Robots, Rights, and Cloud AI Deals
Key Points
- The show kicks off with a discussion on the ultra‑early market for 1x Neo, a new $500‑per‑month (or $20 k one‑time) humanoid robot, highlighting how pricing is essentially a test of market appetite.
- Panelists examine the legal pushback from Japanese copyright holders against OpenAI’s Sora 2, underscoring growing tensions between generative AI tools and existing IP law.
- A major partnership between AWS and OpenAI is announced, signaling deeper cloud‑infrastructure support for OpenAI’s models and services.
- The news roundup covers Perplexity’s $400 M deal to embed AI search in Snapchat, Coinbase’s AI agents with crypto wallets for autonomous purchases, Instacart’s AI suite for real‑time grocery inventory and meal planning, and Google’s launch of solar‑powered AI chips aboard satellites (project “Suncatcher”).
- Throughout, the hosts stress that many of these developments are still experimental, with pricing, regulation, and deployment models evolving rapidly.
Sections
- Early AI Wave: Robots, Laws, Deals - The podcast episode examines the pricing uncertainty of fledgling AI tech while covering the 1x Neo humanoid robot, a Japanese copyright challenge to OpenAI’s Sora 2, the new AWS‑OpenAI partnership, and Perplexity’s $400 million integration with Snapchat.
- Humanoid Home Robots: Hype vs Reality - The speakers contend that, despite impressive marketing videos, current AI and robotics lack the reliability needed for autonomous household humanoids, meaning years of further development and teleoperation are still required.
- From Autonomous Cars to Home Robots - The speaker reflects on how self‑driving vehicles quickly become mundane, draws parallels to the slower timeline for fully automated household robots, and cites Asimov’s Three Laws as a guiding safety framework.
- Pricing Strategy for Early‑Stage Home Robotics - The speaker analyzes why a nascent home‑robot startup charges $20,000 for early access, linking the high price to the value of data, market testing against housekeeper costs, and current teleoperation staffing constraints.
- Risky AI Hardware Costs & Piracy Concerns - The speakers discuss the uncertain pricing and complex physical infrastructure of advanced AI systems, their strategic use for mindshare, and note a recent anti‑piracy complaint from Japan’s CODA to OpenAI.
- Synthetic Data, IP, and Revenue Sharing - The speaker examines how intellectual‑property rights, royalty structures, and data‑permission marketplaces could operate for AI‑generated content, questioning the practicality of revenue‑sharing mechanisms and ownership when synthetic data is used.
- Navigating Style Boundaries and Revenue Shares - The speakers debate the ambiguous definition of artistic styles in synthetic data, model providers’ cautious use of copyrighted material, and the potential for third‑party platforms to broker revenue‑share agreements between artists and AI model owners.
- Balancing Prompt Censorship and Model Interpretability - The speakers discuss the tension between restricting user prompts to prevent harmful content, the lack of clear governance in the AI race, and the need for societal frameworks and interpretability research to guide what is permissible.
- AI Firms Forge Multi‑Cloud Alliances - The speakers discuss how companies like Anthropic and OpenAI are forming overlapping partnerships with AWS, GCP, and Azure, creating a complex multi‑cloud ecosystem that adds technical complexity but offers diversification and strategic advantages.
- OpenAI's Commitment to Nvidia GPUs - The speaker argues that despite diversifying cloud providers, OpenAI remains tied to Nvidia hardware because its models are heavily optimized for those chips, making any switch costly and technically challenging.
- OpenAI Inference Focus for Enterprise - The speakers argue that OpenAI’s new GPU resources will primarily serve inference—enabling agentic workloads for business customers—and debate whether this capability will be offered directly on cloud platforms like AWS or via a marketplace, noting the lack of clarity around proprietary model availability.
