Looking at?
Or should I call you Jane? I mean, maybe I can do that this week.
A lot of big earnings coming through. What are you most curious to see from
these these big tech giants? David, Yvonne, great to have you back.
Thank you very much for having me. Right now, both in the U.S.
and China ecosystem, the air space is racing at a breathtaking speed in terms
of the U.S., of course, in the infrastructure side, in terms of
semiconductor and also a large model as the software infrastructure.
China is still playing catch up a little bit.
And but in terms of application, China's speed is second to none.
So I don't want to steal from the Thunder, but I think today probably the
race change a little bit in terms of performance.
A lot of those earnings impact actually going to come down to including Nvidia.
I know David's one of your favorite topics is going to be impacted by how
efficient those models are going to be, how well performed they're going to be
and how much data processing they require.
So I think not only just in the past few months, we
have witnessed all the big tech both in U.S.
and China ecosystem issued a lot of their own models.
But also we'll start to see this competition of efficiency starts as
well. Kai Fu Yeah, as Jen just mentioned,
maybe the game has changed a little bit with now today your new model.
Tell us a little bit about Wang Jia. Is this really China's answer to GPT and
why do you think some of the other previous sort of ones that we've seen
chat bots in China have not been good enough?
Right. So we are launching today our new model
called E Large, which is the model that has API being released throughout the
world. We intend to be a global company, so not
just for China, but for China. We've developed an app on top of E
large, so e largest to Egypt for as we mentioned, is to charge a T.
We now have both 01.
II. My company was founded only one year ago
and when we started we were at least seven or eight years behind that to open
the AI. But now I'm proud to say that our latest
model E Large is performing comparably with GPG for by third party evaluations
and on top of a powerful model is really the only way to support a product with
product market fit that will excite its users.
Bring about the chatbot moment. So we are hopeful that our model with
this performance will provide a chat bot that will rival chat to be t for China.
But we also think the model will support all kinds of apps throughout the world.
It looks like we are bracing for just explosive growth these next at least
this next year or so, hopefully. And Jen, I want to get your thoughts on
this. But, Kyra, let me just get your thoughts
on this first. You mentioned it carefully, you know,
barely 12 months old and you're in the market.
This applies to other players in this market, to the fact that
the fact that barriers to entry are low, which perhaps is a good thing.
How do you expect competition to change these next 12 months or so?
Should we expect more competition? And how and how are you looking to
innovate your product ahead of perhaps competition coming in Q4?
Right. I actually don't think barriers are low.
You don't see these great models and activity like products and everywhere in
the world. There are maybe half a dozen companies
that are in a similar situation in U.S., China, perhaps France and other
countries. So we are one of the few, and certainly
the Chinese companies have achieved this level of performance at breakneck speed.
And that is the essence of Chinese entrepreneurship, is first, we don't
have as many GPUs. So we need to figure out how to use them
efficiently. So, for example, at 01.
II, we train our models at one half or a third the cost of comparable companies
because we don't have access to the latest GPUs.
And also we have many, many, many fewer of them compared to open the AI.
And despite these challenges, we've reached for performance.
So we're quite proud of that. I think many of our Chinese peers have
also done quite well. So I think a lot of are saying in in
media that China is way behind is not accurate.
Now, that said, companies like Mather, Google, Microsoft are putting 5100 times
more resources in there. So we certainly don't take it lightly.
So as we catch up with currently best model, we realize that even better
models will come from Openai and others and we want to stay as close as we can.
But we also think is about building a great user experience.
It's really understanding what the users want and using that which is China's
advantage. As Jen said earlier, if you think about
Chinese applications, Tik Tok is better than Instagram.
Products like Tamal and Sheean have taken over the world.
Users love it. I think China's ability to develop great
applications that focus on what users want and that product market fit is a
unique attribute to Chinese companies. Carrefour you mentioned about GPUs.
I was I was going to ask you, it seems like you guys have been basically
loading up on these video chips. You know, you foresaw that this was
going to come. I think we spoke to you back in November
saying you may have maybe a year or a year and a half worth of supply left.
What do you think was going to happen after that?
You know, do you think that, you know, basically gender of AI is going to be
for those the deep pockets you have to keep, you know, shelling out dollars for
this? Or do you think that there's going to be
these smaller sort of a newer business models that can actually evolve to give
smaller startups here more access to these compute resources?
Well, I think there are two questions here.
One is accessibility to GPUs. We're in fine shape right now.
We I mean, we've raised a lot of money, but it's a tiny percentage of what the
top American companies have. So we're really more bounded, not so
much by what we can buy, but how much money we have.
So we have to be extremely parsimonious and practical and really build that the
right models with the right size that fit the user's needs.
Now, one could certainly train a, you know, with with the resources that a
Google and Microsoft and matter have, you know, train a 50 trillion parameter
model. But such a model will not run for
everyday applications. Everyday applications require minimal
latency, very good accuracy, and different
applications require models of different sizes.
So today we're also releasing a whole series of models from 60 to 90 to 30, 40
to E large, which is over 100 B, And for every application we've demonstrated
that we have a model size that performs as well as, if not better than the
competition. So I think that demonstrates for real
practical use. We, as you and I, are able to compete on
a global basis. Now there could be a huge model that one
of the top American companies developed that performs better.
But it's still remains to be seen whether such a huge model with
corresponding. Fired.
A large amount of compute resource can be deployed in real applications that
delivers return on investment and has a reasonable infrastructure and latency
for the user. Gen Your reaction to a K fully opened up
many doors there. Just Yeah.
Yeah. First of all, I'd like to congratulate
the co-founder and his team for making such meaningful effort to grow plays to
China's soft software side of infrastructure.
