test
00:26:13
of creativity, right? Where you will essentially ask
00:26:16
people for. I guess one example they gave is
00:26:18
you ask cases around, what would you do?
00:26:21
What are name 100 use cases of this paper clip?
00:26:25
Or name 100 things you could do with this, a rubber band or
00:26:28
something, right? And then they essentially grade
00:26:30
the ideas. And I thought that was pretty
00:26:32
interesting test of creativity and and essentially ChatGPT is
00:26:36
performing, you know, better than most humans including most
00:26:39
like MBA students. So you know I think that there
00:26:43
are ways to like start to test this, but it's underrated the
00:26:46
level of creativity, not just that it's oh, it's cool that it
00:26:49
can create a poem, right. It can create more creative
00:26:52
poems than like poem poetry authors, right.
00:26:54
Like that kind of thing gets to. Be right, it's easy for humans,
00:26:58
he. We just sort of like sticking
00:27:00
our nose up at it. I sort of roll you know change
00:27:02
the goal posts basically. But then yeah, like you're
00:27:04
saying, people will give it these tests like what can an MBA
00:27:08
student do and what can ChatGPT do.
00:27:10
And like people are, you know are impressed I think blind with
00:27:14
the ChatGPT response just to I was yeah, Microsoft in May 20 of
00:27:21
this year said they saw. Sparks of general intelligence
00:27:26
basically in a research paper I think it almost any task right
00:27:30
now ChatGPT 4 is at the level of a pretty solid college student
00:27:35
like maybe in a minus college student in almost any field
00:27:38
which is sort of mind boggling right.
00:27:40
And like how many of us are like at the level of a minus college
00:27:44
student in in more than like maybe one or two things you know
00:27:47
and it's at the level of an A minus college student at
00:27:49
everything and it can serve millions of requests like at any
00:27:51
given time right. So it's like a scaled A minus
00:27:54
college student at basically everything and then particularly
00:27:57
these creative tasks as you're saying, I feel like the one
00:27:59
thing that holds it back is the need to be accurate, right?
00:28:03
And and often it invents facts or it's sort of like misaligns
00:28:08
realworld concepts in ways that aren't really realistic but in a
00:28:11
purely creative endeavor, like poetry or like creating artwork
00:28:14
or coming up with ideas for what to do with the paper clip,
00:28:16
right. Like when it sort of has doesn't
00:28:18
have to be anchored to any sort of like really hard facts.
00:28:21
It's unbelievable. I mean it's it's better than
00:28:23
almost anyone in the world. To what extent do you think this
00:28:26
is all exciting? Because we're like, oh, we're on
00:28:27
the cusp of general intelligence, right?
00:28:31
Like artificial General Intelligence, AGI, this idea
00:28:34
that you know it's. There are different tests for
00:28:38
him, but the idea that it's smarter than a human being to
00:28:41
bring to bring it back to the conference, I mean I would echo
00:28:44
what Ali Goatsy said on stage which is that I just think by so
00:28:47
many. Yeah, back in March at our last
00:28:49
conference and he'll be back for the next conference.
00:28:51
Stay tuned. Yeah, he he said basically,
00:28:53
look, I mean I think it is general intelligence already.
00:28:56
Like it already is, you know as good or better than 99.9% of
00:29:01
humans at almost any task you can throw at it, right?
00:29:02
I mean, how can you not say that's like general
00:29:05
intelligence, right? I just think that holding it to
00:29:08
the standard where it has to be 100% accurate about everything
00:29:10
or it has to be able to go do stuff on its own, which isn't
00:29:14
really that hard of a technical challenge.
00:29:16
I think that's just an it's sort of like goal posts moving
00:29:19
because I think people are like afraid of the idea that we have
00:29:21
created something that's like a smarter, smarter than a human
00:29:24
right. Like, clearly five years ago, if
00:29:27
you had told someone that we're going to have an AI chap out,
00:29:30
they can do everything that Chad JPT can do.
00:29:32
You know, pass the bar, pass AP exams, create beautiful artwork,
00:29:36
talk to you, you know, write poetry by the Dodgers.
00:29:39
But you'd be like so being. Blase Well, that sounds like
00:29:43
pretty freaking close to general intelligence to me.
00:29:45
What more do you want from this thing?
00:29:47
I just think that I don't. Know I wanted to have a through
00:29:50
line of reasoning where it seems to be a thinking being where I
00:29:54
can explain why it generated the answers it has, you know, Yeah,
00:29:58
I mean but can humans really explain why they generate
00:30:00
answers? In most cases?
00:30:01
I just think it's holding it to a really high standard that most
00:30:04
humans cannot meet. And so I would argue is it at
00:30:06
the level of a human in almost every area?
00:30:09
Absolutely. I mean, I would say it's in the
00:30:11
top 1% of humans in almost any area you throw at it.
00:30:14
Yeah, I guess I would agree with that mostly other than the main
00:30:19
area, I think it starts to fail or deteriorate is, is when you
00:30:25
kind of create too much. Memory or context right?
00:30:28
Like essentially humans have this amazing ability to recall
00:30:33
information throughout from their entire life right?
