Exclusive Interview with Jensen Huang: In the Past Six Months, AI Has Finally Become Useful

In a recent in-depth conversation, Jensen Huang, Founder and CEO of NVIDIA, and Harrison Chase, Founder of LangChain, discussed the inflection point of the AI industry, the necessity of open ecosystems, and how enterprises can build their own "super-agents".

The following is a curated transcript of the discussion:

1. The Past Six Months: AI Has Finally Become Useful

Jensen Huang opened with a striking observation:

"We've been working on AI research for 15 years. But in the past six months, everything has changed."

He explained that all foundational technologies — advances in large language models, scaling breakthroughs, and multimodal capabilities — have been building up over the past few years. But only in the last six months has "everything finally fallen into place".

"Now, AI is finally delivering real value. When AI works, every company in the world benefits. Every enterprise wants it."

The question is no longer "is AI useful?" but "how do we implement it?" This is exactly where LangChain comes into play.

Jensen Huang looked back on the history of collaboration with LangChain:

"In the early days, we used LangChain to help turn a large language model into a promptable API. Then we used LangChain to build RAG (Retrieval-Augmented Generation). Step by step, we eventually arrived at the intelligent agents we have today."

He specifically emphasized the biggest breakthrough of the past six months:

"These information and knowledge-based agent systems can use tools to search, have self-managed memory, are equipped with security guardrails, and can iterate repeatedly until tasks are completed."

He believes that OpenClaw has truly ignited people's imagination for agent systems. "Cloud Code made all this possible."

2. Why Must It Be Open?

Jensen Huang clearly articulated the core reason for NVIDIA's investment in the open agent ecosystem:

"AI is a foundational technology. It can only deliver value when applied to a huge variety of different use cases."

He painted a vision:

"Scientists, digital biologists, designers, roboticists, students, researchers, enterprise IT teams — all of us can use AI to solve domain-specific problems."

The key point is that every enterprise has its own unique intellectual property.

"Many of the problems we want to solve involve specialized expertise that is not accessible externally. We believe that when we put AI into a flywheel, it becomes a super-agent — the more we use it, the smarter it gets, and the smarter it gets, the more we use it. It's just like humans, learning continuously over time."

Harrison Chase shared astonishing data comparing Nemotron 3 Ultra to cutting-edge models:

"After fine-tuning, we successfully integrated Nemotron 3 Ultra into DeepAgent. In internal benchmark tests, it achieved a score of about 86%, while Claude Opus scored 87%."

At the same time, open models like DeepSeek and MiniMax have also reached around 82-83% performance levels.

"We are starting to see some newer open-weight models truly reach state-of-the-art performance levels. Equally importantly, it is several times cheaper than Opus."

Jensen Huang responded: "I could not be prouder. This is absolutely incredible."

Jensen Huang deeply analyzed several manifestations of this cost advantage:

"When you have cost-effective intelligence, people simply use it more. When you find cost-efficient agents, you can expand the scope of your search, which in turn leads to better results."

He specifically highlighted Nemotron's advantages:

"Nemotron is cost-effective because it is extremely fast and computationally highly efficient. When computational efficiency is high enough, it can explore a much larger space."

"This is no different from a person with a fast reaction speed — you can explore a far wider space and find better answers. That's why Nemotron has such a huge advantage in the LangChain framework and DeepAgent."

He concluded: "We created a model that is close to the cutting edge, but by re-engineering the environment around it, we enabled it to achieve state-of-the-art capabilities."

When asked "how to best specialize these systems", Jensen Huang's answer was clear:

"The first step in specialization is that you need to have sufficiently capable intelligence. Nemotron Ultra is excellent as an entry-level model, but when combined with the LangChain framework, it becomes an outstanding model — LangChain is built around it to ground it in domain-specific information."

"A smart person becomes extremely useful when we give them access to particularly important information. Access to information is critical."

He emphasized that enterprises should not think in terms of "cutting-edge model vs. open model" but rather "use both":

"Cutting-edge models are constantly improving. They still have a long way to go, and scaling laws will continue to hold. At the same time, we must also create our own specialized super-agents — proprietary agents powered by LangChain and Nemotron."