But also I want to just departure briefly depart from this US-China
framework for a little bit. What we are talking about is the
community where is very GPU rich and the community of the GPU poor, either
because you are open source community or you have because of geopolitical issues,
the start from GPU, the most advanced GPU supplies.
The world actually vast majority of people in the tech industry, not only
just in the US-China framework, in the global
community, vast majority of players don't have a lot of GPUs like, you know,
open openai eye, etc.. So the I see this actually an
opportunity in terms of where I should go, because if you think about the large
model right now, because most of the large models is still kind of black box,
we don't really know which set of data really made a difference in terms of the
algorithm to understand and make sense. Therefore, one of the easier lazier way
to approach this, to add more and more data, use more and more GPU.
But if you think of our human brain, we are
talking about the biggest model in the world is opening.
I charge about for its rumored more than 1 trillion parameters.
Yeah, but our human brain consumes 30 walk and we're process processing 100
trillion parameters. So there's no comparison.
I have always said this You don't get to Mars by building tall and taller,
building on earth. So by building bigger and bigger model,
consuming more and more GPU is not really the way to get truly flexible.
A.I. So I actually think what confused team,
a lot of companies and developers, their pursuit
to more efficient and smaller model is the future.
Microsoft just issued a Phase three as well using, you know, 7 billion
parameter. It's actually outperformed much larger
models over my open Openai I yeah, you mentioned this emergence of smaller
language model and what what implications is that going to have for
these GPU companies? What is their share price and this whole
race right now, I wrote my link to you a while ago.
I when the video's price is highest, I actually did not believe the price will
stay that high because precisely for the reason I mentioned.
But also if you take a look at the what's happening in the global scene.
Right. I think people realizing, you know, in
social media age where you share, you saw you are sharing your information,
pictures to your friends, family, you're actually sharing with Mark Zuckerberg.
But now people are realizing your most intimate questions you're asking.
Judge he charged you be you GBP is actually not private conversation.
You chat with a private owned commercial company opening I run by Sam at MIT.
That's kind of creepy I think. Yeah.
So so so I think what's going to happen because of this combination of
geopolitical situation, GPU distribution, I think this kind of
smaller model, localized, the A.I. local agent is going to happen.
A couple of trends need to notice here, as we have already well address them
that in China, GPU is becoming a resource that's very difficult to
access. Yeah.
However, if you take a look at China's smartphone smartphone life, for example,
OnePlus 12 is selling 24 to 4G and one terabyte phone and 900
U.S. dollars.
And China's EV market is developing so fast, right?
It took China 24 and 27 years to produce the first 10 million EVs.
But you only took 17 months to produce past the 20 million mark.
So all these EVs, we are tablets on wheels.
So what's happened is that all the very, very powerful individual applications,
they will have the capability to hold their own data locally, process locally.
Of course there's a trade off in terms of efficiency and performance, but
eventually I think those kind of eternity of, I suppose very centralized
approach that I open, the AI will start to flourish or will become much more
popular outside of the US system. Who can I bring you in on that?
Just your thoughts on the the trends around efficiency that you think we
should be paying attention to not just this year but in the years to come?
Yeah, certainly. I think the we're currently very fixated
on the training market, which is how to take a whole world's data and train
giant models. And for that you do need the giant
cluster. And currently India is the best
solution. But as Jen said, inference is a very
different story when you actually deploy these models.
Bigger is not necessarily better. You want to be able to look at the app
that you're building. If you're building something with deep
reasoning skills, that generates great content.
You do need a large model. But for a chat bot, customer service and
games, you do not. So there needs to be a model for every
occasion. And that's why 01.8 we've deployed a
models of different sizes and the APIs ISE are being made available globally at
very competitive prices. I think what's important, an indicator
important to watch is really the cost of inference.
I think that is the gating factor, assuming people always find good enough
technology for each application use. It's all about writing down the cost of
inference. We see inference costs dropping about
ten times a year. So that means things that seem
impractical today will be widespread in one year at most in two years.
So I think a proliferation of GPUs models and agents and applications that
run on all different sizes I think will be the key to the next phase of
development of our labs. Genuine to safety come in.
When should safety come in? Conversations around, say, yesterday.
I think that the problem with safety right now, it's actually not so much a
lack of discussion, but it's a lack of understanding where the discussion
there's a lot of distraction in terms of inflating what I can do and therefore
actually describe what need to happen. My personal belief, I'm a huge proponent
of open source and keep localize and have distributed approach instead of
centralized approach, which is that all the large tech is doing right now.
And the end of the day when it comes to safety, everybody have very different
and very nuanced approach. Right.
We we we all see our privacy in the very different way.
So leaving to a handful, few companies or even for regulators to decide every
single aspect to what individuals should use or access and express and interact
with the AI is limited. So therefore, I think the more we have
the open source approach, localised approach, a distributed approach
combined with regulation and combined with this kind of leadership with a with
value and the concern of the society from the large companies is probably the
recipe, the best recipe you can have right now.
Yeah. And coffee, maybe to give you the last
word here, there's the potential for A.I.
to create havoc on the world. I mean, what's the counterargument to
that? For those concerned, the AI may be an
existential threat? Well, one concern is that I will take
over the world like Terminator. Terminator.
I would say that's completely unjustified.
A.I. is a tool that we use.
Now, I do agree that AI's very, very powerful and can be used by the bad guys
that will bring about very serious harm, for example, for terrorist groups and
things like that. So it's important to put all the
check checks and balances in place and safety measures.
I would point out that every great technology always came with big risk.
Elected electricity brought about risk of electrocution.
People invented circuit breakers on Internet, on PC, brought about viruses.
People invented antivirus software. So I would say now is the time to call
to do the for all the people working on their limbs, to not just focus on bigger
a better model, but having a renewed focus on safety.
And I also believe open source is important and we are open source.
It's our biggest model today as well.