Exploring Open‑Source Mixture‑of‑Experts AI
Key Points
- The show opens by questioning the notion of truly autonomous AI, emphasizing that models only predict tokens and require external control to act.
- Recent AI news highlights include OpenAI’s $1 trillion data‑center plan, Alibaba’s partnership with Nvidia on robotics and self‑driving cars, IBM’s PDF‑decoding model topping Hugging Face downloads, and Meta’s AI‑powered digital dating assistant.
- The main discussion centers on Tongi Deep Research, a new open‑source LLM with 30 B total parameters but only 3 B activated per token, optimized for long‑horizon information‑seeking tasks.
- Panelists note that Tongi’s “task‑trained” approach and efficient activation make it a significant step forward for agentic models that can handle extended reasoning.
- The episode also promises to cover additional topics such as AP2, AI‑related safety concerns, the latest AirPods, and a massive $100 B investment in AI initiatives.
Sections
- Debating AI Autonomy & Industry News - The segment opens the Mixture of Experts podcast by questioning autonomous AI, introduces a panel of experts, and previews a rapid‑fire news roundup that includes OpenAI’s trillion‑dollar data‑center plan.
- Closing Gap Between Proprietary and Open‑Source AI - The speakers note that open‑source models are rapidly catching up to proprietary LLMs, but the primary differentiation now lies in specific use cases such as embedded devices, privacy‑focused deployments, and costly deep‑research agents.
- Diverging Paths of AI Ecosystems - The speaker argues that open‑source AI will continue to split from proprietary, product‑centric solutions by focusing on narrow components within broader agent architectures, and wonders if upcoming research papers will spark new trends like model distillation and continual pre‑training.
- Layered Protocols and Business Viability - The speakers debate the difficulty of evolving a niche “punk‑rock” lab into a market‑ready, self‑sustaining company before pivoting to discuss Google’s new AP2 protocol, which stacks atop existing agent frameworks to standardize authentication, authorization, and payment handling for AI agents.
- AI‑Driven Shopping Meets Crypto Payments - The speakers evaluate a camera‑based auto‑shopping concept, argue that major players like Google have a strategic edge in integrating crypto payments and AI, and note how prior smart‑contract knowledge becomes valuable in this emerging market.
- Hidden Protocols, Liability, Market Share - The speakers argue that while everyday users remain unaware of and indifferent to underlying security protocols such as AP2, developers and liable institutions will care, and Google’s integration of an ATA‑based secure payment system could force widespread adoption and boost its market dominance, with rivals like Anthropic possibly working on comparable technology behind the scenes.
- Debating AI Risk Book - The hosts introduce the provocative AI safety book “If Anyone Builds It, Everyone Dies” and examine its claim that super‑intelligent AI poses existential danger, questioning whether AI CEOs are being sufficiently cautious.
- Balancing AI Autonomy and Safety - The speakers argue that strict human oversight and limited use‑case deployment are essential to prevent self‑propagating AI scenarios, citing GPT development safeguards and autonomous‑vehicle mishaps as cautionary examples.
- Untitled Section
- Fiction, Reality, and AI Wearables - The speakers explore how AI blurs the line between fictional storytelling and actual capability, then shift to highlight emerging AI‑enabled wearables such as Meta’s Ray‑Band and Apple’s real‑time translation AirPods.
- Apple’s AI Comeback Debate - The speakers speculate on Apple’s potential turnaround in artificial intelligence, compare its progress to Google’s, consider ecosystem integration, and question the underlying models behind features such as AirPods translation.
- Seamless AI Integration via UX - The speakers argue that AI must become invisible to consumers—abstracting tokens and model details—by delivering flawless, stylish user experiences, a strategy they see Apple poised to capitalize on.
- Nvidia's $100 Billion OpenAI Investment - The speakers debate the ramifications of Nvidia's announced $100 billion infusion into OpenAI, likening it to stock buybacks and exploring its financial and strategic implications.
- AI Alliances, Energy Demands, and Tribal Dynamics - The speakers debate upcoming AI partnerships and the massive 10‑gigawatt power needs of new facilities, likening the shifting loyalties to school‑house rivalries.
- Debating OpenAI's $100B Allocation - The panel speculates whether the massive funding will primarily expand data‑center capacity or fund more efficient next‑generation chips, framing it as a push toward greater raw compute and a hint of forthcoming AGI.
Full Transcript
# Exploring Open‑Source Mixture‑of‑Experts AI **Source:** [https://www.youtube.com/watch?v=h1YqbXQN5N8](https://www.youtube.com/watch?v=h1YqbXQN5N8) **Duration:** 00:52:38 ## Summary - The show opens by questioning the notion of truly autonomous AI, emphasizing that models only predict tokens and require external control to act. - Recent AI news highlights include OpenAI’s $1 trillion data‑center plan, Alibaba’s partnership with Nvidia on robotics and self‑driving cars, IBM’s PDF‑decoding model topping Hugging Face downloads, and Meta’s AI‑powered digital dating assistant. - The main discussion centers on Tongi Deep Research, a new open‑source LLM with 30 B total parameters but only 3 B activated per token, optimized for long‑horizon information‑seeking tasks. - Panelists note that Tongi’s “task‑trained” approach and efficient activation make it a significant step forward for agentic models that can handle extended reasoning. - The episode also promises to cover additional topics such as AP2, AI‑related safety concerns, the latest AirPods, and a massive $100 B investment in AI initiatives. ## Sections - [00:00:00](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=0s) **Debating AI Autonomy & Industry News** - The segment opens the Mixture of Experts podcast by questioning autonomous AI, introduces a panel of experts, and previews a rapid‑fire news roundup that includes OpenAI’s trillion‑dollar data‑center plan. - [00:04:32](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=272s) **Closing Gap Between Proprietary and Open‑Source AI** - The speakers note that open‑source models are rapidly catching up to proprietary LLMs, but the primary differentiation now lies in specific use cases such as embedded devices, privacy‑focused deployments, and costly deep‑research agents. - [00:07:40](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=460s) **Diverging Paths of AI Ecosystems** - The speaker argues that open‑source AI will continue to split from proprietary, product‑centric solutions by focusing on narrow components within broader agent architectures, and wonders if upcoming research papers will spark new trends like model distillation and continual pre‑training. - [00:12:38](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=758s) **Layered Protocols and Business Viability** - The speakers debate the difficulty of evolving a niche “punk‑rock” lab into a market‑ready, self‑sustaining company before pivoting to discuss Google’s new AP2 protocol, which stacks atop existing agent frameworks to standardize authentication, authorization, and payment handling for AI agents. - [00:16:48](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=1008s) **AI‑Driven Shopping Meets Crypto Payments** - The speakers