00:00:00
so nvidia's CEO Jensen Wang did a
00:00:02
special address at an AI Summit in India
00:00:05
now this talk was rather fascinating
00:00:07
because it was one of those talks that
00:00:08
gives you an Insight with as to where we
00:00:10
headed overall in AI I know most talks
00:00:13
are about AI this or AI that but this
00:00:15
one covered three key topics that I
00:00:17
think most people aren't focused on one
00:00:19
of them was the inference time with the
00:00:21
new paradigm that AI is moving towards
00:00:23
as you know the new 01 model thinks
00:00:26
before it actually talks and that means
00:00:27
that the model gets smarter with its
00:00:29
responses there was also the talk about
00:00:31
agents and how they're going to impact
00:00:33
the workplace and lastly we got a really
00:00:35
cool look at how physical AI is going to
00:00:38
change the world in the future with
00:00:39
humanoid robots so this talk is going to
00:00:41
be a shortened summary and I'm going to
00:00:43
give you guys the key details you need
00:00:44
to know so coming in at the first point
00:00:46
one of the first talks that he actually
00:00:48
gives was about the new inference time
00:00:50
AI this is where he speaks about how
00:00:52
This Ti of AI is quite different it's
00:00:54
basically like how you have system one
00:00:56
and system two thinking system one being
00:00:58
that short Snappy where someone ask you
00:01:00
something immediately you immediately
00:01:01
know the response but system 2 is kind
00:01:03
of deliberative and you know planning
00:01:05
and reasoning through certain steps to
00:01:07
get to your response and this is a SCA
00:01:09
scaling law at a time of inference the
00:01:12
longer you think the higher quality
00:01:14
answer you can produce this is not
00:01:16
illogical this is very very intuitive to
00:01:19
all of us if you were to ask me what's
00:01:21
my favorite uh Indian food I would tell
00:01:23
you chicken briani okay and I don't have
00:01:26
to think about that very much and I
00:01:27
don't have to reason about that I just
00:01:29
know it and there are many things that
00:01:31
you can ask it like for example what's
00:01:33
Nvidia good at Nvidia is good at
00:01:34
building AI supercomputers nvidia's
00:01:37
great at building gpus and those are
00:01:40
things that you know that it's encoded
00:01:42
into your knowledge however there are
00:01:44
many things that requires reasoning you
00:01:47
know for example if I had to travel from
00:01:49
uh Mumbai to California I I want to do
00:01:52
it in the in a way that allows me to
00:01:55
enjoy four other cities along the way
00:01:58
you know today uh got here at 3:00 a.m.
00:02:01
this morning uh I got here through
00:02:04
Denmark I I and right before Denmark I
00:02:06
was in Orlando Florida and before
00:02:09
Orlando Florida I was in California that
00:02:11
was two days ago and I'm still trying to
00:02:13
figure out what day we're in right now
00:02:15
but anyways I'm happy to be here uh if I
00:02:19
were to to tell it I would like to go
00:02:21
from California uh to Mumbai uh I would
00:02:24
like to do it within uh 3 days uh and I
00:02:27
give it all kinds of constraints about
00:02:28
what time I'm willing to leave and able
00:02:30
to leave what hotels I like to stay at
00:02:32
so on so forth uh the people I have to
00:02:35
meet the number of permutations of that
00:02:37
of course uh quite high and so the
00:02:39
planning of that process coming up with
00:02:42
a optimal plan is very very complicated
00:02:45
and so that's where thinking reasoning
00:02:48
planning comes in and the more you
00:02:50
compute the higher quality answer uh you
00:02:53
could provide and so we now have two
00:02:55
fundamental scaling laws that is driving
00:02:57
our technology development first for
00:03:00
training and now for inference next this
00:03:04
is where we actually get to speak about
00:03:06
agents now agents are something that are
00:03:08
right on the horizon 2025 will largely
00:03:11
be the year that autonomous AI takes
00:03:13
over and you'll see them in the
00:03:14
workplace it's likely you'll see them
00:03:16
being able to do a various different
00:03:17
amount of things for you as an
00:03:19
individual it's quite likely that
00:03:21
towards the end of 2025 you're going to
00:03:23
see a number of autonomous AI agent
00:03:26
systems paid and free come online and be
00:03:28
able to offer you a variety of different
00:03:31
goods and
00:03:32
services okay so I'm going to introduce
00:03:34
a couple of other ideas and so earlier I
00:03:36
told you that we have Blackwell we have
00:03:40
all of the libraries acceleration
00:03:41
libraries that we were talking about
00:03:43
before but on top there are two very
00:03:46
important platforms we working on one of
00:03:47
them is called Nvidia AI Enterprise and
00:03:49
the other is called Nvidia Omniverse and
00:03:52
I'll explain each one of them very Qui
00:03:54
quickly first Nvidia AI Enterprise this