Full Transcript
# Robots, Rights, and Cloud AI Deals **Source:** [https://www.youtube.com/watch?v=9GhT1mp5Edk](https://www.youtube.com/watch?v=9GhT1mp5Edk) **Duration:** 00:36:17 ## Summary - The show kicks off with a discussion on the ultra‑early market for 1x Neo, a new $500‑per‑month (or $20 k one‑time) humanoid robot, highlighting how pricing is essentially a test of market appetite. - Panelists examine the legal pushback from Japanese copyright holders against OpenAI’s Sora 2, underscoring growing tensions between generative AI tools and existing IP law. - A major partnership between AWS and OpenAI is announced, signaling deeper cloud‑infrastructure support for OpenAI’s models and services. - The news roundup covers Perplexity’s $400 M deal to embed AI search in Snapchat, Coinbase’s AI agents with crypto wallets for autonomous purchases, Instacart’s AI suite for real‑time grocery inventory and meal planning, and Google’s launch of solar‑powered AI chips aboard satellites (project “Suncatcher”). - Throughout, the hosts stress that many of these developments are still experimental, with pricing, regulation, and deployment models evolving rapidly. ## Sections - [00:00:00](https://www.youtube.com/watch?v=9GhT1mp5Edk&t=0s) **Early AI Wave: Robots, Laws, Deals** - The podcast episode examines the pricing uncertainty of fledgling AI tech while covering the 1x Neo humanoid robot, a Japanese copyright challenge to OpenAI’s Sora 2, the new AWS‑OpenAI partnership, and Perplexity’s $400 million integration with Snapchat. - [00:03:03](https://www.youtube.com/watch?v=9GhT1mp5Edk&t=183s) **Humanoid Home Robots: Hype vs Reality** - The speakers contend that, despite impressive marketing videos, current AI and robotics lack the reliability needed for autonomous household humanoids, meaning years of further development and teleoperation are still required. - [00:06:33](https://www.youtube.com/watch?v=9GhT1mp5Edk&t=393s) **From Autonomous Cars to Home Robots** - The speaker reflects on how self‑driving vehicles quickly become mundane, draws parallels to the slower timeline for fully automated household robots, and cites Asimov’s Three Laws as a guiding safety framework. - [00:11:06](https://www.youtube.com/watch?v=9GhT1mp5Edk&t=666s) **Pricing Strategy for Early‑Stage Home Robotics** - The speaker analyzes why a nascent home‑robot startup charges $20,000 for early access, linking the high price to the value of data, market testing against housekeeper costs, and current teleoperation staffing constraints. - [00:14:21](https://www.youtube.com/watch?v=9GhT1mp5Edk&t=861s) **Risky AI Hardware Costs & Piracy Concerns** - The speakers discuss the uncertain pricing and complex physical infrastructure of advanced AI systems, their strategic use for mindshare, and note a recent anti‑piracy complaint from Japan’s CODA to OpenAI. - [00:17:25](https://www.youtube.com/watch?v=9GhT1mp5Edk&t=1045s) **Synthetic Data, IP, and Revenue Sharing** - The speaker examines how intellectual‑property rights, royalty structures, and data‑permission marketplaces could operate for AI‑generated content, questioning the practicality of revenue‑sharing mechanisms and ownership when synthetic data is used. - [00:20:56](https://www.youtube.com/watch?v=9GhT1mp5Edk&t=1256s) **Navigating Style Boundaries and Revenue Shares** - The speakers debate the ambiguous definition of artistic styles in synthetic data, model providers’ cautious use of copyrighted material, and the potential for third‑party platforms to broker revenue‑share agreements between artists and AI model owners. - [00:23:59](https://www.youtube.com/watch?v=9GhT1mp5Edk&t=1439s) **Balancing Prompt Censorship and Model Interpretability** - The speakers discuss the tension between restricting user prompts to prevent harmful content, the lack of clear governance in the AI race, and the need for societal frameworks and interpretability research to guide what is permissible. - [00:27:28](https://www.youtube.com/watch?v=9GhT1mp5Edk&t=1648s) **AI Firms Forge Multi‑Cloud Alliances** - The speakers discuss how companies like Anthropic and OpenAI are forming overlapping partnerships with AWS, GCP, and Azure, creating a complex multi‑cloud ecosystem that adds technical complexity but offers diversification and strategic advantages. - [00:30:44](https://www.youtube.com/watch?v=9GhT1mp5Edk&t=1844s) **OpenAI's Commitment to Nvidia GPUs** - The speaker argues that despite diversifying cloud providers, OpenAI remains tied to Nvidia hardware because its models are heavily optimized for those chips, making any switch costly and technically challenging. - [00:33:56](https://www.youtube.com/watch?v=9GhT1mp5Edk&t=2036s) **OpenAI Inference Focus for Enterprise** - The speakers argue that OpenAI’s new GPU resources will primarily serve inference—enabling agentic workloads for business customers—and debate whether this capability will be offered directly on cloud platforms like AWS or via a marketplace, noting the lack of clarity around proprietary model availability. ## Full Transcript
It's such an early, early product in such an early
space that at this moment in time, I think anyone
who's trying to price this, they really are just trying
to see where the market's at. All that and more
on today's Mixture of Experts. I'm Tim Hoang and welcome
to Mixture of Experts. Each week Moe brings together a
panel of the sharpest minds in technology to distill down
what's important in the latest news in artificial Intell intelligence.