evaluate a camera‑based auto‑shopping concept, argue that major players like Google have a strategic edge in integrating crypto payments and AI, and note how prior smart‑contract knowledge becomes valuable in this emerging market. - [00:20:05](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=1205s) **Hidden Protocols, Liability, Market Share** - The speakers argue that while everyday users remain unaware of and indifferent to underlying security protocols such as AP2, developers and liable institutions will care, and Google’s integration of an ATA‑based secure payment system could force widespread adoption and boost its market dominance, with rivals like Anthropic possibly working on comparable technology behind the scenes. - [00:23:32](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=1412s) **Debating AI Risk Book** - The hosts introduce the provocative AI safety book “If Anyone Builds It, Everyone Dies” and examine its claim that super‑intelligent AI poses existential danger, questioning whether AI CEOs are being sufficiently cautious. - [00:26:54](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=1614s) **Balancing AI Autonomy and Safety** - The speakers argue that strict human oversight and limited use‑case deployment are essential to prevent self‑propagating AI scenarios, citing GPT development safeguards and autonomous‑vehicle mishaps as cautionary examples. - [00:30:08](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=1808s) **Untitled Section** - - [00:33:21](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=2001s) **Fiction, Reality, and AI Wearables** - The speakers explore how AI blurs the line between fictional storytelling and actual capability, then shift to highlight emerging AI‑enabled wearables such as Meta’s Ray‑Band and Apple’s real‑time translation AirPods. - [00:36:40](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=2200s) **Apple’s AI Comeback Debate** - The speakers speculate on Apple’s potential turnaround in artificial intelligence, compare its progress to Google’s, consider ecosystem integration, and question the underlying models behind features such as AirPods translation. - [00:40:41](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=2441s) **Seamless AI Integration via UX** - The speakers argue that AI must become invisible to consumers—abstracting tokens and model details—by delivering flawless, stylish user experiences, a strategy they see Apple poised to capitalize on. - [00:43:51](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=2631s) **Nvidia's $100 Billion OpenAI Investment** - The speakers debate the ramifications of Nvidia's announced $100 billion infusion into OpenAI, likening it to stock buybacks and exploring its financial and strategic implications. - [00:47:01](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=2821s) **AI Alliances, Energy Demands, and Tribal Dynamics** - The speakers debate upcoming AI partnerships and the massive 10‑gigawatt power needs of new facilities, likening the shifting loyalties to school‑house rivalries. - [00:51:12](https://www.youtube.com/watch?v=h1YqbXQN5N8&t=3072s) **Debating OpenAI's $100B Allocation** - The panel speculates whether the massive funding will primarily expand data‑center capacity or fund more efficient next‑generation chips, framing it as a push toward greater raw compute and a hint of forthcoming AGI. ## Full Transcript
premise of a an AI that is capable of
acting autonomously
is one that I struggle with. AI models
fundamentally do nothing except predict
the next token. AI systems
do things and somebody has to turn them
on and turn them off. All that and more
on today's Mixture of Experts.
[Music]
I'm Tim Huang and welcome to Mixture of
Experts. Each week, Moe brings together
a panel of the sharpest minds and
quickest wits to help you digest the
week's news in artificial intelligence.
Today, I'm joined by a championship
lineup crew. We have Mihi Crevetti,
who's a distinguished engineer, Aentic
AI, Gabe Goodart, uh, chief architect,
AI, open innovation, and Sandy Besson,
AI research engineer. Welcome to you
all. We've got a very, very packed
episode. But I think I was just informed
by one of our producers. We're going to
try to wedge another story in. We're
going to talk about Tongi Deep Research.
We're going to talk about AP2. We're
going to talk about AI's killing us.
We're going to talk about the new
AirPods. And finally, we're going to
talk about a small hundred billion
investment. But first, we've got Eiley
with the news. Eiley, over to you.
[Music]
Hey everyone, I'm Eiley McConnan. I'm a
tech news writer for IDM Think. I'm here
with a few AI headlines you might have
missed this week. Another news story
involving trillion with a T. Yes, that's
trillion. Open AI has unveiled plans to
build one trillion dollars worth of data
centers across the US and abroad. Can
tech companies be frenemies? Chinese
tech giant Alibaba is partnering with
American chipmaker Nvidia. Alibaba will
use Nvidia's hardware and software to
develop robotics and self-driving cars.
Turns out there's a world of useful
information in those PDFs that most
people never read. Granite Dockling, a
small but mighty new model from IBM, can
decode those PDFs and has jumped to the
top of the most downloaded models list
on hugging face this week. Are you
suffering from swipe fatigue from online
dating apps? Well, Meta has added a new
digital dating assistant to Facebook so
you can use AI to identify your special
somebody more easily. Want to dive
deeper into some of these topics?
Subscribe to the Think newsletter. It's
linked in the show notes. Now, back to
the episode.
So, for our first segment, I really
wanted to talk about Tongi Deep
Research. Um, so this is a model that
has zoomed to the top of the hugging
face leaderboards. It's an agentic large
language model that quote features 30
billion total parameters with only three
billion activated per token. And it's a
lab model that's specifically designed
for long horizon deep information
seeking tasks. Um, and best of all, it's
all open source. So, Gabe, given our uh
you're our open source guy, maybe I'll
kick it over to you first. Initial
impressions about this model. What do
you think? I think it's a really cool
step forward and I think it the novelty
in this model is that it was taskrained
specifically for this long horizon and
paired with software that was designed
to implement the long horizon search in
an iterative recursive way. Um and I
think that's really cool. So you know at
this point generating tokens is not
novel. um your mixture of experts is not
novel but the shape of those tokens that
are generated and the patterns that it
was trained on are novel and that's
pretty cool. Um I specifically think
their heavy mode uh is a really
interesting doubling down on a specific
agentic architecture around state
management and consolidation that we
haven't seen. We've certainly seen
individual agents built to do this
pattern but we haven't seen them paired
with a purpose-built model. And I think
um you know there are a lot of great
public uh implementations of some flavor
of deepness in research out there that
can run on your laptop. I use them
regularly. But I will say, you know,
they don't measure up to uh public
Frontier model implementations. Uh, and
I think it's really cool that this team
is is pushing the boundary there and
trying to get something that you can run
on a powerful workstation to measure up
to, um, you know, a frontier research
system. So, uh, kudos to, you know,
really pairing the model implementation
and the software ecosystem around it. I
think that's awesome.