00:03:57
is a time now where the large language
00:04:01
models and the fundamental AI
00:04:02
capabilities have reached a level of
00:04:03
capabilities we're able to now create
00:04:06
what is called agents large language
00:04:08
models that understand understand the
00:04:11
data that of course is being presented
00:04:13
it could be it could be streaming data
00:04:15
could video data language model data it
00:04:17
could be data of all kinds the first
00:04:19
stage is perception the second is
00:04:22
reasoning about given its
00:04:25
observations uh what is the mission and
00:04:28
what is the task it has to perform
00:04:29
perform in order to perform that task
00:04:32
the agent would break down that task
00:04:35
into steps of other tasks and uh it
00:04:38
would reason about what it would take
00:04:39
and it would connect with other AI
00:04:42
models some of them are uh good at prod
00:04:45
for example understanding PDF maybe it's
00:04:48
a model that understands how to generate
00:04:50
images maybe it's a model that uh uh is
00:04:53
able to retrieve information AI
00:04:56
information AI semantic data from a uh
00:05:00
proprietary database so each one of
00:05:02
these uh large language models are
00:05:05
connected to the central reasoning large
00:05:07
language model we call agent and so
00:05:09
these
00:05:09
agents are able to perform all kinds of
00:05:12
tasks uh some of them are maybe uh
00:05:15
marketing agents some of them are
00:05:16
customer service agents some of them are
00:05:17
chip design agents Nvidia has Chip
00:05:19
design agents all over our company
00:05:21
helping us design chips maybe there're
00:05:23
software engineering uh agents uh maybe
00:05:26
uh uh maybe they're able to do marketing
00:05:28
campaigns uh Supply Chain management and
00:05:31
so we're going to have agents that are
00:05:33
helping our employees become super
00:05:36
employees these agents or agentic AI
00:05:39
models uh augment all of our employees
00:05:42
to supercharge them make them more
00:05:44
productive now when you think about
00:05:46
these
00:05:48
agents it's really the way you would
00:05:50
bring these agents into your company is
00:05:52
not unlike the way you would onboard uh
00:05:55
someone uh who's a new employee you have
00:05:58
to give them train training curriculum
00:06:01
you have to uh fine-tune them teach them
00:06:03
how to use uh how to perform the skills
00:06:06
and the understand the vocabulary of
00:06:08
your of your company uh you evaluate
00:06:11
them and so they're evaluation systems
00:06:13
and you might guardrail them if you're
00:06:15
accounting agent uh don't do marketing
00:06:18
if you're a marketing agent you know
00:06:19
don't report earnings at the end of the
00:06:21
quarter so on so forth and so each one
00:06:24
of these agents are guard railed um that
00:06:27
entire process we put into to
00:06:30
essentially an agent life cycle Suite of
00:06:34
libraries and we call that Nemo our
00:06:37
partners are working with us integrate
00:06:38
these libraries into their platforms so
00:06:42
that they could enable agents to be
00:06:44
created
00:06:45
onboarded deployed improved into a life
00:06:49
cycle of agents and so this is what we
00:06:52
call Nvidia Nemo we have um on the one
00:06:56
hand the libraries on the other hand
00:06:58
what comes out of the output of it is a
00:07:02
API inference microservice we call Nims
00:07:06
essentially this is a factory that
00:07:09
builds
00:07:10
AIS and Nemo is a suite of libraries
00:07:14
that on board and help you operate the
00:07:16
AIS and ultimately your goal is to
00:07:19
create a whole bunch of agents this is
00:07:20
where we got the very fascinating talk
00:07:23
about how we're going to get physical AI
00:07:25
of course once you do have agents in AI
00:07:28
it's very good because they are digital
00:07:29
and they're able to move at hypers speed
00:07:31
but how do you impact the physical world
00:07:34
how do you manipulate physical objects
00:07:35
and Achieve things in the real physical
00:07:37
world whilst maintaining that scale of
00:07:40
course it's humanoid robot and physical
00:07:42
AI this is where he gives the
00:07:44
interesting insight into where physical
00:07:46
AI is truly headed what happens after
00:07:49
agents now remember every single company
00:07:53
has
00:07:54
employees but most companies the goal is
00:07:57
to build something to produce something
00:07:59
something to make something and that
00:08:02
those things that people make could be
00:08:04
factories it could be warehouses it
00:08:06
could be cars and planes and trains and
00:08:09
uh ships and so on so forth all kinds of
00:08:12
things computers and servers the servers
00:08:15
that Nvidia builds it could be
00:08:17
phones most companies in the largest of
00:08:20
Industries ultimately produces something
00:08:23
sometimes produ production of service
00:08:25
which is the IT industry but many of
00:08:28