Joining us today are three incredible panelists. So a very
warm welcome to Ash Minhas, lead AI advocate, Ambiganasan partner
AI and analytics and Sandy Besson, AI research engineer. Welcome
to you all. Really exciting and interesting episode today. We're
going to cover all sorts of different aspects of what's
been happening in the news. First, we're going to talk
a little bit about 1x Neo, which is a newly
announced humanoid robot and the sort of Wall Street Journal
review of it. We'll talk a little bit about an
interesting challenge to OpenAI's Sora 2 from Japanese copyright holders.
And then finally we'll talk about a big partnership between
AWS and OpenAI. But first we've got Illy with the
news. Hey everyone, I'm Illy McConnon, a tech news writer
for IBM Sync. I'm here with a few AI headlines
you might have missed this week. AI giant Perplexity will
pay $400 million to integrate its AI powered search engine
directly into the Snapchat app so the social media platform
can be used for AI search. In addition to sending
snaps those images and messages that disappear. Two trends collide.
Crypto platform Coinbase is giving AI agents their own crypto
wallets to buy things on behalf of customers instead of
simply recommending purchases. Instacart has launched a suite of AI
enterprise tools for grocery stores so retailers have a real
time view of what's on their shelves at any moment.
This also means that shoppers will have an AI assistant
for personalized meal planning and budgeting. Google is launching AI
chips into space on solar powered satellites to test out
solar powered AI, a project aptly called Suncatcher. Want to
dive deeper into some of these topics? Subscribe to the
Think newsletter linked in the show notes. And now back
to the episode. The first thing I want to start
with is kind of this video that got passed around
wildly on social media this week that was covering the
One X neo. Now, longtime listeners to the show will
remember that we actually talked NEO when it was first
announced, but it is finally Open for sales, you can
go and buy it. It's being offered as a $500
a month subscription or a $20,000 early access fee. And
what 1x Neo is, it's a humanoid robot. The idea
is it's literally a life size, sort of mannequin style
robot which will be in your home and help out
with home tasks. And so I guess, Ash, maybe I'll
kick it over to you first. Does this have legs?
In 2028, are we going to all have sort of
a humanoid robot in our home? Or do you think
this is going to end up being much more of
like a sort of specific use case? As we see
this company launch its. First products, I think that they
did a really, really good job of their video introducing
their product to the world. The Wall Street Journal video
showed huge discrepancies in terms of what its capabilities are
today versus what they showed as the art of the
possible. I think that there's probably more than a year's
worth of development work to go into putting something like
that in my home, where I'm going to trust it
with my glasses in my kitchen. And my casual, breakable
things. In your kitchen? Yeah, yeah, yeah, exactly. I think
that the fact that during the Wall Street Journal interview,
the founder was very open about the fact that there's
teleoperation happening is an eye opener. I think for us
as a society to realize that a lot of these
robotics that they're having to deal with very, very complex,
ambiguous environments, we're just not there yet with the AI
technology to power them to do that job accurately. And
I do think that there's probably quite a number of
years to go yet. Yeah, for sure. And Sandy, I'm
wondering if you can give us kind of like maybe
an intuition for why it is so difficult. Right. Like
we've seen these huge explosions in AI capabilities. But as
yet. Right. Like I think the Wall Street Journal review,
which was widely passed around, you know, the robot is
depicted trying to basically like open and close a dishwasher
door. Dishwasher. And it's like. Yeah, yeah. And it takes,
it takes like minutes to do what's, what's so difficult
about, about this task from an AI standpoint. If we
look at where we've seen robotics prosper, it's in settings
where the job is quite monotonous. Right. It's quite routine.
Like in Amazon factories, is, is like famous for having
largely robotics and robot. What's the word? Roboticized. Is that
a word? Maybe it will become a word, but it's
famous for having that. Right. But when you're taking more
general tasks, everyone's home is unique. Not just has a
different layout, like your Robo vacuum. Right. But has different
handles and different buttons and different makes of washing machines
and gas range stoves versus electric stoves and things like
that. Right. There's so much variance that just like Waymo
experienced with driverless cars, which are still only allowed to
go up to 25 miles an hour. And that's just
starting to change. Right. They're going through this evolution where
they're going to need to collect a lot of training
data. And that kind of poses the question of, like,
how are they going to do that? And do people
want that in their homes for them? Like, he had
this really interesting construct of Big Sister in the Wall
Street Journal view that I thought was a cool way
to look at it. Right. It's like, okay, we have
this negative concept of Big Brother, but is Big Sister
an okay concept because it's going to help you long
term? Yeah, absolutely. I think the Waymo comparison is a
really good one. And Ambie, I wanted to get your
thoughts on this is like, I mean, Waymo works now.