>> Yeah, I think it's a really big kind of
development. And Mihi, maybe I can call
on you to talk a little bit about
trends. I think one of the things we're
always watching ate is like sort of this
frontier right between sort of like the
proprietary model and open source
catching up and it feels like over the
last few months that just continues to
narrow and narrow and narrow and narrow.
Um, and so like I guess maybe this is
always a good data point to check in.
Like do you feel pretty soon like this
this margin is going to disappear,
right? Like it really will be that every
time a proprietary company proprietary
model comes out we're going to see a
open source implementation that's almost
as good like almost immediately.
>> I think it's likely although if you look
at the use cases that's where we see the
biggest differentiation between these
models. A lot of the use cases we have
for these smaller models, especially the
open source models that can compete in
the space, have to do with embedded
devices, with local laptops, with
privacy, being able to run these models,
you know, behind a firewall or running
them at a cost. If you look at the use
case of deep research, since we're
talking about agents and deep research,
that's basically a large language model
that's doing planning to use tools like
both internet search and internet search
as well as searching all of your
document collections and really going
through potentially millions of tokens.
That can get expensive and that can get
very slow if you're using one of the
frontier models. So there's a lot of
organizations who are looking for doing
this cheaper, faster with smaller
models. preferably behind their firewall
especially if you have some of these
models work with financial data or HR
data or internal data. So I definitely
see a sweet spot for many of these uh
smaller
I would say purpose-built models.
>> Yeah, for sure. I think almost you're
saying like we think about it as like oh
is open source catching up? I guess Mi
are you do I hear you right in kind of
saying almost like we should think about
these as like almost two different
markets like what open source trying to
do actually pretty different from like
what like I don't know open AI is trying
to do
>> I believe so and look open AI also has
their you know GPT5 nano which is you
know too cheap to meter 35 cents for a
million token or I'm probably leaving a
comma left or right in their in the
price but again too cheap to meter I
don't care if I'm using it for agents I
know it's fast I'm going to know it's
going to be cheap enough and I can throw
it at use cases like deep researcher and
say, "Yeah, it's fine to crawl through a
hundred of these websites and send every
one of those tokens." But maybe the CIO
office is not going to agree with me
sending all of that internet data to one
of these large language models. So, if
there's a model that can run it
privately or a model I can run on my
laptop or maybe even in the future on my
phone, that's awesome.
>> Sandy, I want to bring you into this
conversation. One of the things that we
talked about on the last episode was
that there was this paper that came out
that was some called something like what
are people using chatbt for? Um, and I
thought the funny part about it was like
everybody looked at it and I think our
guests were all like it's search. People
are using chatpt for search. Um, and I
guess to Mihi's comment, it almost kind
of feels like what we might see if we
had a paper that was like what are
people using open source models for is
like it would look very very different.
And I guess kind of question for you is
like do you feel like these ecosystems
are going to just keep diverging over
time like that over time we should
actually really think about like open
source almost solving like a completely
different set of problems from what you
know the like a chat GPT is trying to
solve. I think it's the the
juxtaposition of productizing versus
plugging into a broader more like bigger
problem, right? Typically in open-source
solutions, they're not solving like the
whole suite. They're solving like a very
narrow piece of the puzzle. So, I could
very much see using this uh deep
research agent as part of a broader
agent team or broader agent
architecture, but it's probably not
going to be like the full picture that
stands alone and is like a hosted
solution somewhere, right? Um, and so I
I kind of see it diverging in that way.
You know, like I also wonder slightly
tangentially, but like I wonder we we
see we saw was it this year or last
year? with deepseek. I can't even
remember timelines anymore. But uh we
see um like once in a while these big
papers come out and I don't know whether
this one will will be that kind of big
paper for this type of of trend or
pattern but
I think just like we saw Deep Seek kind
of emphasize distillation and now
distillation became a big deal after the
Deep Seek paper came out. I wonder
whether this paper would trigger some
sort of trend in terms of that triathlon
of training where you do like continual
pre-training then fine-tuning and then
on policy RL like I wonder whether this
will now become a common pattern because
of this paper.
>> Yeah, that's right. And that's actually
a fun way of thinking about it. I we
haven't really talked about it like that
on the show before. I think frequently
we're very like benchmarked, right?
We're like, "Oh man, this model is like
so much better on this benchmark and not
as good on this benchmark." Sandy,
you're almost saying like we should
almost measure how important a model is
by almost like how influential it is,
right? In terms of like how it kind of
shapes how people are doing things.
>> Totally. Sometimes it's not the first
model that comes out that's actually
like quote unquote the the the best. It
might be soda in some ways, but like
it's actually the trend that it drives
that like points us in a different
direction. Gabe, maybe a final question
on this before we move to our next
topic. I wonder a little bit about
business model here. Uh so if you're
Tongi Labs, right? You've just done this
open source thing. Um at some point, do
you also want to go closed source? Like
I think we've been talking a lot about
the big kind of like open AI and
anthropic, you know, kind of all sort of
leaning and thinking more about open
source. But I wonder if the ecosystem
goes the other way at some point where
like leaders in open source eventually
start to go more closed as well or if
you know almost like constitutionally
it's like a different direction.
>> So I I actually really like something
that you said Sandia and I'm going to
pick on it here which is like it's
really this two directional like the
closed source products are trying to
give something to you like here's a
thing you should have it the open-source
tools software and models are trying to
plug into an existing ecosystem. So, do
I think that labs that are primarily
open source are eventually going to want
to go closed source, too? I think that's
a possibility. I think if they come up
with something genuinely novel or
something that they really think, you
know, we could actually put this in
front of people and people would be
willing to pay us for it and they can't
get it anywhere else, we have a shot at
taking the crown on some novel
capability, they might. I think the
other route and especially that I've
seen when talking with other uh small
organizations that have an open core uh
the other route is really leaning into
that plug-in portion and that by that I
mean an enterprise tier of some variety
right and so engaging with your same
open source componentry possibly with
some additional componentry around it uh
for meeting enterprise needs like better
authentication management better data
sovereignty etc um and engaging in
almost like a consulting type of view
with a large enterprise client that
allows them to have a revenue stream for
their software while continuing to push
the envelope in the open. So, I think
that's the business model that I have
seen more often for these types of
shops. Um, and the other thing you said,
Sandy, that I also loved was like
there's also just the influence game
like especially depending on the life
cycle of these startups. Um, a lot of
times the metric isn't dollars, it's
likes, clicks, retweets, uh, download
and download, stars, all all of the
social metrics, right? And I think, um,
I don't frankly know quite where uh, you
know, we're at here, but I think
depending on the journey of whether or
not they're being measured on dollars,
um, it may still just be in the can we
move the needle on the influence game.