00:30:36
And to like sort of be able to maintain the context of a hour
00:30:43
long multi hour long movie or 20 page book right?
00:30:48
And sometimes it feels like ChatGPT it's stretching to that
00:30:53
with the amount of tokens, context, windows it can
00:30:56
understand but. It really does show
00:30:58
deterioration as you add more and more context, and then there
00:31:01
is a actual hard limit of context you can include in your
00:31:05
prompts. I hate the fact that it doesn't
00:31:09
remember when we've talked before, even the same thread,
00:31:12
and it just starts lying about what it said before.
00:31:14
There are parts of what human beings do that humans would
00:31:19
never do This sort of just totally make.
00:31:21
I mean, some would, but just like totally bullshitting, like
00:31:25
when it's like, why? Yeah.
00:31:27
Anyway, I wanted to get into more of the business question
00:31:30
from this sort of same framing with is it all Open AI chat CBT
00:31:35
On the one hand, you know, I I feel like we're seeing you know
00:31:39
people do these like Elo tests where they like compare
00:31:42
different foundation models and we do see like Llama, like
00:31:46
Facebook's model, open source model and other models like sort
00:31:50
of being competitive at times with chat CBT though 4.5.
00:31:55
ChatGPT 4.5 remains the gold standard.
00:31:59
I guess I'm curious, it's sort of like a 2/1.
00:32:01
Do you think other people will catch up and like how much do
00:32:05
you think there's sort of a Moat here?
00:32:07
Like how much do you think being ahead slightly or like having
00:32:11
been sort of the the one that consumers know about is is like
00:32:16
a Moat, like how, how, how defended do you think they are
00:32:19
with their position? That's a really hard question.
00:32:24
I mean, first of all, they're basically a subsidiary of
00:32:26
Microsoft, right. So you're asking for.
00:32:28
You're asking for. Yeah, Yeah.
00:32:29
You're asking, you know, are they going to be a huge, you
00:32:33
know, strategic value add to Microsoft going forward?
00:32:35
I think obviously, yes, right. I mean, are you asking is
00:32:38
ChatGPT always going to be the gold standard for text
00:32:42
generation models? Like, I don't know, it seems
00:32:44
like everyone's catching up. To your point, it also seems
00:32:47
like they have the best people and they're moving the fastest
00:32:49
on releasing new things. So they'll always stay, you know
00:32:52
6 to 12 months ahead. You know, does a Moat really
00:32:55
matter in this context again where like you're, you know,
00:32:58
you're you're the best at least and you're ahead of everyone
00:33:01
else. And again you're a subsidiary of
00:33:03
Microsoft, so there's no real business benefit to, you know,
00:33:05
winning anyway to you know it doesn't.
00:33:09
It's a sort of hard business question.
00:33:10
I think what's more interesting to ask is if all the knockoffs
00:33:13
or the competitors or the various image models or the
00:33:16
various, you know, versions of ChatGPT that are out there are
00:33:19
going to be successful. Because I think ChatGPT will
00:33:22
clearly be successful, I think in some contexts, I mean it's
00:33:24
going to be in Microsoft Word 10 years from now, right?
00:33:27
But what about anthropic? What about Llama?
00:33:29
What about Google Gemini or Lambda?
00:33:32
Or what about, you know, whatever Amazon's cooking up,
00:33:35
what about a million startups got funded in the last 10
00:33:37
seconds, right? I mean, I think those are like
00:33:39
more interesting questions because it's to your to the
00:33:43
earlier discussion it seems like a pretty commoditized concept
00:33:46
like chat with a large language model and unless someone can be
00:33:49
way better or have a very different business strategy than
00:33:52
ChatGPT, it's hard to see where they're gonna win, right.
00:33:56
And so people are trying different angles on this
00:33:58
concept, but the concept itself is not that different in
00:34:00
different companies. I do think this really gets to
00:34:04
the question of will there be one best AI essentially 1
00:34:09
ChatGPT like that we all use as our personal assistant, right?
00:34:13
And that is the most general, most high powerful model of.
00:34:17
Most highly intelligent model right that exists in the world
00:34:20
you know because it's so generalizable and then you know
00:34:24
I think there's a good. I think there's that's
00:34:26
plausible. I think that certainly you know
00:34:28
we use. 11 you know browser and one e-mail client, I mean I
00:34:33
don't know like we don't, we're not switching between them a
00:34:35
lot, right. But I think it starts to ask you
00:34:38
know, will that model also be better at all other you know
00:34:42
vertical tasks as well? I don't that seems harder to
00:34:44
believe right? Like that.
00:34:45
It will also be the best model at reading legal documents and.
00:34:50
Being I was just gonna bring up case tax, sold the Thompson
00:34:53
Reuters except for hundreds of millions of dollars and.
00:34:57
They were like, I mean much more sophisticated than ChatGPT
00:35:00
wrapper, but they were using ChatGPT largely.
00:35:04
Sure, regardless of what their specific model was doing, it
00:35:09
seems possible you could put a ton of compute and specifically
00:35:14
trained legal data into a model that would outperform ChatGPT at
00:35:19
that use case of the law, right? Or same for medicine or
00:35:22
something, but. I don't know.