I mean, if you've been in San Francisco recently, you
can call a robot car, it rolls up, you can
jump in it. And I think for me, what made
me very confident about the future of it is that
after the second or third ride, it's completely boring. Like,
you don't even think about it at all. But that
took a really long time. And I guess the question
for you is if you think that, like, you know,
we're going to see a different, similar timeline on this
kind of thing, right? Where I think Google's been talking
about autonomous vehicles for, I don't know, over a decade.
I actually don't know when they started talking about it.
You know, do you. Do you think there's a similar
timeline here before we have really kind of like fully
automated kind of home robots? Yeah. And that's the way
that I feel about the WAYMOS as well. I've taken
a bunch of those. And, you know, I always say
that the most surprising thing about the WEMOS is that
it's so unsurprising. Right. Like, it's so normal that you
just don't feel anything about it out of the ordinary.
Right. Like, we've got a ways to go before we
get to that stage on home robotics. And, you know,
for me personally, and I think a lot of us
would feel this way as well. Growing up, I was
a Huge Isaac Asimov fan. And I, I always think
about the three laws of robotics, right? Don't do any
harm or through inaction, don't cause any harm, always safeguard
humanity. Don't do any stuff that's going to put yourself
in danger, things of that nature. I think there are
ways to go before we encapsulate all of those and
then make home robotics approach that space, right? So to
Sandy's point, Waymo and the autonomous vehicles work to some
extent because the search space is a little bit structured,
right? Like you're always going to have street signs, like
you're always going to have roads marked with lanes, right?
There are a bunch of these factors that are structured.
Whereas when it comes to home robotics, right, the search
space is so unstructured and vast and infinite, it's going
to take some time. I think there are some tantalizing
clues and aspects that I think we are saying, hey,
maybe we can use world models to simulate and synthesize
data and then maybe we'll do. But there's a bunch
of maybes here that we got to figure out. So
there is. If you look at it from multiple factors,
are you going to get enough data to do the
training? Are you going to have enough? Just think about
all the compute capacity that every time we hop on
this podcast we complain about the compute capacity getting constrained.
Infrastructure becoming the moat. No, just think about all the
infrastructure that you'll need to run these robots at scale,
right? There's so much that's pending to be built out
from that infrastructure layer perspective. And then the third piece
is, like I mentioned, all the safety aspects, right? Whether
you subscribe to the three laws of robotics or some
fashion of it, we'll have to come and codify and
say, okay, here's how we need to regulate and leverage
home robotics, right? And there is a lot to be
figured out there. So, yeah, this is going to be
a multi year journey. It's not a one and done
deal in a couple of years. Honestly, one thing I
was thinking of when I watched the Wall Street Journal
video was I was thinking, I wonder if the folks
of NIO have thought maybe we should just send the
robots to IKEA stores and just get them to walk
around the whole IKEA store for a while and just
get them to train. Like they could just walk around
opening and closing all the drawers and sitting on the
furniture and doing all that stuff and then probably collect
a mountain of training data doing that. And then they
come to your house, which doesn't have ikea and then
they'll fail. And Ash, if I could stay with you,
I do want to talk a little bit about the
data aspects of this, right? Because I think as you
mentioned, it came out in the review that a lot
of it is still tele operated and ultimately I think
the teller operation is to collect data on how you
might ultimately kind of navigate these complex spaces. I mean,
I think one of the things I hear from Ambi's
comments is basically that if anything, homes are even more
complex than trying to navigate the road. I guess, Ash,
with that all in mind, like why are they pricing
it so expensively, right? Like $500 a month is like
that's like two and a half ChatGPT Pro subscriptions, right?
And if you want to buy like the main one,
it's like $20,000 for early access. But the data is
so valuable, why aren't they just making this product like
20 bucks a month, Right? Because like the data is
really what they need in order to get this thing
to work. I really have no idea how the financing
has worked to get the startup to where it is
today. To know what they're looking at achieving by charging
this price. I think that it's such an early, early
product in such an early space that at this moment
in time, I think anyone who's trying to price this,
they really are just trying to see where the market's
at. And I think they probably looked at this and
thought, what does a housekeeper cost? What's that working out
to for a household every month? And they're looking at
what's that point at which, you know, this becomes more
cost effective than a housekeeper. And they're kind of trying
to anchor on the price and they'll get the product.