Um, and I think that in and of itself is
a business model that has a decent
chance of panning out either in an
acquisition or in enterprise deals. Um,
I think the leap to being a private
selfpropelling entity that has products
you offer to the market is a tough one
to make.
>> Yeah, for sure. Yeah. I want to create
like the most like punk rock lab which
is like a disaster from a business
standpoint but is just like incredibly
influential.
>> Exactly. Exactly. most stars of any lab.
You know,
>> you're not alone there. There are labs
that are literally doing that.
>> That's right.
>> I'm going to move us on to our next
topic. Um I always joke that, you know,
a space is really maturing when people
are introducing protocols to sit on top
of other protocols. Um and we have a
good example of that from this week. Uh
Google announced a new protocol that
they are calling AP2. Um, and AB2 builds
upon a lot of existing agent frameworks,
uh, like model context protocol and
other things that we've we've talked a
lot about. Um, and what AP2 is
attempting to do is basically set up a
common structure for agents to engage in
commerce and payments online. Um, and
sort of, you know, in the blog post,
Google sort of describes what they're
really trying to do is make sure that
agents can, you know, have proper
authorization, be authentic, right, and
be sort of accountable to the fact that
they're going to move money around for
people in the future. And the most fun
part about it is that AP2 explicitly
contemplates a situation where you are
not there when your agent goes out and
does its stuff on the internet. Um, and
what they're proposing is a thing that
they call the intent mandate, which is a
cryptographically signed record of what
you want the agent to do. And that's
kind of how they instantiate like, okay,
what's the scope of authority you're
giving the agent to do when it goes out
there and, I don't know, you know, puts
a bunch of books into a shopping cart
and, you know, checks out for you. So,
Mihi, I guess you're our agents guy, uh,
I think on this panel. Um, how big of a
deal is AP2? Like, do you think it's
going to get big up adoption? Do you
think people should be paying attention
to it? I think it's got huge potential
because it kind of tries to solve a
problem that neither A2A or MCP solve
today. And the reality is if you look at
the ecosystem, MCP is solving the
problem of building your tool once and
having that tool work with any kind of
agentic framework. So you're kind of
decoupling the agent and the tool. But
all of the respective security,
authentication, authorization mechanisms
that you're going to use, how you're
handling secrets, secrets management,
data management, certificate management
is really left up to you in terms of how
you implement it or how you make it
work. And there's already like five
versions of MCP and they all implement,
you know, more and more and more around
the area of security. But I don't think
we're at a point where we can say, "Yep,
this is something I can easily build a
payments processing system on top of or,
you know, have an agent go off and book
my travel to some island and, you know,
make all the payments and trust it with
that information. Not even from the
perspective of trusting it to get it
right, but from a perspective of how
it's handling that data. You can't put
your payment information, your credit
card information in the text that goes
into the model. There needs to be some
kind of a side mechanism on how the data
is transferred cryptographically
verified, not up to the interpretation
of a large language model. So I think
what Google did here was very smart. Um
they tried to leapfrog the effect that
MCP got because MCP has been adopted
virtually everywhere. Microsoft, OpenAI,
IBM and so on. 8way probably less. So,
so now they went to all of the banks,
all the financial institutions and
especially some of their core client
base both in the finance industry but
also in the advertisement industry cuz I
can see this take up for example for you
know those single person shops who are
using Tik Tok as their mechanism of
advertisement and say hey talk to your
personalized shopping assistant. He's
going to pick the right clothes for you.
you take a picture with your camera and
it's going to buy you a new, you know,
kind of dress or pants or whatever every
single month and ship it back to you and
you give it the authority to make those
purchases on your behalf. So, I I think
it could potentially open up a new
market, a market they're very competent
in, where they have potential both
clients and relationships in where I
don't think Entropic, for example, has
the same kind of leverage within the
same institutions. they're not a payment
provider like Google has Google Pay. So
I think it's a brilliant move. I do
think some of the items are so so
they're also releasing X4 for two
extensions for 8way.
>> So this is going back to crypto
payments.
>> So we're going back to web free and
MetaMask and the Ethereum foundation and
coin bags. It's like wait I thought we
were done with the web free stuff or
we're now doing AI. Come on one trend at
a time. So, I think it'll be
interesting. Uh, it'll be even more
interesting if we see somebody like
Facebook and their metaverse get excited
about it, but time will tell.
>> I'm so glad I spent those two weeks
about a year and a half ago figuring out
how to write smart contracts. That might
be useful now.
>> Yeah, exactly. Exactly. It was all worth
the investment of time and effort. Well,
so there's two directions to go and I
think I want to hit on both of them. I
guess one of them that just to build on
Mah what you just said was thinking a
little bit about like I hadn't really
thought about basically like this is
sort of competition between um kind of
anthropic and Google in some sense. Um
if you're anthropic like what is the
counter move here? Do you really care?
>> I mean we'll have to see if entropic
adopts it.
>> That would be an interesting move
because you know we've seen that even
OpenAI for example has adopted MCP now
and you can go in chat GBD and use that.
So we're going to see if we see similar
adoption from Entropic. Maybe the
protocols are going to merge and we're
going to have A2A plus MCB plus AP2 all
in one big
>> one protocol to them all. Yeah.
>> One protocol to roll them all for AI
agents securely doing anything. But
we'll see.
>> Sandy, can I ask you a question about um
so I used to do a lot of work in the
privacy space. Um and I think like one
of the things that we always thought for
privacy is surely everybody will wake up
one day and care about their privacy.
And so we're going to create lots of
things for people to kind of like tweak
like the data that they share and then
every time there's a large data breach
you'd say surely now people will want to
care about their privacy. And so I guess
I have a question for you on just like I
see some of this stuff and maybe I'm
just like very skeptical at this point.
I'm kind of like do agents need this
kind of thing in order to succeed? like
implicitly we're always like oh well you
need trust in order to transact online
but like isn't one counter-argument that
like kind of people don't care and so
you know this will be maybe good at the
margin but it's not really like
necessary for agents to start really
interacting in the economy
>> I think in a if you want to it really
shifts the liability right like it
shifts the liability from the security
aspect which hopefully uh AP2 has
covered right to the ability of the
model itself to make the right choices
like Mihi said. And I think people don't
care as long as it goes right. Right.
Like if if AP2 is already in place,
yeah, people won't really care. They
won't even know it's in place. That's at
the protocol level. Most people don't
even know what's happening, right? They
you're using an agent and interacting
with uh an agent on a platform. you
don't know whether it's an Ato agent or
it's using MCP servers or what's going
on behind the scenes, right?