00:35:24
I'm not 100% confident in that. I think that it's very possible,
00:35:27
just with the you know that chat TBT as a foundational general
00:35:32
model will just always be better at all of those key use cases.
00:35:36
I mean Jason Warner, you know who's speaking in the
00:35:38
conference, who's now the CEO of Poolside is betting that he can
00:35:42
build a foundation model for code.
00:35:44
Obviously, you know, yeah, you can see a bazillion companies
00:35:47
that are sort of trying to be like I will be a foundation
00:35:50
model. Yeah, for, for XI mean I think
00:35:54
it's, I think it's sort of instructive to think back to
00:35:55
like social networks, right. And for me, I always find it
00:35:58
interesting analogy of, OK, you know, 1520 years ago would you
00:36:04
predicted there would be a social network for work and
00:36:08
there would be a social network for gaming and there would be a
00:36:11
social network for people under 25.
00:36:13
That's Snapchat. And there'd be a social network
00:36:14
for people between 25 and 45 and that's Instagram and there'd be
00:36:17
a social network for people who are over 45.
00:36:19
And that's Facebook blue, right. I mean like and and and it's a,
00:36:24
it's sort of interesting to think about what the cleavages
00:36:27
are in the the user need. Right.
00:36:29
And in social networking, it's often just age cohorts sort of
00:36:33
build these network effects with each other.
00:36:34
But then there's also this LinkedIn, which is, oh, actually
00:36:37
work is a completely different social concept in your life that
00:36:40
you need to keep separate from everything else.
00:36:42
And then you have Discord, which is like actually gaming is like
00:36:45
a completely different concept that you need to keep distinct
00:36:47
from everything else. Right.
00:36:48
And I don't think it would have been that easy to predict those
00:36:50
things. I mean maybe like Reed Hoffman
00:36:52
will say, it was super easy to predict, right?
00:36:53
But but I think with these foundational models, it's
00:36:56
similar, right? It's, you know, there probably
00:36:57
aren't really network effects other than just like who can eat
00:37:00
the most data the fastest, right, Which seems like it's
00:37:02
going to be GPT. And so the data, quote UN quote,
00:37:05
network effects, are the data scale effects are probably
00:37:08
always going to be 1 by 1 by GPT, right.
00:37:11
Then the question is, are there other use cases where there's
00:37:14
some kind of network effect or there's some sort of different
00:37:17
business concepts? You know, Facebook seems to be
00:37:19
going the angle of we're going to open source this all.
00:37:21
So people will just build it into all these things like it's
00:37:23
Linux back in the day and that'll be the way we win is
00:37:26
that it'll be the free open source version and you can just
00:37:29
throw it into everything if you want to, which I think it's a
00:37:31
pretty interesting like business concept, right.
00:37:33
And then I don't know that much about the, you know, enthropics
00:37:35
or the pool sides or whatever. You know, how are they going to
00:37:37
win? What is the cleavage in the the
00:37:40
use case or the way these models are built that is going to allow
00:37:43
someone other than ChatGPT to win?
00:37:45
Because it seems like they're gonna win on data and they're
00:37:47
gonna win on hardware. So like, where are you gonna win
00:37:50
if you're not them? I guess, and I'm sure everyone
00:37:52
has an answer to this, but that's me.
00:37:53
It's the hard question. We've spent a lot of time
00:37:55
talking about the foundation models.
00:37:57
I mean, Chad, I mean open a Eye is sort of a combination, right,
00:38:00
where it's like they have the foundation model, they apply it
00:38:03
to these use cases. I mean people talk a lot about
00:38:06
you know, applications and like infrastructure, right.
00:38:09
I mean, there are companies, they're all these sort of wonky
00:38:12
companies like. The vector databases have been
00:38:15
super hot of the last couple months, right?
00:38:17
People talk about companies like Pine Cone, We V8 there.
00:38:21
There's like a whole list of them which are just like trying
00:38:23
to organize your data in a better way to get it into
00:38:27
foundation models. I wanted to get to the actual
00:38:30
like applications right and you sort of Max earlier referenced
00:38:34
the sort of chat bot application which also goes back to you
00:38:39
know, your both of your early days at volley trying to build.
00:38:44
Chat bots. I'm curious getting away from
00:38:47
who has the technical expertise, like what you think is
00:38:51
interesting in the sort of what its character.
00:38:54
Is it replica? You guys know this world much
00:38:57
better than I do. So you're sort of getting at do
00:39:00
we feel these are successful use cases or?
00:39:03
Right. Yeah.
00:39:04
Yeah. Do you think there's sustained
00:39:06
promise there or what do you see in terms of people actually
00:39:10
using AI and applications that? That excites you.
00:39:15
There's a challenge with these models being accurate, obviously
00:39:20
100% of the time, and you can debate whether that's necessary
00:39:24
or not, right? To consider it True general
00:39:27
intelligence. But in the entertainment space,
00:39:29
like, it's just less of a problem, like conversing with a
00:39:33
fictional character or historical character, right?
00:39:36
Like these things don't need to be extremely accurate because
00:39:38
they're essentially entertainment anyway, I think.
