Extract captions

[dongho204]

URL

https://www.youtube.com/watch?v=Fo-6BWKeerw

[github-actions[bot]]

[youtube] Extracting URL: https://www.youtube.com/watch?v=Fo-6BWKeerw
[youtube] Fo-6BWKeerw: Downloading webpage
[youtube] Fo-6BWKeerw: Downloading ios player API JSON
[youtube] Fo-6BWKeerw: Downloading android player API JSON
[youtube] Fo-6BWKeerw: Downloading m3u8 information
[info] Fo-6BWKeerw: Downloading 1 format(s): 22
[info] There are no subtitles for the requested languages

[youtube] Extracting URL: https://www.youtube.com/watch?v=Fo-6BWKeerw
[youtube] Fo-6BWKeerw: Downloading webpage
[youtube] Fo-6BWKeerw: Downloading ios player API JSON
[youtube] Fo-6BWKeerw: Downloading android player API JSON
[youtube] Fo-6BWKeerw: Downloading m3u8 information
[info] Fo-6BWKeerw: Downloading subtitles: en
[info] Fo-6BWKeerw: Downloading 1 format(s): 22
[info] Writing video subtitles to: 챗GPT가 불러온 변화, 차세대 AI 대세는 AI 풀스택？ [북툰 과학다큐] [Fo-6BWKeerw].en.ttml
[download] Destination: 챗GPT가 불러온 변화, 차세대 AI 대세는 AI 풀스택？ [북툰 과학다큐] [Fo-6BWKeerw].en.ttml

[download]    1.00KiB at  Unknown B/s (00:00:00)
[download]    3.00KiB at  Unknown B/s (00:00:00)
[download]    7.00KiB at    4.66MiB/s (00:00:00)
[download]   15.00KiB at    2.24MiB/s (00:00:00)
[download]   31.00KiB at  989.64KiB/s (00:00:00)
[download]   32.66KiB at  981.95KiB/s (00:00:00)
[download] 100% of   32.66KiB in 00:00:00 at 88.85KiB/s

[github-actions[bot]]

In 1737, French mechanical engineer Jacques
Bosqueson created an amazing masterpiece: a
flute-
playing automaton. This
automaton played the flute like a real person, by precisely moving its fingers and
blowing air through its mouth.

Although there were dolls that made movements, it

was the first time that a person's movements could be elaborated like the Jade Hwang's doll. It
looked as if the
doll was imitating human intelligence.
People called these automatic dolls
automata. Automata are large  It
gained popularity and became a hot topic wherever it went.

[Music]
In modern times, a tomata refers to an
abstract machine that automatically processes certain tasks. In
that sense, artificial intelligence
can be said to be the digital automata of the 21st century.

Digital automata are like chess.  It made an impressive debut for
the public by defeating the champion and the Go team one
after another. As a result, it

was quickly applied to various industries, including self-driving cars. When AlphaGo beat people,
the public
did not really realize the changes brought about by artificial intelligence.
However, at the end of last year.  The first gpt that appeared in
completely changed the atmosphere. The

era has come where not only the industry but also general consumers can experience artificial intelligence.

Artificial intelligence is now

becoming a part of daily life, not a vague technology. The
artificial intelligence craze brought about is strong.
In the past, the iPhone  Just as the full-fledged
mobile Internet world has opened
with the advent of Whip PT, expectations are high that a full-fledged AI
world will arrive with the advent of Whip PT.
Accordingly, global tech
companies are also moving quickly.
Microsoft is
using technology to develop everything from search engines to PowerPoint.
All of its products are armed with AI. As
Microsoft
took the lead, Google also began to jump in.
Meta, Amazon, and Apple
are also introducing AI services. In fact,
almost all global tech companies seem to be jumping into AI.

Domestic companies are responding quickly as well.
All of this

happened only half a year after the collection PT was released.
Yes, today, we will summarize the changes in the AI ​​market that have occurred over the past six months

focusing on four keywords.
Knowing these four keywords

will help you understand the swirling situation.  I think that
even in a whirlpool, updates are necessary anyway. The

most frequently encountered term since the collection PT came out

is probably generative AI. Not only in the
new language field of AR, but also
in the art, music and video field AI, there

is no place where the modifier "generative" is not used.