>> In fact, it's good that you don't know,
you know,
>> and and so I think in a way people won't
care, but the people developing the
tools and the institutions that are
responsible for the liability will care.
Um and and so everyday consumer, no, I
don't think they care. um they like to
be told that it's trustworthy and that
their stuff is secure, but they don't
care how. Uh the people that are liable,
I think they will. And I think they like
Mihi was saying it's such a smart move
for Google because this will force
adoption of ATA as well because in order
to use it it's built on top of ATA and
so
it it will become the de facto that if
you want to use this secure payment
protocol you're also making it an ATA
compatible agent. Um and so I think that
they will gain a lot of market share
there and there was some chatter to your
comment about Anthropic developing a
similar agent agent type protocol. Um
but we haven't seen it yet and so maybe
they will develop it but they started
talking about that like a good 6 months
ago. Maybe they're developed media edit
internally and they haven't released it.
Or maybe they've kind of taken the
OpenAI approach where they were like,
"Yeah, MCP like OpenAI was like, "Yeah,
we'll just adopt MCP. Maybe that will
happen too."
>> Yeah. I think one other item here is the
market is targeting as well because I'm
probably not the demographic or the
market who's going to set up a
personalized shopper on TikTok based on
some ad I've seen scrolling down and go,
"Oh, sure. a $100 every month, you can
go off and do some random shopping for
me. Or I'm not going to delegate my
vacation to some assistant. I'm going to
be there. I'm going to check every
single hotel. I'm going to do the
numbers. I'm going to ed it up. I'm
going to go through the payment. I'm
going to look has the payment gone true.
So, I I I think consumer habits are
changing as well. And there's certainly
a market for it. It's it's emerging.
It's not quite there yet, but maybe in
10 years everybody's going to just go
off and say, you know, hand it over to
the AI to do my shopping. you'll figure
out what I need.
>> For sure. And I think it's like it's not
too far away, you know, like people set
up recurring payments and the whole
point of that is to offload, you know, a
transaction each month. The question is
just like how much broader are you happy
with that going, you know, over time?
>> I I've never set up a single recurring
payment ever. Whenever I see one of
those like I don't trust it.
>> Yeah. Okay. Something that I do think
will like this will propel is the use of
computer use. You know, until now,
computer use has been um kind of like
researchbased, right? Just collect
information in a way that you can't
access other ways. But once we have this
protocol in place and it's able to
actually check things out for you
securely and manage secrets and
credentials, and I think that that will
propel the actual usability of computer
use where right now it's like a cool
thing, but honestly, no one's using it
that much in practice. Yeah, that's
really interesting. It's kind of like
the downstream effects. It maybe turns
out like once you get payments right,
like lots of other things kind of come
out of it.
I'm going to move us on to our next
topic. Um, we're going to do a little
bit of a book review uh for our next
segment. Um, so a book came out that has
been getting a lot of chatter online.
It's dramatically titled If Anyone
Builds It, Everyone Dies. Um, and it's a
book authored by two, uh, long-standing,
you know, people who I think have
participated in the AI policy, AI safety
discussion. Um, Elzar Udicowski and Nate
Sorz. Um, and, uh, I guess Gabe, maybe
I'll kick it over to you first on this.
Um, the core of the book is the idea
that at some point we may build super
intelligent AI and in which case the
minute we build it, it becomes really
dangerous and everybody might die. Hence
the title, if anyone builds it, everyone
dies. uh we can talk about the validity
of that claim but I think I actually
want to start somewhere different which
is I think part of the argument of the
book which I found quite interesting is
just to kind of take the CEOs of all
these AI companies at face value where
we have had kind of a cycle of CEOs
coming out and saying AI is probably the
most powerful most dangerous most risky
most uh most promising technology um and
you know I think almost an argument of
the the book is basically like if you
buy what they're saying, shouldn't they
be a lot more careful than they are
right now? And I'm curious about what
you think about that as as an argument.
>> Shouldn't they be a lot more careful
than they are right now? It's a very
good question. Uh, you know, certainly
some some introspection here as we are
also hopefully helping to build some
technology in this domain. Um,
I think there are a couple of aspects to
this that
the
premise of a an AI that is capable of
acting autonomously
is one that I struggle with. Um, because
somebody has to write a for loop. Now
that somebody might be another AI model,
but somebody had to write the for loop
for that AI model, etc. Um,
AI models fundamentally do nothing
except predict the next token. AI
systems
do things and somebody has to turn them
on and turn them off. And so yes
absolutely at a meta level care needs to
be taken to ensure that the systems we
are building do not have the aggregate
set of capabilities to hit any of these
escape points that we have considered. I
think from a hypothetical
what could go wrong perspective,
there are real risks here, right?
Fundamentally, the issue here is a a
much more holistic one than some
simplistic evolution of a sensient AI.
It is a system of humans, companies,
software, hardware, and yes, a whole
pile of floatingoint numbers coming
together. and the humans are the input
to that system, right? And I think one
of the big worries is obviously that the
models themselves, the system itself
becomes self-propagating.
Um, and that only happens if you give it
the tools and the capabilities to become
self-propagating. And so this
fear about that if somebody did it, it's
possible, it could happen. But the
safety procedures, the use cases that
these systems are being set up for um
can and should be limited to avoid these
scenarios. Um and I think at least what
we're seeing currently, you know, I
highly highly doubt that open AAI is
letting GPT5 write code that goes into
GPT 5.1, at least not cart blanch
without any engineers looking at it. So,
it's really the humans in the loop that
we have to be careful of and as stewards
of this technology, we have to make sure
that we're being diligent about that.
Um, but I don't think we're anywhere
close to escape velocity on doomsday.
>> Gabe, I saw you wrapping up and then
Mihi and Sandy immediately went off
mute. Um, so, uh, Sandy, how about you
go first?
>> Well, I was just I was I had a more of a
question to pose to Kim's opinions. Do
you think we're going to get close
enough where we go oopsie and then like
backtrack and be like, okay, maybe that
was like a a little far, that was a
little scary. Uh, we learned a lesson
from that.