00:39:42
I guess I always come back to some lessons I've learned in the
00:39:47
gaming world, where there's a clear difference between an
00:39:52
entertaining demo that is fun to do and fun to play with.
00:39:55
And I think we've seen a lot of those that are actually really
00:39:58
amazing and impressive demos, but they don't have long staying
00:40:02
power or retention, right? So thinking of all of the apps
00:40:06
that create fun. Photos that put you into their
00:40:10
photos or. Yeah, what's can of soup.
00:40:13
Can of soup is the. Latest I keep meaning to write
00:40:15
about them. They're like super buzzy, right?
00:40:17
Yeah. And you can put yourself in AI
00:40:19
generated photos with your friends.
00:40:21
I think it's amazing. It's really cool.
00:40:23
But the question is, does that really have staying power?
00:40:26
Do you build a social network around it in order to make it
00:40:28
have staying power and network effects?
00:40:30
These are real challenges to create like a sustainable
00:40:33
business and startup that achieves a great outcome.
00:40:37
So and similarly with talking to characters, character AI I think
00:40:40
clearly has some product market fit there with people wanting to
00:40:44
come back and talk to those characters.
00:40:46
But I think what Max and I like to think about is how do you
00:40:48
even build more retention around that?
00:40:50
How do you build like game mechanics or features in the
00:40:53
concept of a virtual pet that you might, we know from looking
00:40:57
at mobile gaming and previous eras of gaming that you can
00:41:01
build long running retention into that if you add game
00:41:05
mechanics to the experience. So yeah, I'm less bullish on
00:41:09
just like general like character conversations that don't have
00:41:14
any are sort of aimless or don't have an end point or a purpose
00:41:17
and more bullish on, you know, how do we bring those characters
00:41:21
or Npc's into a gaming context with like normal gaming
00:41:25
objectives. I do think in interacting with
00:41:28
these underlying characters, I think to your point, I think a
00:41:30
lot of the real value is when you start losing sight of it
00:41:33
being an AI, whether it's like. Virtual boyfriend, girlfriend
00:41:36
type thing, companionship, you know, sex chat.
00:41:39
As we said, being a big opportunity there.
00:41:41
Are those companies allowing it or they cracking down on it?
00:41:44
They claim they're cracking down, but then if you go on
00:41:46
Reddit and you look at all the screenshots from the last week
00:41:49
of what people have been talking to these, I mean, go on, go on
00:41:52
the Reddit of any large language, character driven
00:41:55
experience, you know, you'll see what people are really
00:41:58
passionate about using it for, you know, replica or character
00:42:01
or whatever, right? So I think that, yeah, losing
00:42:05
sight of whether or not it's an A I is, is important in some
00:42:07
contexts. But in the end, I mean, in any
00:42:10
story experience, from reading a book to playing a video game to
00:42:16
chatting with an A I like the goal or watching a movie is to
00:42:20
become immersed in a world that doesn't actually exist, right?
00:42:22
And you know, when you're watching a movie, you're
00:42:24
watching Lord of the Rings, you know it doesn't actually exist.
00:42:26
Those characters aren't real, but you sort of become swept up
00:42:29
in the narrative in the world, right?
00:42:30
And. I think similarly, I think
00:42:32
that's where the opportunity is with a lot of these character
00:42:35
driven AI entertainment experiences.
00:42:37
Do you guys want to try this other one that I have?
00:42:40
Oh yeah, you have a game? Yeah.
00:42:42
Sure. So I thought it'd be fun to see
00:42:47
what GPT for ChatGPT for thinks your lives will be like in the
00:42:51
year 2028 and see if you either of you agree with that.
00:42:56
The prediction there so. I say it again.
00:43:00
I I asked Chachi BT to predict a day in the life of Eric
00:43:06
Newcomer, specifically a American journalist reporting on
00:43:11
the Silicon Valley and startup industry, which I borrowed from
00:43:13
your Wikipedia and, you know, gave it a little bit more
00:43:17
context of sure about your sub stack.
00:43:19
And yeah, it came up with A day in the life in the year 2028.
00:43:22
So I'm going to read it to you and you guys can assess whether
00:43:26
you think this is accurate. Day in the life of Eric
00:43:30
Newcomer, 2028. The first hints of Don kissed
00:43:34
the San Francisco skyline. As Eric newcomers, smart blinds
00:43:38
slowly begin to rise. The ambient sounds of birds
00:43:41
chirping echo softly through the smart speakers in his apartment.
00:43:45
The AI? The AI.
00:43:47
Everything's just smart. The AI driven home system has
00:43:50
analyzed his REM sleep and calculated the optimal wake up
00:43:55
time to ensure he starts the day with peak cognitive efficiency.
00:44:00
Lifting himself out of bed, Eric's AR glasses are laying on
00:44:04
the night stand feeding him the day's analytics.
00:44:07
Newcomers Media Outlet has grown tremendously.
00:44:10
AI. But we also have to be bullished
00:44:11
by AR augmentedly out here and you're seeing okay.
00:44:14
Well, you can. You can assess whether you think
00:44:16
that's accurate, but you should. You should at least appreciate
00:44:19
this newcomers Media outlet has grown tremendously and AI Bot
00:44:23
the. New like, what?