I do think that there probably will be some volatility
in that business model and pricing as this actually matures
over the coming years. But as I said, it's so
nascent right now. You're not going to have one of
these in your home. I was just going to say
I think they're constrained by their own constraints. Right? It's
like you said, the teleoperation. They have to have people
that are. Right now you're scheduling this on an app.
So if they make it so widely available now, they
have to get hundreds of teleoperators to do these things,
right. And train those people and scale it up. So
for an early access product, I think it makes perfect
sense. Even if it means that it takes longer for
the product to improve, that they're doing so in A
really scoped way that they can control. One interesting thing
that I noticed was that they were already anchored on
the fact that people should have this expectation that you're
going to have a companion application on your phone or
of anti climatic as it were. Why would you have
a robot in your house that you've got a schedule
to do something, you know? And it comes down to
that same constraint that you implied, Sandy. Right. They've got
to hire people to teleoperate these things. Right. And they
need shifts, I guess. And there's a whole labor model,
a. Virtual housekeeper right now that's kind of what it
is. You're outsourcing the body of a housekeeper, but you
have someone behind the scenes that's still doing the work
for right now. But that will change overall. And to
be honest, if I don't have to fold my clothes
and wash them and put them away in two years
time, I would be happy to give it access to
the viewings of my wardrobe. Well, we'll have to have
you back on in 24 months. That's a good prediction.
I mean, I think a lot of this is like
we probably shouldn't worry too much about the price piece
here. A lot of this I think is also to
capture the mindshare. Right. I mean there is a reason
now we are talking about all of this and they're
capturing the mindshare. Right. I think robotics as Optimus has
been edging at the top of the mind share figure
has been pushing it. Yes, they're operating in the industrial
setting. The home robotics space was a little bit open.
I think this has gotten a lot of us talking
about it and capturing mindshare. Right. Pricing to Ash's point
I think is a little bit of a dart on
the boat right now. No one really knows. Yes, granted
there is a lot of sophisticated instrumentation that's needed to
make something of this nature work. Right. So it's not
just purely, hey, I'm collecting data and then I'm training
data and I'm just inferencing it somewhere. There is a
lot of physical equipment that's needed in order to make
this work. Right. Like you'll have sophisticated actuators and gears
doing this in precise fashion. So there is, I get
the cost of operations and cost of manufacturing probably is
a factor in there. But bottom line, right, it is
a little bit of a dart on the board. This
whole, I think push and declaration is a little bit
to capture the mindshare and then declare to the world
that, okay, we are, we are taking a step forward
and we are going to go proceed forward. I'm going
to move us on to our next topic. So, super
interesting story came out of Japan earlier a week about,
about a week ago. There's an industry organization in Japan
called the Content Overseas Distribution association, or CODA for short.
It's an anti piracy organization that represents basically Japanese IP
holders. So Bandai, Namco, Studio Ghibli, like a lot of
your kind of favorite brands out of Japan are kind
of represented by this organization. And what's interesting is that
they sent a letter to OpenAI expressing concern about essentially
the use of their intellectual property in Sora 2 and
the generation of videos that kind of like implicitly sort
of like rely on that ip. And I think it
was such an interesting case. And I guess, Sandy, like,
I guess the question for you is how do you
think OpenAI should kind of navigate these types of discussions?
It's a really hard and tricky thing about how the
rights of these rights holders should be taken accounted for,
but also to give the freedom to innovate on the
technology. And I think there's just a really interesting set
of questions there. Totally. And the more I was thinking
about it, the more I was like, oh my God,
this is a large task. And there's a, a few
ways, if I was OpenAI, I would think about it.
One is, do we actually have to alter, and I'm
sure they're already thinking about this one, do we have
to actually alter our pipeline of how we tag and
add metadata to our training data to be able to
flag exactly where everything is coming from? And to some
extent they might do this, but they might not have
IP owners on there or certain things like that. And
then do I have to potentially, potentially in order to
in some ways make this go away, potentially do some
sort of revenue share in the end of the day?
That's what IP and royalties are about, right? They're about
sharing revenue. So is it possible that if I say
like, hey, make me a cartoon version in the style
Studio Ghibli that's tagged somewhere, if it's using Studio Ghibli
and there's some sort of revenue share there, that, that's
one way, if I were OpenAI, I would be thinking
about it, to kind of escape all of these large
lawsuits. But I think that also opens up the other
end of the side where it's like, okay, well who's
going to actually collect all this information? Is there opportunity
for a market share, a marketplace out there that essentially
has the rights of people that say, yes, I allow
you to use this data or you are not allowed
to use this data. But the further I kept thinking
about that, the more, and I'm a little bit on
a monologue here, but the further I kept thinking about
that, the more I realized is this just more argument
to move into synthetic data and then what happens to
IP and ownership? If it's almost like inception, right? Like
IP ownership, inception, where is it still owned by that
person? If it's influenced because it's created synthetically via that?