>> Maybe one example and then Gabe would be
interested in your response. You know,
in the autonomous vehicles context, it's
true that like, you know, there was a
company called Cruz, right, that was in
San Francisco operating. they had a
number of accidents and had to pull back
and then you know but you know I guess
luckily there was like the technology
retoled and now Whimo and a number of
other companies are kind of running in
the space and so I guess we like at
least maybe by way of responding to your
question it's like almost like we we do
see that pullback and iteration I guess
the question is whether or not like you
know AI broadly writ I guess is maybe
different in that respect um Gabe do you
want to jump in? Yeah, I mean I think
you know we're still holding the
steering wheel. So absolutely we will
definitely see cases where we have real
world harm done um and have to dial it
back. In fact we're seeing that now. Um
you know there are whole huge
discussions to be had around the
negative role we've had it on this
podcast to some degree of these models
in the context of educating children um
or mental health in terms of folks going
down rabbit holes that they wouldn't
with another human. uh and have, you
know, an infinite self-reassuring uh
hallucination machine in front of them.
So, there are real real risks to these
things. And I do not mean to minimize
those. Um but it's it's particulate. I
it's um here's an area where these
things have risk. Now, let's go address
it by trying to tackle this risk. um the
sort of runaway acceleration um you know
end of the world scenario still feels um
I I don't really see us sort of knocking
on that door short of some like
genuinely bad actors entering and trying
to break things uh and trying their best
to uh you know be a villain in a deep
underground layer and launch the
apocalypse. So
>> in other news I've heard that cloud code
is writing 95% of the code for cloud
code.
>> Yeah. Yeah.
>> So AI is writing.
>> All right. Bye. Okay. You're right.
>> Like I think it is writing itself at
this point.
>> Yeah. On a more serious note, I think
this has been a trend in every single
piece of literature, media, movies, and
games. So I I personally really enjoyed,
for example, some of the games uh like
Mass Effect where you have the Reapers,
which is like a sentient species of AI
where humanity built AI at some point
they realized it was sentient and then
they went, "Whoops,
we're going to stop that." And when the
AI saw that it was being stopped, it
reacted and said, "Oops." And you know,
humans are our enemies. Um, there's also
Dune, like if you remember in Dune, one
one of the premises was thou shalt not
make a machine in the likeness of a
human mind because that's how it the
whole thing started. There was the
Berian jihad and then when you know AI
and humans fought and since then you
were not allowed to build AI. Um there's
showdown in system shock if you've
played like you know DOSS game 199596
whenever that came out uh where again
somebody built a super intelligent AI
and it turned on mankind or matrix so I
think in every popular media games
movies books this has been a trend
whether it's going to happen we don't
really know should we be more careful
well if we're careful then our
competition is not going to be careful
and they're going to get ahead of us So
everybody's pushing to create more and
more innovation. I think at some point
there need to be guardrails in terms of
how these AI systems are allowed to
connect and interact with the real
world. Should you have an AI system, for
example, in charge of medical equipment
or make life like and life and death
decisions. So for example, if you have
an accident and you're in the hospital,
the AI can go in and say, "We're going
to turn off your life support because we
can use your organs." And that decision
is made by an AI. no longer by your
family and the doctor. So, it can take
take some very dark turns if we don't
create rules and regulations in terms of
what systems were allowed to to connect
these AI these AI applications to. But
as long as you're connecting it to your
GitHub repo tools or your, you know, web
search tools, I think we're going to be
fine.
>> Yeah, exactly. It doesn't feel like deep
research presents this threat.
>> Mihi made a made a good point. He
referenced a lot of uh like fictional
fears of humans. Um and I think this is
what this plays into whether it's true
or not. It it is it is clearly a a fear
of humanity is that we will be overcome
or become extinct or taken over. Um and
then if you think about okay well what
is what what is artificial artificial
intelligence trained on? Okay, all of
the data comes from humans, at least
right now, right? Like all of the data
comes from the internet and our own, you
know, fears and our own publications.
And so clearly, like maybe in some ways
we're uh it's chicken and egg. We're
kind of like enabling this aspect of it
not wanting to be shut down because it's
fed on all of our fears.
Yeah, I love that because I was about to
say like I was like, "Well, Miha, you're
just talking about fiction. This book is
supposed to be a piece of non-fiction."
I guess Sandy, you've already kind of
anticipated me because you're kind of
saying, "Well, there's sort of a weird
part to this which is regardless of what
AI is, it's informed by all this data
that has this fiction." And so like it
itself is kind of reenacting this in
real life. And so there's kind of this
weird kind of muddling between like the
fiction of what these AIs do and the
reality of what they will do under a
number of conditions.
>> Cool. Well, we've got, as I promised, a
lot to cover. So, I'm going to move us
on to our our next story of the day. Um,
this was a fun one and also building
again from the discussion that we had
last episode. So last episode we talked
a little bit about a startup called
Alterra Ego um and uh Meta doing a bunch
of demos with wearables, right? They
launched a new Ray-B band that has a
bunch of AI features built into it and
they had some problems on the demo, but
the features are are pretty cool. Um and
I guess not to be left out, um Apple is
dropping its new uh AirPods, which I'm
wearing one of them here today. The cool
thing about the sort of new AirPods
though is that they're going to have
this built-in feature for uh real time
translation. Um, which I think is like
in some ways been like one of the really
cool dreams of AI that I've been waiting
for for many years is the idea that like
you can hear audio and it just
translates into a language you know
autonomously. Um, and so I guess uh Gabe
maybe I'll kick it over to you. you
know, this is this is pretty cool and it
feels very different from a lot of the
other wearable AI features that have
been launched lately. Um, but I'm having
a hard time articulating why this feels
so different. Um, and so I'm wondering
if you could help me kind of like think
a little bit about like why is this
maybe different or maybe it's actually
quite same to like what Meta is
attempting to kind of portray as like
the future of how AI is going to be with
you. I think this is a
solution to a problem instead of a
solution in search of a problem. And I
think that's why this feels different.
You know, we work for an international
company. Um I I think most discourse
that I engage in happens in English, but
uh there are many folks for whom that's
not their first language. Um, and uh, as
a native English speaker, I feel both
privileged and a little bit uh, uh, bad
all the time about forcing everyone to
conform to my language, right? Um, I
think
language translation, whether it's for
personal travel, whether it's for
international commerce, whether it's for
just about anything, is a real real
issue in a global world and people face
it every day. And so, yes, um, lots of
cool possible features out there that
could use AI in real time on some kind
of a device that's attached to my body
that I might like try out the demo and
think, man, this is super cool, and then
put away, but I could see, certainly not
for everyone, not everybody lives in a
multilingual world, but for folks that
do, a a earpiece that you stick in that
can give you that translation with
minimal effort and a smooth UX is a real
game changer. So, it's an actual problem
and an actual solution to the problem. I
think that's why this feels different.