00:44:25
Is it like a horoscope? You're reading my Yeah, I'm
00:44:29
predicting your feature and AI bot had curated and condensed
00:44:33
the most relevant news from the vast ocean of information, all
00:44:37
of this streamlined and visualized for maximum
00:44:39
absorption. Eric's morning routine was
00:44:42
synced seamlessly with his devices as he finishes his quick
00:44:46
morning workout, suggested and supervised by his virtual
00:44:49
personal trainer. His smart kitchen is already
00:44:52
brewing, his favorite blend of coffee tailored to his genetic
00:44:55
makeup and current health metrics.
00:44:58
Settling into his study, the day's agenda was projected
00:45:01
holographically. The AI had lined up interviews,
00:45:04
articles to review, and podcast episodes to record.
00:45:08
Eric's writing has also evolved. With the help of AI enhanced
00:45:11
tools, he could construct and edit stories with increased
00:45:14
speed and position. His A I assistant named Nora was
00:45:17
able to give real time feedback on the readability, engagement,
00:45:21
and impact of his writing. Nora also had a mode where she
00:45:24
could mimic Eric's style, allowing her to draft articles
00:45:27
for him. It was a collaboration that Eric
00:45:30
had grown to appreciate midday. As Eric prepares for his podcast
00:45:34
recording, he remembers the days when he used to worry about
00:45:37
Internet connections and sound quality.
00:45:39
Now with quantum driven communication technology, my
00:45:42
God, not only was the. Connections, but whole
00:45:45
technology here. But he could also record in
00:45:47
virtual environments, giving his audience an immersive
00:45:50
experience. Today's episode was recorded in
00:45:53
a simulation of a venture capital firm in the heart of
00:45:57
Silicon Valley. Yeah, this both he and his.
00:46:00
Sandhill Rd. The most beautiful both he and
00:46:03
his. Guests wants to be there.
00:46:05
Both he and his guests appear as lifelike holograms afterwards.
00:46:09
Eric took a moment to reflect. He looked out over the city,
00:46:12
remembering the early days of Newcomer Now.
00:46:15
He wasn't just delivering news. He was shaping the future
00:46:18
narrative of startups and venture capital.
00:46:21
The weight of that responsibility was not lost on
00:46:23
him. Yeah, that's about it.
00:46:27
Certainly. I mean, it's funny, I find a lot
00:46:30
of the predictions about non AI stuff to be the most annoying,
00:46:33
that it's so certain hardware is hard, artificial augmented
00:46:37
reality. And I mean, there was some
00:46:38
other. I'm pretty short quantum, right?
00:46:41
Quantum Internet? Exactly.
00:46:42
Yeah, can. I take the under on quantum
00:46:44
powered Internet or whatever that was.
00:46:46
Well, it's an interesting thing because we were talking about,
00:46:48
you know, believing in AGI, right?
00:46:51
And can you, you know, if you believe in AGI or general?
00:46:54
Well, then it would be about much more, you know?
00:46:57
Yeah, I get the point. We could have.
00:46:58
We could wish everything. Yeah, exactly.
00:47:01
Or. That's that's that sounds like a
00:47:03
story an AI would tell me to convince.
00:47:05
Me. That we wouldn't all be dead in
00:47:06
five years, thanks to AI. And it was like, the future's
00:47:10
gonna be right. Yeah, that would not be AI
00:47:12
having killed us. Too optimistic.
00:47:14
You'll have smart blinds. I know they.
00:47:15
Try to like I mean I do think a problem with some of this stuff
00:47:19
is like it's programmed to be like too benali optimistic like
00:47:23
I find it like I like beg chat you video be more like George
00:47:27
Carla. I don't know, just behave with
00:47:29
some free thought. And it's so why isn't it like
00:47:32
Eric, I don't know, at that age he's probably has some cardiac,
00:47:35
you know, whatever. Where's the like medical thing
00:47:39
where it's five years from now, you know, anyway is easy to
00:47:42
start to hurt when he goes on runs or whatever.
00:47:44
And oh, and I would be interested in sort of the AI
00:47:46
piece, obviously of the medicine.
00:47:48
I mean the idea that I would have a writing assistant.
00:47:52
Seems like basically plausible. Today I go into chat literally.
00:47:57
When I did, I wrote my own vows, but they proofread them and
00:48:00
Chatchi video tweaked, told me to move one thing to active
00:48:04
voice from passive voice. You know, it's like, I feel
00:48:06
like, I feel like this is assuming that there's gonna be
00:48:09
this like perfect equilibrium of you reporting the news and doing
00:48:13
interviews and using your AI assisted.
00:48:15
I just kind of find that to be, like, not that plausible.
00:48:18
Like it's either gonna be one or the other, Like you're gonna
00:48:21
still be doing most of the work, or you're gonna be doing almost
00:48:24
no work and won't have a job, or you'll essentially have evolved
00:48:29
into a brand instead of a writer, right?
00:48:31
I mean, it seems hard to believe that we're gonna thread the
00:48:34
needle here, that you will still be doing all this.
