# Jensen Huang: Why Open Agent Systems and AI Specialization are Key
**Source:** https://youtu.be/Yy3JH6dDugc?si=v5fN1-smYYMgupRJ

## Summary

NVIDIA CEO Jensen Huang, in conversation with the likely LangChain founder, discusses the recent breakthroughs making AI truly useful, emphasizing the critical role of open agent systems and frameworks like LangChain. He highlights NVIDIA's commitment to fostering an open ecosystem that enables the creation of specialized, domain-specific AIs through a combination of powerful base models and adaptive harnesses.

- The last six months have been pivotal, making AI finally 'useful' and accessible for enterprises, a culmination of 15 years of AI work.
- LangChain is deemed critical for transforming raw Large Language Models into useful products and agentic systems, providing an essential 'harness' for RAGs, tools, and memory.
- Jensen Huang emphasizes NVIDIA's dedication to building open AI systems to enable widespread application and the creation of specialized, domain-specific AIs.
- Agentic systems are highlighted as the 'big breakthrough,' characterized by grounding on knowledge, tool usage, memory management, safeguards, and iterative problem-solving capabilities.
- AI specialization requires both capable base models, such as NVIDIA's Nemotron Ultra, and robust harnessing frameworks like LangChain to embed domain-specific knowledge and enable post-training.
- NVIDIA's Nemotron 3 Ultra, when integrated with Deep Agents and a tweaked LangChain harness, achieved 86% on an internal benchmark, closely trailing Claude Opus's 87%.
- The 'flashpoint' for advanced agentic systems capabilities was realized with models like Claude Code, which sparked imagination in this domain.
- The envisioned future involves an 'AI flywheel' where systems continuously learn and improve through use, becoming smarter and more useful over time, akin to human learning.

## Transcript

[[0:11]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=11s) Excited to be here with Jensen. There has been a ton of advancements
in AI and agents over the past year, but few months in particular, I feel. We've seen a lot of these advancements
come in the form of better performance, but at the same time, we've also seen
that openness and control and trust in a lot of these models and systems around
them has become more and more important. And so the first thing I want to start
with is how and why are you guys at NVIDIA investing in an open agent ecosystem- and stack? first, before I answer,  I
wanna congratulate you for all the work that you do. In fact, if you look at the last
six months, we could both agree that,  although we've been working
in AI for 15 years, the last six months changed everything,  and so,
all of the technology, all the large language model advances, all the
scaling, all of the breakthroughs, all the omni models,  multimodality
stuff, all that stuff is fantastic. But in the end, it was the last six
months that where everything came together, and now finally AI is useful.

[[1:15]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=75s) And when AI is useful, every company
in the world, every enterprise in the world wants to get their hands on it. And now the question is how? And this is where LangChain comes in. And you always had a vision that the
large language model was the essential ingredient, the essential enabling
technology, but in order to turn it into a useful product, you have to surround
it with what is now known as a harness. There's so much more. So much more. And, back in the old days,  we used
LangChain to help us turn a large language model into a promptable API. And we would turn we used LangChain to,
build our RAGs, and we used LangChain to, step by step which led to today's agents. And really what happened in the last
six months, the big breakthrough are these agentic systems that are grounded
on info, grounded on knowledge, that can use tools to do search and has
memory that it manages and, it has safeguards and,  has the ability to
iterate until it gets the job done. But it ultimately needed some models
that have reached a level of capability where everything comes together
into that flashpoint, and that's where Claude Code, really brought
the, imagination of agentic systems.

[[2:28]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=148s) OpenClaw, of course, was a big deal and,
all the work that you did, with, what, Deep Agents and, we use that ourselves
and, all of that kind of came together and bam, here we are with agentic systems. The reason why we do it, we've
dedicated ourselves, for many, years to build open systems. And the reason for that is
because ultimately, AI is a fundamental technology. It can only be useful if applied in a
whole  whole bunch of different use cases. Now, of course, it's the first use case is
just language and cognitive intelligence, and that's very important, of course. We imagine a world where, scientists
and digital biologists and designers and, roboticists and, students and
researchers, enterprise IT, all of us could use agentic systems, AIs,
to solve domain-specific problems. And many of the problems that we wanna
solve, either we have specialized domain knowledge that is just simply not
available outside that we have to embed into, imbue into, our, AI or it's because,
we believe that AI becomes ultimately great, become a super agent when we put
it into a flywheel where, we use it, it gets smarter, it becomes more useful.