>> Sandy, is this the uh beginning of Apple
making a comeback on the AI side of
things. Um, you know, I think the most
interesting reversal of fortunes we've
been observing for the last 12 months
has been, oh my god, Apple's going to
kill it on AI. Oh man, they seem so far
behind. And also Google, they seem so
behind to, wow, they seem to be really
doing all good all of a sudden. Um, I
guess the question is if we're about to
like go through another inflection point
where by December next year, we're going
to be like, "Oh, Apple's got this." We
should have never have doubted them.
>> You know, I've been getting a lot of ads
for um Google's new phones on my TV. I'm
being targeted for that. So, and and I
was just talking to my husband
yesterday. I was like, I wonder whether
they're going to take over um Apple's
device sales very shortly because
they're so integrated in with the entire
ecosystem, right? That that's Google's
MMO is to integrate everything together.
Um I do we know the model actually being
used for the translation in the AirPods
is has that been released? The article
mentioned Apple Intelligence, which is
their like multi-layered ondevice plus
escalate to a secure cloud plus escalate
to a frontier model at at all costs. So
I I doubt we know specifically, but I
suspect there's an ondevice component
for the latest iPhones.
>> Sure. Like with the latency that makes
sense, but but I I also wonder like how
much they're partnering to make this
happen because you know there's a
history of Apple partnering which I
think is a good thing, right? if if you
don't have it in house at the moment.
You don't want to like just not do it,
not keep up, right? And so they've
partnered with um Google before to make
Chrome the default browser rather than
Safari. Um they've partnered with um and
I they think they have a big partnership
right now with Gemini, right, to bring
it potentially and to catch them up a
bit um since they delayed their Apple
Intelligence release for the next
iPhone. And so, you know, I I want to
say yes. I I am hopeful because I think
more players in the game is a good
thing. Um I don't think we should have
incumbents that capture like 75% market
share when it comes to things like
devices. Um which maybe is what Apple
has kind of done in the past. Um maybe
this is allowing them to share that a
little bit more. this probably not a
favorable if an Apple exec is listening
is probably like I don't agree but but I
think if in terms of a fair market that
is reality is that we should have
competition and maybe this is allowing
some competition to shine through. So
I'm not sure it's the worst thing. Look,
I'm just kind of um going to make a
reference back to a book again,
Hitchhiker's Guide to the Galaxy,
Douglas Adams, 1979.
And I think he predicted a lot of this
um with the Babelish, which was a fish
you put in your ear from the
hitchhiker's guide and it would
translate between languages in real
time. I think he's credited with even
creating the concept of a tablet because
the hitchhiker's guide was, you know,
the size of a small pocket book. It had
all the words information. and you could
put the babel fish in your ear to give
you real time translation. So I think
these are things that humans have wanted
for a very very long time. It came up in
fiction, it came up in movies. Uh we've
seen various implementations over the
years. I remember this feature was
present in Skype which is no longer a
thing right now. But in Skype you could
do real time translation with your I
would say you know if you have a family
and they speak a different language you
can do realtime translation. I think
it's down to the user experience. Making
it part of something like your
headphones is going to be a lot more
conducive to folks who are not used to
technology. They're not going to go and
download the model. They're not going to
go into Skype. Um I think the first
couple of iterations are going to have
some interesting challenges. So I expect
a lot of funny situations in a lot of
countries where you think I can now
speak Italian. It's like no not quite
>> and you can cannot.
Um, so, so I expect there's going to be
a couple of challenges, especially if
you're going to the local model first.
Uh, but I see it as a good thing. So, I
see it as AI becoming integrated into
technology to the point where from a
consumer spe perspective, it disappears.
You don't care what model is being used.
You don't care. You What's a token? I
don't care what the token is, right? I
just want this thing to translate what
you're saying in real time. Done.
>> I I I think your your point about UX is
spot on, Nihi. And I think um it really
it to your earlier question Tim about
what feels different. Um I think that's
what Apple is getting right with this
and to the question also about whether
or not Apple has a comeback in store. I
think that is their avenue is UX that as
you aptly put it Mihi uh has no
reference to the fact that there are
tokens anywhere. It just fits into
something that's already part of your
ecosystem and it makes it better and it
solves a problem. Um, you know, one
other incumbency that Apple has, uh, is
sort of style and popularity. You know,
frankly, Tim, seeing you in those nice
white headphones, um, I I see people
walking down the street with those
headphones in all the time, especially
all the teens in my neighborhood. It's
it's a look. And so now that those solve
an additional problem, that is not
saying I have to now put on a clunky
pair of additional glasses or uh some
heads up display that makes me look like
a geek uh to actually, you know, bring
this technology into my life. It's
fitting into something that's already an
accepted thing in everyday life. And so
the user experience is really spot-on.
So I think that's the avenue um that has
some some real traction here. Yeah, just
look up on the internet United Nations
translation device because you see them
at the United Nations and there's like
this awkward cheap looking thing with
the antennas and I think there's a
person behind it but still it's it's not
a look.
>> Yeah, for sure. Um Sandy, final thought
on this?
>> Yeah, I I I'll wrap this up with a a
quick story. Um earlier this year I was
in Japan doing an awesome solo trip,
getting lost in so many train stations
and having to ask for directions, right?
And um and so I pull out my and I was
using GPT because it just is
semantically better translation um and
voice mode in a way, right? And so I was
doing the translation, but then I would
hand it back to like the older gentleman
and he had no concept how to like
how to participate, right? And so it
wasn't he would speak but not press the
button and then I'd press the button too
late and it would just it was it didn't
work, right? because of the learning gap
and the skills gap and so this I think
is a if they can nail a way to solve
this right that would bridge that gap
where you don't have to change your
learning pattern you just change the
experience
>> yeah this is going to be a super
interesting interface and I want to kind
of keep coming back to it as we go about
like where where the AI will be in this
wearable ecosystem I think it's really
interesting and I think yeah as Sandy
you're pointing out in some ways like
even sort of quote unquote seamless
interface faces have a lot sort of built
into them. Um, so a lot to talk about
there.
Uh, I'm going to end this on a final
story which is breaking as of this week.
Um, it's a headline that could only come
from 2025. And that headline is that
Nvidia is investing $100 billion. I'm
laughing as I say it because it's such
an absurd number in some ways. A hundred
billion dollars in open AI. Um, and so I
want to take the maybe like last five
minutes of this episode to just talk a
little bit about this news. Obviously
the topline number is mind-boggling. Um,
but I guess Mihi, maybe I'll get you to
respond to one thing that occurred to
me, which is Nvidia's about to get or
OpenAI is about to get all this money
from Nvidia. Isn't OpenAI just turning
around and giving it back to Nvidia?