00:48:37
Intellectual and the temptation, I mean.
00:48:39
I do think one of the real fears I have about AI that will it's
00:48:44
just like. The temptation not to think if
00:48:47
it can do your task for you, right?
00:48:50
I mean that's what we're sort of seeing with that's what I'm.
00:48:52
There are a lot of like some types of cheating where it's OK
00:48:55
you can bring in the formulas, we still have to use them or
00:48:58
whatever. At least you still have to
00:48:59
think. Whereas like when chat GBD can
00:49:02
produce the final text for you. Man, that is like a path to.
00:49:07
Not progressing as a writer anymore, right?
00:49:09
Like as soon as you're just like you're not human out of the
00:49:12
loop. You know what?
00:49:13
Yeah, exactly. What am I?
00:49:14
Besides, what are you providing to this business?
00:49:17
Right. Yeah.
00:49:17
I agree though, it's just so hard to get anything edgy or
00:49:20
funny out of it. It's like kind of an, I don't
00:49:22
know, it's tough to it's. Very driven.
00:49:27
By scifi, I feel like one of the child, I think.
00:49:31
I think this is the prompt, right?
00:49:32
It's hard to get it to think like, I know you could say the
00:49:35
same thing about humans. There is no independent thought.
00:49:37
You're parodying a bunch of stuff you've heard, but it does
00:49:39
feel like have it generate a really new, an actually new
00:49:45
idea. You know, I think it's, I think
00:49:46
it's very capable of that, actually.
00:49:48
I I found that if you tell it to be extremely original, you know,
00:49:53
ignore. I feel like it just gets like
00:49:56
rhymy. I don't know.
00:49:57
I feel like it. I do think there's like a
00:49:59
prompting element to this feeling that it's not able to be
00:50:03
as creative or original as you might think.
00:50:05
In the same exercise that ChatGPT is there.
00:50:09
We can compare and contrast in five years.
00:50:11
What are specific predictions, less colorfully said, that you
00:50:16
would make in five years? I mean, I definitely think to
00:50:23
come back to the self driving thing, it seems like we're
00:50:25
actually going to have fully self driving cars in five years.
00:50:27
I mean, I know you could say. We are in cities.
00:50:29
I think the big question there is just how much you know Waymo
00:50:33
and. People think.
00:50:34
Take what they've. Learned I think if today they
00:50:37
can drive around San Francisco I think they'll be able to drive
00:50:39
between cities and and in in the vast majority of cities or
00:50:43
whatever at that point at least. I don't know if the United
00:50:45
States it has to be trained on different data than other
00:50:47
countries or whatever or something, but I think that.
00:50:50
I mean the fact that we have live way Mo's like dropping
00:50:53
people off on my street every day just makes you think you're
00:50:55
going to be able to go anywhere in a self driving car in five
00:50:58
years. Which again, I lost the bet the
00:50:59
other way on that last time, so I'll probably lose the bet
00:51:02
somehow this time. It's a problem.
00:51:04
And also, just like car turnover is such a long life cycle.
00:51:08
Depends how many you're what. Percent, It's not that I'm
00:51:11
arguing for 100% penetration of those, It's more that I will.
00:51:14
I will argue that you know you will be able to take one you
00:51:17
know, as an Uber or whatever in any major US city.
00:51:20
And you know, I don't know whether or not you'll be able to
00:51:23
buy one. That's sort of an interesting
00:51:24
question as to whether or not it'll be an ownership model
00:51:26
versus an Uber type model or a lease model or whatever you want
00:51:29
to call it. But I think they'll be like you
00:51:31
know, available for use. So you know at large scale
00:51:35
across the major U.S. cities and also between cities, right.
00:51:39
It just seems like we're clearly like pretty much they're
00:51:43
assuming way mo and crews aren't, like lying about the
00:51:45
capabilities of their their vehicles.
00:51:47
Interesting that your main predictions around self driving.
00:51:50
I was thinking about something. I feel pretty good about and I
00:51:52
agree with that. I do think this view of you will
00:51:56
wake up and there will be essentially your personalized
00:51:59
assistant. Maybe it'll be in the smart
00:52:01
speaker or in your wall or your mirror or something, I guess.
00:52:05
I just, I don't know if it'll be Alexa or ChatGPT or something
00:52:10
brand new, but I do think, you know it's going to be a voice
00:52:15
probably driven experience. And bias.
00:52:18
Bias. Well, at least in your home and
00:52:20
probably bias. You're going to be typing to it.
00:52:22
I didn't love it. It doesn't make.
00:52:24
Sense. Very efficient.
00:52:25
People love it. It's very so people do not love
00:52:27
typing, no. Typing is way less efficient
00:52:30
than dogging. No you're I mean it'll you'll be
00:52:33
able to type to it. But I just think like most of us
00:52:37
will be like talking to this AI assistant and it'll be I
00:52:41
generally think there will just be one that I use every day
00:52:43
maybe maybe there will be more than one in the market like that
00:52:46
people use. But I will just have one that
00:52:49
learns my preferences and becomes personalized to me and
00:52:53
creates kind of a history with me.