[[3:49]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=229s) We use it even more, it gets even smarter. Kinda like us, kinda like
humans, learns over time. Learns over time. And so, we imagine this future
where AI has a foundation and, the work that Anthropic and OpenAI and,
Google's doing is all fantastic. But there's specialized AIs and
domain-specific AIs and proprietary AIs that people wanna build, and we
wanna enable that world to happen. Maybe digging into that for a second
on this topic of specialization. How exactly do you think it's
best to specialize these systems? Is it gonna be purely the model? Is it gonna be the harness
as well, the context outside? What goes into the specialization? The specialization starts with you need to
have intelligence that's good enough, and this is what why we worked on Nemotron and
really love the fact that you're part of the founding team of Nemotron Coalition. We made Nemotron Ultra pretty incredible. Now, Nemotron Ultra is, a great model
as a start, but it becomes an incredible model when you put the LangChain framework
around it, the LangChain harness around it, so that you ground it on information
that is domain-specific, and, a person is, an intelligent person becomes
super useful when we give them access to particularly important information.

[[5:07]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=307s) And so access to information's important. Putting it into a flywheel where,
maybe you're even training the model, post-training the model inside the
LangChain harness against a harness so that, the model becomes good
at applying the harness around it. What you want it to do for that task. What you want it to do. Yeah. And so I think that this moment
has now arrived, but we need a open harnessing system that we can build
ourselves, that we can, apply and then, of course, improve against over time. I love what you said about
the model being good enough. I feel like that threshold- Yeah … was
crossed, I don't know, maybe a year ago by some of the frontier models, six months
ago by some of the open weight models. Yeah. You talked about Nemotron 3 Ultra. We've done a lot of work with that to
make that really good in Deep Agents. some of the things we did is tweak
the harness to make it best for this model, 'cause we found that different
models need different prompts- Yeah different tools.

[[6:07]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=367s) And with that tweaking, we managed
to get Nemotron 3 Ultra in Deep Agents to we have, an internal
benchmark, and it managed to get up to, eighty-six percent on that. Ooh. Claude Opus on, for
comparison's at eighty-seven. you've got DeepSeek and, one of
the Minimax models at eighty-two, eighty-three down there. So we're starting to see that some of
the more recent open weight models are really getting to frontier performance. I know. I am so proud. But it is so incredible. Thank you. But the, but the It is so incredible
… one of the just as important thing is it's, 10 times as cheap as Opus. Yeah. And I think open weight models really
strike are starting to really strike a good balance between performance and cost. So I'd be curious how you see this cost
part changing the equation for builders. The benefit of cost is it comes
in a couple of, different ways. I happen to think that, when you
have cost-effective intelligence, people just use more of it. When you have a cost-effective
agent, then you can iterate across a larger search space. And as a result, the answer
could actually be better. And in the case of Nemotron, it's
cost-effective because it's so fast.

[[7:10]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=430s) It's so computationally efficient. When it's computationally efficient,
it could explore larger spaces. And it's no different than
when somebody can think fast, you could explore more space. When you can try things more
quickly, you could become… you can find a better answer. And so this is the incredible
benefit of Nemotron 3 Ultra inside the LangChain framework and the
LangChain harness inside Deep Agents. It could think so quickly, it could
explore so quickly, it could iterate so quickly and effic and efficiently
that it's gonna find better answers. And so I'm, just really,
excited that we created a model that was near the frontier. But adapting the environment
around Nemotron, you made it deliver frontier capabilities. Now insight, for humans, it's the same. Of course, we like to hire the
smartest people in the world. but beyond that, we also give them
access to tools, we give them access to information, and we also create the world
around them so that we, we allow them, enable them to create the conditions for
them to achieve their full potential.

[[8:14]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=494s) And so you adjust the
environment, not just the model. And this is where LangChain came in. What you said about using more of
the intelligence as it's cheaper and faster, we, see that to be so true. I think one of the things that  I
like to think I'm AI forward. One of the things that I've underestimated
is just the demand for intelligence and for tokens and how big and massive
that market is, and I think that's become especially true recently. With these models getting good and being
really fast and really cheap, how should we think about using frontier models? Should we just use these open
source models all the time? is there a time and place for both? the frontier models are getting better
all the time, and I fully expect the frontier models to be unbelievably good. And they still have a long,
runway of improving the models. The scaling, scaling laws,
of course, are gonna sustain. Their harnesses are
improving all the time. their technology for dealing with memory,
whether it's working memory or long-term memory, is advancing incredibly quickly. The compaction technologies, the, all of,
the advancements in, retrieval, augmented generation and, knowledge graphs,
and there's still a lot of incredible advances that are being implemented
into these frontier, model APIs.