And so you kind of move into this like ownership,
inception. Right. So truthfully, I don't know where this is
going to end up. Ambi, what do you think about
all of this stuff? I think yeah, where I was
really hard to figure out how you would set up
kind of like some kind of payments infrastructure here. And
so is one possibility that a lot of these companies
just start investing a lot more in synthetic data. Right.
Like kind of the era of like we need to
scoop up all this data to train our models is
going to give way to. Well, just we're going to
try to do synthetic as much as possible. I don't
think anyone has solved this. Right. No one really has
a clear understanding or has a clear solution to any
of this at this point in time. It's really, really
murky waters. I'm going to put a little bit of
an enterprise lens on it because I talk to clients
and I deal with enterprises on a day to day
basis and when I look at it from their perspective,
you can't afford to have any of these being transmitted
back to enterprises. And there is two, there are two
forks in the road that I'm seeing. There's one set
of model providers, very early on they always said I'm
just going to go for completely copyrighted data, I'm not
going to go and scrape anything and I'm going to
go only use that. And that becomes a safe approach
and a safe path for enterprises to consume. And then
there is the other fork which is the likes of
Gemini and OpenAI have. Clearly it's very consumer facing and
therefore a lot of the imitation aspects that creep in
here. But for enterprises outside of a text modality, if
you're getting into image or video modality, something like this
is still a Landmine that they wouldn't want to touch
on so that that piece has to be solved. I
don't think there is a clear answer to it yet.
Maybe synthetic data, but even there, who draws the boundary?
At what point do you say this looks like a
Studio Ghibli style? No, this doesn't look like a Studio
Ghibli style. There is no quantitative way for you to
actually go and delineate that. So it's a little bit
of a fuzzy aspect over there. No hard and clear
answers over there. I think the safest approach some of
the model providers like Adobe have taken as saying, I'm
not even going to touch any of that before all
of this thing gets sorted out. I'm just going to
go train on just purely copyrighted data and then I'll
deal with that for the time being. So it's still
those two folks in the road that hasn't fundamentally changed
over the last year or year and a half, Right?
I think there is an opportunity to be made. Like
Sandy is saying, if the model providers are okay with
some of these rev share agreements, and I think there
could be an enterprising layer that creeps up between the
artists and the model providers. It may not even be
the model providers themselves. There may be a third party
that says, okay, you know what, I'm going to help
the artist form a consortium or form a network and
then I will help transact all of these between the
model providers and the artists. So I think there's some
new spaces to be covered here. Yeah, absolutely. And Ash,
maybe I'll give you the last word on this. I
mean, I think part of the big question is. Yeah,
I think it's like part of what's difficult about this
discussion is to what Sandy's saying. In principle it makes
a lot of sense, right? Like you contribute some data
and you should be compensated for the use of it.
In practice, it's really, really complicated. And I think one
of the questions is like how we say something is
so close enough in style that you deserve to have
some kind of compensation. Do you think at that point
we're going to just have to draw an arbitrary line?
It'll be like, oh, we're going to calculate a difference
vector between this video and this video and if you
are within this threshold, then you have to pay and
if you're outside this threshold, then you don't have to
pay. Some of this might just ultimately be maybe arbitrarily
resolved. Do you think that's where we land. With some
of This, I don't know if it is somewhere where
we're going to arbitrarily land. I think that there's too
much at play here, both commercially as well as culturally,
for there to be a place that either the technology
companies or the people who own the rights to a
lot of this art are going to let that happen.
I think that ultimately, at this moment in time, we're
still in a very nascent space, right? I remember when
stable diffusion came out in the first instance. And, you
know, as background, I'm a photographer, okay? And I've always,
like, over the last sort of 10 years or so,
shared very few pieces of my photography on the social
networks, because I'm like, hey, I'm transferring my rights away
from my work. And so we've now got to this
place now where these generative models, we need them to
be good, to be able to get value from them.
them to be able to respond accurately to what people
are asking for, right? And as Ambie already touched on,
at what point does that become censorship, right? Like at
one point when someone's prompting the model and saying, hey,
I want you to make me a picture of this,
at what point, if we just keep on putting in
sort of protections there at the prompt level to go,
hey, you can't prompt for this, you can't prompt for
this, you can't prompt for this. What I think about
is who's making those decisions, right? And I guess at
this moment in time, in this. This huge AI race
that we've got going on, no one really wants to
do that too much because that makes their model less
helpful. And so I think that this won't be arbitrary.