00:52:56
And it might even actually recommend other assistance for
00:53:00
certain use cases, right? If I have to go prepare for
00:53:03
Max's congressional testimony, I will maybe use a separate bot
00:53:07
for that or something. But yeah, specialize in.
00:53:11
Testimony bot. But I do think, yeah, I feel
00:53:14
pretty strongly we're going to have conversations, voice
00:53:18
conversations with our personalized assistants every
00:53:20
day. The prediction I'll make this
00:53:23
sort of different than what's been said is underlining that I
00:53:27
think the average person wants consumption more than creation.
00:53:32
And that TikTok is in some ways the actual most used thing in in
00:53:37
the AI world and and this sort of yeah, chat GBD got at this a
00:53:41
little and its prediction about me where it's very good at
00:53:44
sorting things that I want. I I think in five years we'll
00:53:49
see at least the beginnings of like pure AI generated social
00:53:54
account. I mean, you're already seeing
00:53:56
sort of like these, like, women and cartoons that are like.
00:53:59
This seems like they're gaming like the Instagram algorithm
00:54:02
with like totally sort of machine made.
00:54:05
But like the I think even like video within five years I think
00:54:09
we'll have, yeah, you're on TikTok and it's just here is a
00:54:12
generated video and it's warring with actual creativism.
00:54:16
I think that would be like potentially A generationally
00:54:19
culturally interesting period where like you have like young
00:54:23
people coming up where they're just like.
00:54:25
Being fed sort of what they want outside of it could create like
00:54:29
a really weird type of humor where like they have they're.
00:54:32
Used to? Yeah.
00:54:33
Good. There's like an AI.
00:54:34
There's like a bunch of AI sort of celebrities and accounts and
00:54:38
they create content and they interact and they host, you
00:54:41
know, podcasts and. It's all like effectively a be
00:54:44
tested. It's like they put out like 100
00:54:46
version build, you know, a ton of versions of the video and
00:54:49
they see which one is getting engagement and then they slowly
00:54:52
funnel into those just TikTok chooses which videos get
00:54:54
surfaced. Yeah, that approach is in the
00:54:57
creation. It makes most sense for short
00:55:00
term, short form and but you know the day you can make like a
00:55:03
movie about it, like then we're like killing American industry
00:55:06
but. Yeah.
00:55:07
Well, the question is, do you think we're going to get
00:55:09
personalized? I mean, you mentioned they're
00:55:10
going to a B test everything or whatever.
00:55:12
But is it that there are literally be 8 billion different
00:55:16
versions of each piece of content for each person?
00:55:19
If there is no such thing as a piece of content anymore, like,
00:55:23
it's just you get a version of some concepts like that is
00:55:27
perfect for you, right? I mean or do we still have some
00:55:30
sort of like value in being like, oh, did you see that video
00:55:33
the other night about that thing and you can talk about it and
00:55:35
you can do, I think? Relationships, I mean, people
00:55:37
are going to try both. I mean, I think one, there's
00:55:40
like a limited, there's like A at some point there aren't
00:55:44
enough humans, right? Or there's like there there is
00:55:48
like a data shortage, right? It's hard on some level.
00:55:50
If you do something population wide you can really test and see
00:55:54
what works like broadly. Whereas like running experiments
00:55:57
on me, you just don't not get enough shots on goal necessarily
00:56:02
to be great at like the TikTok style.
00:56:04
So to me that leans a little bit more less this sort of obsessive
00:56:08
personalization and more. I mean, obviously they're
00:56:11
subgroups, but not like to a person, more like subgroups.
00:56:14
Yeah, I think subgroups, like, I think it'll not just be like
00:56:18
training on your information, but other people who have
00:56:20
similar experience, similar behaviors that you do on the
00:56:24
app, right? It's able to use that's that
00:56:26
data for training too. My last AI a GS in prediction
00:56:29
is. I mean this was in some of these
00:56:30
forward-looking GBT scenarios is?
00:56:33
I mean, I do think you take five more years of development on AR
00:56:36
glasses or whatever, You know, I don't think they'll be as
00:56:38
lightweight as the glasses we wear every day.
00:56:41
It'll probably still kind of look like a VR headset in many
00:56:44
ways. But I do think people are going
00:56:46
to spend like hours per day inside a high quality Apple
00:56:49
Vision pro type experience. Because I think again, I mean, I
00:56:53
mean if the average Americans watching five or six hours of TV
00:56:56
a day, which they are like. Why wouldn't you watch two of
00:56:59
those hours on 100 foot screen in front of Mount Hood,
00:57:01
Washington or whatever, right. Which is basically the pitch
00:57:03
from Apple, you know or why wouldn't you watch personalized
00:57:07
AI generated TikTok content like from Eric's, you know, feed or
00:57:09
whatever, right. I just think that, you know,
00:57:12
screens getting better has been one of the true constants of our
00:57:15
lifetime, right? And the Vision Pro which is sort
00:57:19
of the infinite screen, the infinite canvas for visual
00:57:22
content is, is kind of the ultimate expression of that,
00:57:25
right. So.
00:57:26
I think. That I'm more on the tenure.
00:57:27
There will be quite a lot of time.