[[9:30]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=570s) The thing that… The way I think about
it is, frankly, I always start all of my work starting with the frontier. Okay. and the reason for that
is because, it's useful. I know what's the potential. It costs a little bit more money, but
it's incredibly… the, my, my time to getting the work done is fast. However you know, over time, I find
that I want to add sub-agents to them. I want to connect sub-agents that
are super agents at certain skills. And so we have, we have optimization
problems inside our company that, relate to supply chain. Maybe it's related to, chip design
optimization, floor planning optimization. And these problems, these optimization
problems are insanely hard. And so you're not going to just
have an, a general AI go off and crunch on it and think that you're
going to find a great answer. So we create super sub-agents, and
these super sub-agents we would create with Deep Agents, LangChain Deep Agents
with Nemotron 3 inside, and we'll even connect them to specialized tools.

[[10:31]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=631s) And that thing is built for one job. That super agent is not trying
to, book me travel, appointments. It's just trying to
optimize our supply chain. And in that case, I really
do need to have LangChain. I really do need to have Nemotron 3 Ultra,
and I connect it to a lot of proprietary knowledge and proprietary skills. I've got a whole team who's
just dedicated to refining that. Now, I think that defines a company. A company is really about a collection
of a whole bunch of these super proprietary, super important workflows. And now we can have LangChain with
Deep Agent and Ultra, Nemotron 3 inside, and it gives them all
of the control they need, super efficient access to incredible tools. That's the future. Do you have any advice for enterprises
if they're following your practice of starting with the frontier
and then starting to specialize? When should they think about specializing? what are some triggers that you
look for that- As soon as it gets good enough, So I would take I would
start with Cloud Code and Codex and, use it for as long as I can.

[[11:36]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=696s) And frankly, for a lot of things you
never have to replace, because it's get- they're getting better all the
time, and they're on a, they're on a trajectory that's gonna, that's gonna
bring capabilities insanely fast. And so, I think that, in the future,
just like companies are today, we have employees that we hire because their
domain specialization and the refinement of the work and the work process and
all of their life learnings here in the company is just too valuable. but we also hire consultants,
and we license external tools, and we outsource work to other
people, and so on and so forth. I think this is the future
for AI, and are we going to continue to use, frontier models? Absolutely, and tons of it. But are we also going to create
specialized, super agents with LangChain and Nemotron 3 Ultra that in fact
arguably could be your crown jewels? And the answer is absolutely true. I think even for the consultants that
you bring in just like when you bring in a consultant, you need to get them
up to speed in your organization and give them context on the organization,
how things work, what tools do they need that have access to data that's
only inside your organization.

[[12:46]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=766s) And so I think, one of the things we've
seen is as enterprises start to adopt AI, there's all of these kind of like
systems that they have to, build around them in order to make the, agentic
systems as a whole trustworthy and safe and proper kind of like governance. I'm, curious, how do you see and just
to add on that, just, today, most companies are built on business processes. Yeah. In the future most companies
will be built on harnesses. And so the idea, LangChain would just
become the, tool that creates the operating system for the company, and
everybody will be  using LangChain to create their specialized harness, which
represents a workflow of the past. And now this harness inside that
workflow becomes autonomous, agentic, much more efficient. I think we see that these things are…
There's, the harness, there's the model, and then there's all the context
around it, and all of these can be optimized at different points in time. That's right. And so the work that, that we did
with Nemotron 3, I think was a great example of doing, some, pretty high
ROI things around the harness, changing the prompt, changing the tools.

[[13:55]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=835s) One of the things we're looking
forward to is experimenting with post-training Nemotron to, It, it,
takes a little bit more time, but I think it really raises the ceiling
of what this overall system can do. This is incredible. This is the big breakthrough. And so what you just described is a
future where, once you get the harness built and it's, built, it's doing the
work, and it's now part of the business process and it's very successful. Now the question is: how do we
get it even better than that? Of course, you can keep improving
the, information that you provided. You can tune the, harness, but you
can now also improve the AI model, the large language model, Nemotron
3 Ultra, inside the harness. I think that's a complete breakthrough. That's a capability that's never existed
before, and I'm super excited about that. And it's really gonna take all of these
enterprise-specific business processes and really start to tune this flywheel. And I think one of the things that
we've heard when talking to enterprises is the demand or need for this
to be built on an open ecosystem. This is, all this enterprise's
knowledge and processes that they're putting in there.