I think that this is just a lot of back
and forth that's going on between so many different stakeholders
right now. And we really need some sort of real
frameworks in place, ideally from a government and a societal
level, to make some decisions here as to what is
and isn't allowed. And people should abide by that. I'd
like to just add one little point is something that
Ash and Ambie both touched on is that I think
a lot of us look at it from the output
perspective, like the user perspective, like, what are we prompting
for? What does the model output show? But in reality,
there's an entire field called modern interpretability, right, where they're
trying to understand behind the scenes the reasoning of the
model. And they have. The famous example, I think, is
talking about what is the capital of France, right? And
they're seeing the nodes light up throughout the model. Right.
And so are we going to accelerate our interpretability of
the reasoning of the models to understand what's actually being
used to produce this output, or are we going to
assess it from the perspective of the output? And I
think right now we don't have a choice but to
perspective but to assess it from the output because we
don't understand the innards. Right. But as we start to
understand the innards, there's going to be even more of
a case for these other companies to say, hey, you're
using my stuff because some node got activated somewhere deep
in the neural network and then it will become even
more gray. Yeah, exactly. Yeah. I think that somebody almost
like, it's like the illusion of clarity is like once
we start digging, it'll be like, okay, well presumably some
people will come up with prompts that can get very
similar styles that don't activate certain kinds of neurons, and
that ends up becoming a new game. Right. Even as
the field kind of improves. So. All right, I'm going
to move us on to our last topic of the
day. Super interesting story, particularly on the backdrop of the
last MOE that we recorded for Halloween last week. Week
there's basically news that OpenAI has announced a new partnership
with AWS. And I will always kind of like quote
the numbers here because they continue to be mind blowing
to me. So OpenAI is doing a deal with AWS
that represents a $38 billion commitment, expanding its compute capacity
and basically working with AWS to expand on its infrastructure.
This is an interesting one. Amazon, of course, has been
touting its Trainium, its, its specific chip, but this one
seems very much still focused on Nvidia GPUs, at least
from the blog post and I guess maybe Ambi, I'll
throw it over to you first. It seems like one
thing we keep talking about on MOE is how you're
starting to see these alliances form where you're like, okay,
you've got Anthropic and they're getting close with GCP, but
they're also a little bit close with AWS. And then
of course OpenAI is very close with Azure and Microsoft,
and now they're also getting close with aws. I think
the most interesting thing and maybe a good place to
start is it seems like a lot of these companies
are going multi cloud ultimately, which seems like, I guess
if you want to give us an intuition for why
they're doing that, because my sense of it is that
that actually adds quite a Lot of complexity. Yeah. Yeah.
Well, yeah, I think alliances is the right word. It
always reminds me of like Game of Thrones and it's
so confusing. Who's alliance, who's in a royal marriage with
who. Exactly. Right. It's so exciting, right? A new episode
comes out and new. Things, you know, it's the nerdiest
possible Game of Thrones you could imagine. But yeah, I
mean, the $38 billion, if you think about the $1
trillion supposed IPO, I mean, this is like a drop
in the bucket. Right? But you know, leaving that aside,
right. I mean, sort of makes sense, right? You want
to diversify, you don't want to put all eggs in
one basket. So. Yeah, I mean, you have to diversify
like we were talking about in a previous episode, right.
The moat is shifting from pure model layer into the
infrastructure layer and some of the depths of the infrastructure
layer. Right. So knowing that, I think it makes sense
to ensure that you don't put all your eggs in
one basket and then just go with one infrastructure provider
and therefore you diversify into aws. What I found surprising
was there wasn't a reciprocal arrangement for AWS to host
the proprietary models from OpenAI and then expose them. That
is still an exclusive arrangement between OpenAI and Azure. So
it's a little bit of a one way imbalanced relationship,
I would say. What would have made it more interesting
was the models are also getting hosted on multiple different
environments and it becomes truly diversified. And you're getting to.
When I talk to enterprises, they've all moved on to
just going with one model, to having a mixture of
or a choice of picking their own models and cloud
providers. Just like we went with the hybrid cloud approach,
we are fully seeing the hybrid model approach, so I
would have liked to see some of that manifest here.
But bottom line, I think this is purely hedging your
bets and making sure that you're not getting stuck in
the infrastructure mode and getting caught there. Right. That's the
simple way to look at it. Sandy. One interesting aspect
of this is in contrast to the anthropic deal, which
very much kind of touted like we're working with TPUs,
I think the language is like we will use up
to a million TPUs or some crazy number like that.
