"AI Is Evolving Beyond Generation to Actually Performing Work" – Jensen Huang [Global AI Briefing]

[The Financial News] Nvidia CEO Jensen Huang stressed that "the era has begun in which AI goes beyond simply generating answers and starts handling tasks on behalf of users." He noted that while AI had long been limited to generating and analyzing text, since the emergence of OpenClaw it has been evolving into a stage where it can actually carry out real-world work.
On the 18th (local time) at San Jose Civic in California, Huang spoke in a panel discussion on open source with the heads of major AI companies. He predicted, "Every application will be powered by AI, and every country and company will build and use AI," adding that open models will play a crucial role in this transition.

Mira Murati, founder of Thinking Machines Lab, who joined the discussion, remarked, "The pace of AI progress is exponential, and large research labs alone cannot carry out all the necessary research." She continued, "This is precisely where open models can be tremendously helpful, and this applies not only to the models themselves but also to the underlying infrastructure."
Jensen Huang described the AI agent platform OpenClaw as a "reinvention of the computer." He called it "the most popular open-source project in history" and emphasized that "in just a few weeks, it has shown the world the potential and vision of agentic systems." At the same time, he cautioned that "trying to solve everything with a single general-purpose agent makes systems unnecessarily expensive and slow," arguing that a systems-level approach that combines agents in series and in parallel is essential. Separately, Nvidia introduced an AI agent platform called "NemoClaw" at the GPU Technology Conference (GTC) on the 16th (local time).
China’s DeepSeek has stunned the U.S. tech industry by unveiling its large language model "DeepSeek-R1." The model delivers performance on par with competing U.S. systems, yet it was developed at a far lower cost and has been released free of charge.
According to MIT Technology Review on the 18th (local time), the debut of DeepSeek-R1 wiped out roughly 1 trillion dollars (about 1,502 trillion won) in market value from U.S. stock markets. U.S. President Donald John Trump called it "a wake-up call." Marc Andreessen, a leading Silicon Valley investor, stated, "DeepSeek-R1 is one of the most impressive breakthroughs I have ever seen, and as open source it is a gift to the entire world."
DeepSeek-R1 was trained using an automated reinforcement learning approach that minimizes human intervention. Instead of people, computers evaluate the quality of answers and feed those evaluations back into training. Large language model (LLM) AI systems are built through a pretraining phase followed by post-training. In post-training, they typically undergo supervised fine-tuning (SFT) on human question-and-answer data and reinforcement learning from human feedback (RLHF) to become viable for real-world services. DeepSeek-R1 reduces costs by minimizing human involvement in this post-training stage. MIT Technology Review noted, "Computers are strong at evaluating problems with clear right answers, such as math or code, but they are weak at judging creative or subjective questions." The outlet added, "This method has the advantage of dramatically lowering costs, but as a result DeepSeek-R1 excels particularly in mathematics and coding, while for general questions or creative tasks some degree of human feedback is still needed."

cafe9@fnnews.com Lee Gu-soon Reporter