00:57:29
And I feel like that space has been dogged by limits of optics
00:57:34
and just sort of hard constraints.
00:57:36
And so a lot of our intuitions about software level
00:57:39
improvements don't translate and I would say I would take the
00:57:43
longer time horizon. On that maybe, Yeah, I don't
00:57:47
know. Apple seems to think they got
00:57:48
it, so I guess we'll all see. Yeah.
00:57:49
This is a smart Max take. When's Apple been wrong before?
00:57:52
I don't know. Trust.
00:57:53
I'm just saying that when Apple ships something like.
00:57:56
I mean, again with the iPhone, like, was there a touchscreen
00:57:58
that had ever worked before? No, there never had been, right?
00:58:01
I mean, I just think that are you gonna get?
00:58:02
That wrong? Is that out yet?
00:58:04
No. Next year.
00:58:05
Oh, I. No, you can't get it until next
00:58:06
year, but I'll definitely get one for sure.
00:58:09
Yeah, you're like, I can expense.
00:58:10
I mean, for God's sake, right? Like you're like, we're a game
00:58:13
company. I would buy it out of my hard
00:58:15
earns personal money. OK, I would.
00:58:18
I I don't know. I I guess, yeah.
00:58:20
I I trust Apple when they make a.
00:58:23
I mean, this is a decade long hardware bet and they waited a
00:58:26
decade instead of shipping five or seven years ago because they
00:58:30
didn't feel like it was good. Enough is there, besides
00:58:32
obviously that Chad GBT was getting us thinking about this
00:58:36
technology, Do you see any sort of AI connection or AI utility
00:58:41
or how do you see? I just again, think with these
00:58:44
characters that we're talking about, whether it's an assistant
00:58:47
or whether it's an in game character or whether it's a
00:58:49
virtual companion or whatever, a pet or a girlfriend or
00:58:52
boyfriend. Like if they are the size of a
00:58:56
real object that you know, if they're the size of an actual
00:58:59
virtual boyfriend, girlfriend, or a size of a, you know, a
00:59:02
Pokémon, whatever that actual size is and you can talk to
00:59:06
them, that's going to be a better experience than.
00:59:08
Looking at a tiny little screen in your hand, right?
00:59:11
And. I wanna.
00:59:12
Yeah, I wanna. Write I mean the.
00:59:14
Companion I've got. Yeah, yeah.
00:59:16
I mean, have you read the Golden Confidence?
00:59:17
Right? They all have.
00:59:18
Like, yeah, yeah. Yeah, how do I get that?
00:59:21
But you want that's part of self projection though also.
00:59:25
I mean, you want everybody to see that kind of thing, not
00:59:27
just. So then, yeah, sure, if you're
00:59:30
both in Apple vision class headsets.
00:59:33
Right. Amazing.
00:59:35
All right, this is what this is basically our first episode.
00:59:39
We're gonna come out with probably five more.
00:59:43
I think next episode will be, what do we think of sort of the
00:59:47
apocalyptic vision of AI? Yeah.
00:59:50
Anything you guys would add on what to look forward to you in
00:59:53
the next couple episodes? I'm excited to run this
00:59:59
conference. I think it was amazing the first
01:00:01
time. I think it's gonna be even
01:00:03
bigger and. Better this time.
01:00:05
Yeah, I genuinely believe this is like the biggest thing since
01:00:08
the Internet or the iPhone. So I think that we're all pretty
01:00:12
authentically pumped about what's happening in the space.
01:00:14
And I think the amount of stuff that's happened in the last six
01:00:16
or seven months has been mind boggling.
01:00:18
And it just feels like things are moving faster than at any
01:00:21
time in my entire life in any technology space.
01:00:23
So it's just super exciting to. Even be like remotely adjacent
01:00:26
to any of this, so. I'm at the heart of it at
01:00:28
Cerebral. At the heart of it, yeah, at
01:00:30
least physically. At the heart of it, working on
01:00:35
working on the other elements and being at the heart of it.
01:00:37
But yeah, physically for sure. Great.
01:00:39
Well, that's our episode. I'm Eric.
01:00:41
Newcomer, Max Child, James Wilsterman are my Co hosts of
01:00:46
cofounders of Volley. Thanks so much to Scott Brody,
01:00:50
who's been producing the episodes.
01:00:52
Shout out to Riley Kinsella, my Chief of staff who's super
01:00:55
involved, Gabby Caliendo who works at Volley, whose quarter
01:01:00
to the conference and making everything happen.
01:01:02
I think she made sure you guys had microphones and lights so
01:01:06
you could actually see and hear you.
01:01:08
Thank you to young Chomsky as always for the theme music
01:01:13
Please got to build the feeds? Like, Comment, Subscribe on
01:01:17
YouTube, give us a nice review on Apple Podcasts and.
01:01:21
Go play Yes Sire or song quiz. And for me, subscribe to the sub
01:01:28
stack newcomer.co That's the most important thing.
01:01:32
Thank you. All right.
01:01:33
We'll see you next week. Goodbye.
01:01:36
Goodbye. Goodbye.
01:01:37
Goodbye. Goodbye.
01:01:38
Goodbye.