How is type AI different from existing traditional AI? While traditional
AI aims to better analyze data, the

difference between generative AI is that it derives new results based on the obtained data. For
example, photos of dogs and cats.  In
this case, traditional AI is better at determining which of the two is
stronger and which is
the hometown. On the
other hand, generative AI can

draw pictures of dogs and cats, write novels, or compose music. It is

strong in analysis.  If AlphaGo was close to traditional AI,

Whippet, which communicates skillfully,
is a completely generative AI. To
learn more about generative AI, let's

take a look at the name of Whip
PD. It is a generative AI that communicates in natural language like a whip holder and has

a vast  This is a Transformer pre-trained with data.

Transformer
Of course, I'm not talking about Bumblebee or Optimus Prime.

Transformer is one of the deep learning models.

Transformer looks at a given sentence and
probabilistically predicts what the next word will be. For
example, 1+1 is
the answer.  Rather than obtaining it through calculation like a human, it

selects the word that appears most frequently after 1+1 from the data it has learned. Because of this,
1+1 easily answers the same question, but there is

a lot of data like 29,486 + 5,959.  It may
give wrong answers to unlikely questions. In other words, the
Transformer is not a model that answers life and lies,
but is a model trained to give the most plausible answer.
If there is a Transformer
trained only with data related to Korean entertainment,
1+1 is
the answer.  You might say that.
Anyway, most of the recent generative AI
uses the transformer model. I

said earlier that the transformer is one of the deep learning models. Deep learning

is a learning method that imitates the neural network in the human brain among machine learning. Our
brains are made up of neurons.
Information is transmitted and processed through synapses that connect neurons.

Deep learning also
processes information through parameters corresponding to synapses.
So the more parameters deep learning has, the
more sophisticated learning can be done. The
parameters that are whip pts are
There are as many as 175 billion parameters. When we

pre-trained with 300 billion words and 5 trillion documents for these parameters, we
got results that surprised people.

AI that has countless parameters and is trained with massive data like this is
usually called ultra-advanced AI.
Transformers can ultimately be said to be ultra-large AI.

The performance of ultra-large AI depends
on the scale of parameters and data rather than the excellence of the model itself.

Even for an ordinary model, if it has many parameters,
a lot of data, or both, it can
excel at artificial intelligence.  This means that it can be intelligent.
For this reason, the
race to develop super-large AI is similar to an arms race. The

more military you invest, the
stronger your military power becomes than the other side. The
reason everyone is unable to develop ultra-large AI is not because of
lack of technology, but because they are confident of winning the arms race.
Because there is no such thing, it

costs an astronomical amount of money to develop and operate ultra-advanced AI. It

costs a lot of money to provide massive data and pre-train, but the
real money-maker is in the hardware side. It

is the AI ​​semiconductor optimized for artificial intelligence calculation.  In some ways,
all industries that face change
always have money at the center of that change.
Therefore, if you know about AI semiconductors, which are money-consuming creatures, you

can see changes in the AI ​​industry at a glance.

Computers have a CPU that is in charge of calculations. The

CPU's calculation method.  is a
serial operation that processes data in order.
AI algorithms
can be calculated using serial operations, but the
speed is slow.

Parallel operation is essential, especially for algorithms that need to process data simultaneously, such as deep learning. An
unexpected trick for engineers who were struggling.  It
occurred to me that it uses none other than the
GPU used as a graphics card.

GPU is a semiconductor that is specialized in calculating
numerous numbers simultaneously in order to quickly display images or videos on a monitor.

This simultaneous calculation
ability is perfect for parallel computing in artificial intelligence.  They say

people who are good at drawing are also good at drawing, but
I didn't know this also applies to semiconductors. When the

parallel computing performance of GPU was confirmed at an AI competition held in 2012,

GPU was used in most artificial intelligence learning from then on.
Currently, nbdi artificial intelligence  The intelligent semiconductor market
share is about 80%. The
problem is that GPUs are too expensive.
NVIDIA's GPU called A100

can add and subtract as many as 312 times per second.
Harvest PT uses as many as 10,000 of these A10s.
10,000  The price of the product is
not cheap, but if you

include the huge amount of electricity and coolant costs, engineer technical support costs, and management and operation costs, it

will cost hundreds of billions of won for large-scale AI learning.In addition,
whip bits cost about 26 won per search,
even though one person per month  Just by
searching 10 times, it costs 26 billion won.