[[14:58]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=898s) And having full control over that
seems paramount to a lot of them. So I'm curious if you can touch on how
you see open stacks really empowering enterprises going further with AI. Every company is built fundamentally
on domain-specific or some specialized intellectual property. The reason why we call it intellectual
property, intellectual, it's intelligence. Every single company is built on
intelligence, some foundation of intelligence that's specialized. Our company is specialized in something. We're not good at everything, but
we're very, good at one thing, and every company is built that
way, and that specialization, your company's intelligence is who you are. You can't possibly not continue to control
it, improve it, make it better, right? And, somehow, outsourcing that
intelligence, whether you're a person, company, country, makes no sense to me.

[[15:59]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=959s) And of course, there's general
intelligence, and there are general things that we all do,
and it's a lot of the economy. And for example, software coding
is actually a general thing. We all program in Python, we all
program in C++, we all program, right? And, so you're applying it to
different things, but the, skill of coding is largely the same,
and that's a general skill. Writing is a general skill. But those are foundational skills
that we then apply for our specialized domain intelligence, and that's where
LangChain and Nemotron comes in. I think the foundation, of society
is going to have these foundational models, and they're gonna be general,
and they're gonna be, available in the cloud, and it's gonna be incredible. but on top of that platform,
we're gonna have to build our own specialized capabilities,
and you need open tools for that. And you can't outsource it. You can't… I, can't imagine
calling a third party, when I need to enhance my intelligence.

[[17:00]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=1020s) I need to enhance it right
here inside the company. And so, I think that future
is not one or the other. It's a completely complementary
vision and really what we're doing is just making sure that we, automated
intelligence is integrated, into all aspects of everything that we do. And as a result, we're
all gonna be better. Completely agree, and I think
it's still hard to get that integration up and running. And so one of the things that we're
announcing today is, a blueprint with Deep Agents and OpenShell inside of
the, NemoClaw, blueprints out there. And so this will let enterprises run
Deep Agents with Nemotron 3 Ultra inside of OpenShell, which is a
secure and open runtime- That's right and take advantage of that. This is one of just- such a, such a huge
deal … hopefully it makes it way easier for enterprises to get up and running. Such a huge deal. Yeah, all of the key ingredients
necessary for you to build your personal domain-specific, proprietary, your super
agent, all of the technologies, all the components, all the tooling, all of the
harnessing, everyth- and the blueprint, a great example, all put together for you.

[[18:11]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=1091s) How do you guys think about blueprints? You have many of them. This is- this is obviously the best one. I won't make you say that. Yeah. But I'll say that. This is the best one out there. But you have a ton of blueprints. Why, what is the, why
invest so heavily in them? Because the tools are, the
tools are arcane still, and, there are a lot of pieces to it. Building, building an agentic
system, building AI is not simple. And there's a lot of different
pieces of technology, and we already talked about some of them. There's the large language model, there's
the tool, the tools it uses, and, the knowledge graph that it has to deal with,
its memory system, and its, guardrailing system, and its fine-tuning system, and
now, the technology you're gonna create, the post-training against the harness. And then of course,
there's the harness itself. But what about the runtime? When you're done, you
still have the runtime. You have to keep it in a sandbox
so it's secure, it's private, that, that is access control. It's something that IT
organizations can control. Is that the hardest thing about
the runtime, you think, inside of enterprises, all the security
things that go alongside it?

[[19:13]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=1153s) Without, solving the security, the access
control, it's impossible to deploy. It's no different than it's
impossible to hire a new employee into the company if you don't onboard
them, give them access control. We don't give every, employee access
to every file and every network, right? And so you have every single
employee, based on their job and their responsibility and what they need to
have access to, we give them access to tools, the laptops and, design tools
and programming tools and whatnot. We give them access to
certain parts of the network. We give them access to information. We give, we connect them to other agents. We connect them to other colleagues
that they work within, and we provide them a skills file. You know we essentially, give
them a document about this is the, this is your mission. this is how it's previously
been done, and, and now, help, do it even better than that. And, so in a lot of ways, we are
creating an, HR system, if you will- You know, for AI that allows the IT
organizations and all of the different, business units inside the companies
to be able to build, improve, and deploy these agents inside companies.

[[20:21]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=1221s) This is more of a philosophical
question, but you're talking, and I think a lot of people talk about these
agents and anthropomorphize them- Yeah a lot, bring them into human systems. But, agents aren't human, and they
have some things that are better than, what humans are, and, they
have other places where they're, they are, very different and maybe not
as good as what humans are good at. What is the right level to
anthropomorphize these agents? It's, it's electrons it's
electrons, not atoms. and, it's not biological,
has no consciousness. It, it's not awake. and so it's not any of that. it's a tool that, it's like my vacuum
cleaner that's, roaming around the house. And it's, of course, roaming around
the house, cleaning up the house, doing something that I used to do. And, you now have autonomous lawnmowers,
and you have  and, so I, you could just imagine, a hundred years ago
when, the first dishwasher came along, and now it's doing dishes by itself.