your customers are about producing
00:08:29
something those that next generation of
00:08:33
AI needs to understand the physical
00:08:35
world we call it physical AI in order to
00:08:39
create physical AI we need three
00:08:41
computers and we created three computers
00:08:43
to do so the dgx computer which
00:08:47
Blackwell for example is is a reference
00:08:49
design an architecture for to create
00:08:51
things like dgx computers for training
00:08:53
the model that model needs a place to be
00:08:58
refined it needs needs a place to learn
00:09:00
and needs the place to apply its
00:09:03
physical capability its robotics
00:09:05
capability we call that Omniverse a
00:09:08
virtual world that obeys the laws of
00:09:11
physics where robots can learn to be
00:09:14
robots and then when you're done with
00:09:17
the training of it that AI model could
00:09:20
then run in the actual robotic system
00:09:23
that robotic system could be a car it
00:09:24
could be a robot it could be AV it could
00:09:27
be a autonomous moving robotic could be
00:09:29
a a picking arm uh it could be an entire
00:09:33
Factory or an entire Warehouse that's
00:09:35
robotic and that computer we call agx
00:09:38
Jetson agx dgx for training and then
00:09:42
Omniverse for doing the digital twin now
00:09:45
here here in India we've got a really
00:09:47
great ecosystem who is working with us
00:09:50
to take this infrastructure take this
00:09:53
ecosystem of capabilities to help the
00:09:56
world build physical AI systems then we
00:09:59
got a very short summary about the
00:10:01
entire talk this is from software 1 to
00:10:03
software 2.0 about her AI agents and
00:10:05
mainly about how the humanoid robot is
00:10:08
going to go completely crazy Nvidia are
00:10:10
actually doing so much in that area that
00:10:12
I can't wait to show you guys a new
00:10:13
video I've been working on that covers
00:10:15
how the Nvidia company is about to
00:10:16
change the entire AI ecosystem so take a
00:10:19
look at this because this one gives you
00:10:21
pretty much everything you need to know
00:10:23
for 60 years software 1.0 code written
00:10:27
by programmers ran on general purpose
00:10:29
purp
00:10:30
CPUs then software 2.0 arrived machine
00:10:34
learning neural networks running on
00:10:36
gpus this led to the Big Bang of
00:10:39
generative AI models that learn and
00:10:41
generate
00:10:42
anything today generative AI is
00:10:45
revolutionizing $ 100 trillion in
00:10:48
Industries knowledge Enterprises use
00:10:51
agentic AI to automate digital work
00:10:53
hello I'm James a digital human
00:10:56
Industrial Enterprises use physical AI
00:10:58
to autom physical work physical AI
00:11:02
embodies robots like self-driving cars
00:11:04
that safely navigate the real world
00:11:07
manipulators that perform complex
00:11:09
industrial tasks and humanoid robots who
00:11:13
work collaboratively alongside us plants
00:11:16
and factories will be embodied by
00:11:18
physical AI capable of monitoring and
00:11:21
adjusting its operations or speaking to
00:11:23
us Nvidia builds three computers to
00:11:26
enable developers to create physical AI
00:11:29
the models are first trained on djx then
00:11:32
the AI is fine-tuned and tested using
00:11:35
reinforcement learning physics feedback
00:11:37
in Omniverse and the trained AI runs on
00:11:40
Nvidia Jetson agx robotics
00:11:43
computers Nvidia Omniverse is a
00:11:46
physics-based operating system for
00:11:48
physical AI simulation robots learn and
00:11:51
fine-tune their skills in Isaac lab a
00:11:54
robot gym built on Omniverse this is
00:11:57
just one robot future Factor fact will
00:11:59
orchestrate teams of robots and monitor
00:12:02
entire operations through thousands of
00:12:05
sensors for factory digital twins they
00:12:08
use an Omniverse blueprint called Mega
00:12:11
with mega the factory digital twin is
00:12:13
populated with virtual robots and their
00:12:15
AI models the robots brains the robots
00:12:19
execute a task by perceiving their
00:12:21
environment reasoning planning their
00:12:24
next motion and finally converting it to
00:12:27
actions these actions are simulated in
00:12:29
the environment by the world simulator
00:12:31
in Omniverse and the results are
00:12:33
perceived by the robot brains through
00:12:35
Omniverse sensor
00:12:36
simulation based on the sensor
00:12:38
simulations the robot brains decide the
00:12:41
next action and the loop continues while
00:12:44
Mega precisely tracks the state and
00:12:46
position of everything in the factory
00:12:48
digital twin this software in the loop
00:12:51
testing brings software Define processes
00:12:54
to physical spaces and embodiments
00:12:56
letting Industrial Enterprises simulate
00:12:58
and validate changes in an Omniverse
00:13:01
digital twin before deploying to the
00:13:03
physical world saving massive risk and
00:13:07
cost the era of physical AI is here
00:13:11
transforming the world's heavy
00:13:12
Industries and Robotics