From learning costs to operating costs, it is

difficult for most companies to develop ultra-large AI using GPUs. The
reason it costs so much is

because GPUs are not true AI semiconductors.
GPU is bound to be
less efficient in AI calculations.
To solve this problem, we

need to develop AI semiconductors to replace GPU. There

may not be many gamers who made their own GPU because it is expensive and inefficient, but
it is possible for companies.
Google aims to respond to GPU.
In 2016, we developed a semiconductor tpu dedicated to AI learning.

AWS also developed AI semiconductors separately.
Other companies are also

aiming to develop their own AI semiconductors in the long term. The
generational replacement of AI semiconductors is much more
complicated than this, but the
core  It is simple.
Move away from general-purpose and expensive GPUs and

move toward dedicated and efficient AI semiconductors.
Once you have your own AI semiconductor, the rest is relatively
easy. So, the future AI competition is
likely to unfold like this.

Develop your own AI semiconductors and build your own
AI infrastructure.  We build and provide
integrated AI services to customers based on this.
This

is the AI ​​full stack, which has been attracting attention recently. Let's take a look at the AI ​​full stack, which is the last
keyword and the end point of the whirlwind. The

term full stack is mainly used in IT-related
areas.  This
means a developer who knows how to handle

all aspects of software and hardware. However, in
reality, it
is close to impossible to have a full-stack developer who can really handle all of these areas. A realistic meaning would be a
developer who has a high understanding of the overall area.
However,
people  What if it is a system that is not like that?

Wouldn't it be possible to achieve a full stack in the true sense of the word? A

system that builds its own AI infrastructure, including AI semiconductors, and provides customized AI services to businesses and consumers based on this. It is an AI full stack.
AI  Building a full stack
requires enormous investment. Just

like growing the device industry, such as petrochemicals or steel,
long-term and continuous investment must be made.
Nevertheless,
interest in the AI ​​full stack is high. As
competition in AI semiconductor development intensifies, it will
eventually lead to the expansion of overall AI services.  This is

because there is a high possibility that it will expand into a battle of size.
Even in Korea, companies with capabilities are
showing moves to build an AI full stack.
Among them, KT already has the
necessary environment to build an AI full stack.

KT was a reporter at the end of last year.  We held a meeting and

announced plans to complete the Korean

AI Full Stack in 2023. If so, let's take a look at the progress of the Korean AI Full Stack.

Kate made strategic investments in Rebellion, an AI semiconductor
design company, and Moret, an AI infrastructure
solution company, last year.  We have

established an AI infrastructure environment. We are

researching the latest AI algorithms through industry-academic cooperation.
Based on this, we plan to commercialize super-large AI within this year. In
addition, we are already
applying AI technology to
existing business
areas such as artificial intelligence contact centers, robots, logistics AI, etc.  We plan to apply ultra-advanced AI.

KT Cloud, KT's data center,

launched a pay-as-you-go infrastructure service that allows you to borrow GPUs as needed.
On the 28th, KT Cloud launched

next-generation memory technology with the goal of implementing a Korean-style AI full stack with Samsung Electronics.  A
business agreement has been signed for cooperation.
The most meaningful progress is that
the development of AI semiconductors has received a green light.

Rebellion's AR
semiconductor Atombe was first introduced on April 6.
Atom is the first in Korea to develop a Transformer-based
natural language model.  It is an AI semiconductor that supports not only

language models but also vision modelers such as image search.

Atom

was up to 2 times faster than Qualcomm and NVIDIA with language models, and
up to 3 times faster with vision models.
Power consumption is lower than NVIDIA.

If Atom is commercialized, it will

be able to gradually replace NVIDIA's GPU.
If that happens, I think the Korean AI full stack will also
gradually become competitive.
[Music] [
Music]
Thank you for watching
[Music]