>> It's like stock buybacks. It's perfect.
I mean, it's it's kind of strange,
right? Like like that actually we should
read this as open AI gives Nvidia
hundred billion dollars back, right?
I mean, some of it will be spent on
personnel and that kind of thing, I'm
sure. But yeah, um what do we what do we
make of that? That sounds that seems
weird. It does, but then you look at
things like stock buybacks and
everybody's doing it and then you're
making billions of dollars. If you look
at the Oracle stock, which kind of blew
up as well from a very similar event, um
you're going to see it has a big
financial impact. Then second, they're
sponsoring their biggest client. You
know, if OpenAI goes down, they're
they're going to be in trouble. I think
there was an article going around that
Nvidia has like, you know, three, four
large clients and then the rest. And
then there's a couple of gamers in there
with GPUs, but those are not as
important as those three, four, you
know, whale of customers. And
there's also the opportunity for AMD or
Intel or some other company to come in
and say we have found a solution and
OpenAI is going to diversify. So by
making the investment they're locking in
their biggest potential target. There is
no indication that the OpenAI is going
to reduce the number of GPUs they need.
They're still making their models, you
know, CH GBT freely available for a
billion users or whatever the numbers
are up to right now and they're going to
continue doing it on Nvidia GPUs. So, I
think it's a brilliant move. I think
it's going to
have massive effects on the industry. I
would have wished to see a bit more, I
would say, diversity in terms of what's
supported for inference providers. So
AMD and Intel give us more GPUs, make
them better. I've got an AMD GPU right
there. I love it. But when it comes to
the software,
that's where I think CUDA and what open
what um Nvidia is doing still still has
the lead.
>> Yeah. Sandy, is um is Enthropic in
trouble? They're sort of left out of
this, right? Like
>> there seems to be unlimited money right
now. I I I'm I'm not sure. They they
plan to raise I think a $8 billion round
just recently and raised a $13 billion
round.
>> They'll be okay is what you're saying.
>> Maybe they are in trouble. Maybe this
will be a catalyst for them to to in
order to compete and stay in the ring to
need to raise even more um than
anticipated. But
I I I feel like we're just going to see
more partnerships emerging. We're going
to see more alliances like clubs. You
know, something I did want to comment on
though is the amount of energy that the
facilities that they're investing in are
putting out is enormous, right? Like I
think typically Meta has notoriously
some of the bigger biggest scale
projects in this industry and this like
out competes that tenfold. Um I think
there I think the number was that they
expect 10 gawatts of power at least from
the facility which is like a billion
light bulbs at a time.
>> It's a lot of light bulbs.
>> It's kind of insane.
>> Yeah. Well, and Gabe maybe you can take
us home for the final comment. I mean
this kind of idea of tribes I think is
really interesting because it does sort
of feel like we're getting like it's
going to be Gryffindor and Ravenclaw.
it's going to be, you know, Haronin and
what have you. Like I I like it feels
like initially it was oh okay well
Anthropic is going to kind of pair off
with say Amazon and OpenAI is going to
pair off with Microsoft and that's for
the the sort of cloud right and then now
it feels like okay well the next step in
the game is open AI is going to go with
Nvidia and I don't know maybe anthropic
is going to go with AMD right and then
like I guess the next step in that game
would be Sandy what you're talking about
is I guess like open AI is going to go
with like northeastern nuclear power you
know company and Anthropic is going to
go with you know southwestern you know,
uh, wind power or something like that. I
guess Gabe, like where does this all go
in terms of market structure? Like, do
we feel like we're going to almost have
these like vertically integrated
alliances start to emerge over time?
Because I mean, traditionally, we've
seen like Nvidia selling its chips to
everybody. And I think this does really
seem to put the the thumb on the scale a
little bit, right? I think there's this
is a big bet, frankly, uh, that bigger
is still better, right? I think um we've
hit a point now with some really really
good frontier models trained on the
infrastructure that already exists and
um we're seeing that huge gains in
capability are being achieved not
necessarily just by training the newest
models but by training by by building
the systems around the models. And 2025
the year of the agents and you know
we've we've covered everything under the
sun about and I keep coming back to you
know models are just generating tokens.
It's what you put around them that
actually adds the value. And I think,
you know, there's a real obvious loser
in moving the innovation from the models
to the systems and that's Nvidia, right?
Because you don't need to buy more chips
if you've already got enough, right? You
don't need to boost your data center
size if your models if you can keep
refining the models with the same
hardware you already have. And so I
think this is a huge bet on Nvidia's
part to say, "Hey, we believe that
bigger is still better and we're going
to make bigger happen and see what
happens." Um, so they're putting their
money where their mouth is very
self-interestedly as you put, Tim, and
it's going to come back to them in the
chips anyway. But I think, um,
does this where does this shake the
industry out? It's really going to it's
going to come down to the technology of
whether big is in fact better. Um and
Sandy to your point eventually these
this power consumption the environmental
impact um you know do we can can we
source this power responsibly like we're
going to hit some breaking points there
when you're talking about orders of
magnitude this big and I'd be curious to
see I mean again it's really uncharted
territory. It's a push to explicitly try
to push into further uncharted territory
so that um this critical resource, the
next generation of GPUs, is still a
commodity that is desperately
desperately needed. We'll see where
those uncharted territories lead.
>> Gabe, to your point before we wrap up,
I'm I'm curious whether the hundred
billion will completely go to facility
or they're maybe making that next
investment in the next type of chip or
the next architecture. like is that is
this a play for what's whatever is next?
>> Very possible. But I think you know the
topline power numbers and the topline
scale numbers still really hint at
bigger is better. I think you know
theoretically you could invest $und00
billion in chips that are 10x more
efficient rather than 10x more powerful.
But I don't think that's what they're
doing here. Um, and so it seems like
this is clearly just trying to push the
upper envelope of, you know, raw
compute. Um, but I'm sure some of that
will go into the ecosystem. And I don't
know, I didn't get into the details of
the article about how much is up to the
discretion of OpenAI versus how much of
this is just build the biggest data
center that's ever been built and staff
it up with our GPUs.
>> Can you smell the AGI?
>> Yes, I suppose we all can. Uh, it's a
great note to end on. Um, that's all the
time that we have for today. Uh, so, uh,
Mihi, Sandy, Gabe, thank you for joining
us as always. And, uh, thanks to all you
listeners. Uh, if you enjoyed what you
heard, you can get us on Apple Podcast,
Spotify, and podcast platforms
everywhere. And we'll see you next week
on Mixture of Experts.
[Music]