[[21:23]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=1283s) It must have been magical to watch
it, and, we call it a dishwasher, which is a little bit like a human. Yeah. Yeah. And, we have dishwashers. when… my first job, I was a dishwasher. And and so in a lot of
ways, we'll get used to it. I think it's right now we tend to
imbue too much, human properties to it. It's nothing close to that. It's software. It's computers. We know exactly how it's
working because obviously, we created the harnesses around it. We obviously know how it works because
it's getting better all the time. If we don't understand how
something works, how do we make it better every time? And if we don't understand how
something works, how do we improve it? How do we fix it? And so obviously, we understand how
these things work and, I, I think that we ought to keep it there, And,
meanwhile, one of the things that we know is that the more AI we use,
somehow the more people we have to hire. And the reason for that is because
these agentic systems are new skills and, now we have a lot of
software engineers, building agents.

[[22:30]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=1350s) They used to code software, but
now they're building agents. If you ask me, every one of my software
engineers prefer to be building agents than to be writing Python code. Coding is like typing, and so
they're gonna do less typing. They're gonna they're gonna
be more systems engineers and more building engine… building and creating these autonomous
systems that are super cool. They're creating evals. They're creating benchmarks. They're creating guardrails. Isn't that right? And so the amount of work that we
have to do to bring AI into the world is really quite incredible. And so it's creating
a whole bunch of jobs. And, my software
engineers love, love this. I think we've seen-- You
mentioned evals briefly. I think we see that being a key
part to- Yeah … unlocking a lot of agentic usage inside an enterprise. You need to have some sense of, how
it's doing, and, quantifying whether it's good or not is oftentimes best
done by subject matter experts who already live inside the enterprise and
can easily give feedback and work with these systems to automate a lot of the
tedious parts of their job and then spend time on the really intellectually
stimulating parts and, the creative parts.

[[23:31]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=1411s) And so I- That's right. In a lot of ways, whether you're a doctor
or a designer or software engineer, you are creating an agent And, you're
taking, all the mundane work, and you're trying to get this agent to do it. But meanwhile, we're all trying to
get our agents elevated to do things with us that we couldn't do before. And so that, that requires
imagination, that requires creativity, a lot of technology. I think, that's spot on. I think currently the a lot of the best
usages that we see of agents are giving ourselves more leverage to do more things. Yeah. But I think a lot of that approach
is thinking about what did we do previously, and can we automate that? But I think a lot of the unlock
will come in the future of what just-- what couldn't we do before-
That's right … that now we can do. And so maybe, maybe- Ambition helps. A hundred percent. Right? Ambition, agency. Ambitions help. Yeah. Yeah, yeah. Maybe, on that vein, wrapping up, as you
think about how to help drive towards this future, what are some of the
missing pieces of this agentic stack? Today we're announcing a very big deal. This is a very big deal
thing that we're doing today.

[[24:33]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=1473s) We are providing, the basic building
blocks, the foundation, all of the the key ingredients, all of the key
ingredients to build super agents. These-- When I say super
agents, they're domain specific. They belong to you. You could… You build them, you
improve them, you refine them over time. You give them access to proprietary
information, knowledge, maybe it's super private to you. and as a result, this super agent will
be able to do things, that, you can't imagine, and it will be extremely, good. We've created all of the all of the
key parts, a, world-class language model, a framework called LangChain
Deep Agents that has now been also fine-tuned to expose the full potential
of Nemotron 3 Ultra, a blueprint that helps everybody do that, and of course,
the runtime, the OpenShell runtime that keeps it secure, the, acceleration
stacks that are all integrated into it.

[[25:42]](https://www.youtube.com/watch?v=Yy3JH6dDugc&t=1542s) And so every company in the world
should be able to, every developer in the world should be able to now
create these super agents, deploy it anywhere in the cloud, on-prem. a, good friend of mine just
built one for DGX Spark. And so now you have these agents running
on DGX Spark right next to your laptop. You could have it
running on a DGX station. you could build your own
supercomputer inside your company if you like, or do it in the cloud. We now have agentic, capabilities that
you can now build for yourself everywhere. All the pieces are now here. There are no excuses not to engage it. I think that's a perfect way to end it. You, got me so pumped up
when you were speaking. That was a great motivational speech. So I'm gonna go out and build some…
I'm gonna go out and build some agents. Thank you, Jensen, for sitting down. Congratulations. Thank- Good job. Thank you. Proud of you guys.


---
*Transcribed with [Tacit](https://www.trytacit.app)*