This one really does still seem focused on Nvidia GPUs.
And I guess maybe it seems like one part of
this deal is OpenAI is not really. Maybe it's changing
and diversifying its infrastructure providers. But at the end of
the day, the Chips are still the same. That was
100%. What I was going to hit on is that
they've seemed to have chosen their hardware provider. They can
diversify in their cloud infrastructure, they can diversify in how
they offer it to end users, but they seem to
have clearly chosen their hardware providers. And the reason for
that is not all chips are created equal. Number 1
and 2 is models are optimized specifically for the hardware
that they run on. So clearly OpenAI has done a
lot of work in optimizing their models to run most
efficiently on these Nvidia chips. And so to make that
switch is no easy feat. That's like teams and many
hours and many weeks and potentially months of switching cost.
And is it worth that if AWS also offers the
thing that they're already optimized on? It's a really hard
call. And I guess part of it is just like,
of optimize outside of it, I guess. Ash, is this
ultimately, like, how much of a difference do you think
this makes going forwards? This is sort of the direction
you'd expect these companies to be moving in, or, you
know, how much of this, like, I think a little
bit about how, like Nvidia or not Nvidia, but Netflix
used to run a lot of its infrastructure on aws.
And, you know, it seems possible to me that over
time, like, OpenAI could become like much more of an
AWS provider or a consumer. You know, do you, I
guess, in the battle of the clouds, do you feel
like one has an advantage over the other? I don't
know whether I would look at it with that lens
of it being a battle of the clouds. The way
I'm looking at it is at a macro level, they're
doing so much stuff with AI that they need all
the GPUs that they can get their hands on. Okay.
And what I found really, really interesting in this specific
this to rapidly scale agentic workloads. And so I'm wondering,
what are these agents doing? Right. Okay. And I don't
think it's sort of a battle of the cloud per
se, Tim, but more a case of they're using different
infrastructure providers for doing different types of work. Right. And
I mean, I honestly don't know what those agentic workloads
are doing, but I think it just comes down to
there's not enough GPUs available. They need to get something
done. They're a rapidly scaling company. AWS has got loads
of GPUs that they're willing to sell them and they're
like, great, we can take this portion of work that
we need to do and we can just go scale
it over there quickly. Right. It's like almost like any
GPU in a storm. Just like whatever is available, we'll
take it. If I say what Ash said, I think
you've hit the nail on the head in terms of.
It's for a different point. I don't think they're doing
training on these GPUs. I think they're doing inference and
that means that they're moving more into their OpenAI for
business play with it. Because what do businesses use? They
use cloud. Right. And so as they scale more into
their enterprise piece, if I had to take an educated
guess, I would guess that they're using this compute power
more for building agentic workloads with their clients and that's
where they're going to use that inference for. Yeah, tbd,
I have slightly different take on that. Like I said.
Right. Again, they haven't given us a lot of details.
So yes, enterprises will use, Azure will use AWS, all
of that, but unless OpenAI is exposing them on the
AWS environment, then it's not going to be consumable. So
the big question is, are you just using it to
run your workloads or are you going to actually offer
a marketplace or are you going to offer a way
to consume it on aws? If it is the latter,
then yeah, I mean, that's fantastic. Right? I mean, that
would be fantastic for enterprise consumers. So I think they're
still figuring out a bunch of things here as well.
Maybe it's a simple. They do have OSS models. They
do have OSS models. Not the proprietary version. Right. That's
what I'm. Proprietary ones. That's what I was hoping to
look for. And I found that there is no mention
of that there. Yeah, mysteriously not mentioned. And this is
what led me to make my point to say that
I just think that they're like, hey, They've got enough
GPUs for these workloads that we want to run. Let's
do a deal and let's just use these GPUs because
that's why they specifically called out these agentic workloads. But
like I said, I really want to know what those
agentic workloads are. Yeah. It's like we're all looking for
this N dimensional chess, but it's just like we just
need more chips. We just need more cloud as maybe
the primary driver. Yeah. What are they doing with them?
That's what I want to know. Well, we're always asking
the hard questions here on Moe, and that's a great
note to end on. So that's all the time that
we have for today. Ash, Ambi, Sandy, always great to
see you on the show. And hopefully we'll have you
on at MOE very soon. And thanks to all you
listeners. If you enjoyed what you heard, you can get
us on Apple podcasts, Spotify and podcast platforms everywhere. And
we'll see you next week on Mixture of Experts.