"In fact, they can be considered your most precious assets."

3. The Future of Enterprises: Built on Frameworks

Jensen Huang made a bold prediction about the future of enterprise AI:

"Today most companies are built on business processes. In the future, most companies will be built on frameworks."

"The vision of LangChain is that it will become a tool for companies to create operating systems — everyone will use LangChain to build their own proprietary frameworks that represent their legacy workflows."

"Now, the components in that workflow become autonomous — more proactive and far more efficient."

He further elaborated on the different layers of this vision:

"The guardrails are there, the model is there, and all the contextual information around it is there. All of these can be optimized at different points in time. We are tuning prompts, tuning tools. In the future, we will also perform post-training model optimization — which truly raises the upper limit of what the entire system can achieve. This is a major breakthrough."

Jensen Huang clarified why enterprises must build on top of open ecosystems:

"Every company is built on domain-specific intellectual property. The reason we call it intellectual property is — intelligence is wisdom. Every company focuses on a particular field, rather than trying to be good at everything. That's how every company is built."

"This specialization, the wisdom of your company, defines who you are. You cannot possibly not continue to control it, improve it, and make it better. Outsourcing that wisdom makes no sense to me."

He acknowledged that general skills exist: "Software coding is a very universal thing, we all program in Python. Writing is a universal skill. But these are foundational skills. We can apply these skills to our domain-specific wisdom — this is exactly where LangChain and Nemotron fit in."

"Society will have foundational models that are universally applicable, available in the cloud, and extremely powerful. But beyond that, we must build our own specialized capabilities, and that requires open-source tools."

"I can't imagine turning to a third party when I need to enhance my company's wisdom. I need to — inside the company — augment it."

Harrison Chase announced the latest achievement of the NVIDIA and LangChain collaboration:

"We are announcing today — NemoCore — a blueprint that includes DeepAgent and OpenShell. This will enable enterprises to run DeepAgent with Nemotron inside OpenShell — a secure and open runtime."

Jensen Huang emphasized why this matters:

"These tools are still complex, with many different components. Building an agent system is no easy task: language models, tools, knowledge graphs, memory systems, guardrail systems, fine-tuning systems, runtime environments — all of these pieces are required."

"And the runtime is the hardest part — you have to sandbox it, ensure it's secure and private, and make access control something the IT organization can manage. You can't deploy it without solving the security problem. It's exactly like how you can't onboard new employees into a company without solving the onboarding process."

"Now, all the key pieces are here — world-class language models, frameworks fine-tuned to unlock their full potential, blueprints to help everyone get started, and runtime environments that ensure security. There's no reason not to get involved."

4. Don't Anthropomorphize, But Be Ambitious

When asked "how to properly anthropomorphize these agents", Jensen Huang's answer was straightforward:

"This is electrons, not atoms. It's not biological. It has no consciousness. It's a tool — a bit like my vacuum cleaner. We call it a dishwasher, it's a little anthropomorphic, but we know exactly how it works."

"If we don't understand how something works, how can we improve it? So obviously we understand how these systems operate."

But at the same time, he emphasized the value of ambition:

"The more AI we use, the more people we end up needing to hire. These agent systems represent entirely new skill sets — previously we had many software engineers writing code, now they are building agents. Every one of my software engineers prefers building agents over writing Python code."

"Programming is becoming as common as typing — they are evolving into systems engineers and architect engineers, building these super cool autonomous systems. They are defining evaluation criteria, creating benchmarks, and designing guardrails."

Finally, he encouraged everyone:

"Now developers all over the world should be able to create these super-agents — deployed anywhere, in the cloud or on-premises. All the key pieces are here. There's no reason not to get started."

Harrison Chase responded: "While you were talking, I got so excited I felt my blood pumping. That was an amazing inspirational speech, so I'm going to go out and build something, I'm going to go out and recruit some agents."

From language models to agent systems, from general AI to proprietary super-agents — the future Jensen Huang depicts is one where every company can build its own continuously learning intelligent agent. And the foundation of all this is an open ecosystem.

This article is from the WeChat public account "AIGC Index", author: Mark, published with authorization from 36Kr.