NVIDIA CEO Jen-Hsun Huang recently made another blockbuster statement in public, claiming that "we have achieved artificial general intelligence (AGI)." However, his expressions and examples in different interviews also revealed that there is still a big gap between the emerging stage of artificial intelligence capabilities and "human-level intelligence."
On March 22, Huang was asked in an interview with podcast host Lex Fridman when AGI would be achieved. According to Fridman's definition, AGI should be "a system that can do your job, including starting, running and growing a billion-dollar-plus technology company from scratch." When asked to give a timetable of 5, 10 or even 20 years, Huang responded without hesitation, "I think it's now." However, he immediately added that the other party mentioned "1 billion U.S. dollars" but did not say "how long it can last." This is equivalent to understanding AGI as a staged threshold for commercial achievement, rather than a long-term stable "mind" with human-like cognitive abilities.
Huang Renxun made it clear in the same interview: "I think we have achieved AGI." He regarded the current class of AI agent systems with high hopes as examples, believing that these systems have shown strong entrepreneurial potential in launching social applications and conducting creative experiments. He mentioned that open source projects including OpenClaw (an open source AI agent platform being acquired by OpenAI) are spawning a wave of "entrepreneurial waves" using digital agents to create virtual influencers and automated digital communities, which seems to prove that AGI has "arrived".
But while emphasizing this wave, Huang also admitted that the current success of these systems is highly accidental. He pointed out that many projects "cool down after two months of fire" and bluntly said that "the probability of actually making one NVIDIA out of 100,000 such agents is zero." This statement actually acknowledges that the current AI agents are still very far away from truly possessing the comprehensive ability to systematically build and operate ultra-large-scale technology enterprises for a long time. It also weakens his previous radical judgment that "AGI has been achieved".
The statement surrounding "whether AGI has been realized" has also touched upon long-standing differences in the industry. The concept of general artificial intelligence itself has long been highly "politicized" and "capitalized." In the contract terms between companies such as OpenAI and Microsoft and their partners, whether "AGI has been achieved" is often directly tied to huge amounts of money and strategic direction. Therefore, any statement claiming that "AGI has been achieved" can easily inflame debate. Technology leaders and researchers have been debating for years whether current large-scale model systems embody "true general intelligence" or are simply highly simulated fragments of human intelligence.
It is worth noting that just three days before the release of Fridman’s interview on March 19, Huang Jen-Hsun gave a completely different emphasis on the practical application of AI when he was a guest on the All-In Podcast during the Nvidia GTC (GPU Technology Conference). In that conversation, he did not focus on whether AGI was "realized," but instead focused on whether human engineers were making full use of existing AI tools, and even warned in strong terms that he would be "deeply shocked" if highly paid engineers did not spend enough money on AI.
“If an engineer with an annual salary of US$500,000 does not consume at least US$250,000 in AI tokens a year, I would be very vigilant,” Huang said. He explained “token” as the basic unit of measurement for large model processing and language generation, and also a direct reflection of AI computing costs and work capabilities. In his view, insufficient use of tokens means that engineers do not have the ability to fully utilize AI, which is equivalent to wasting potential productivity.
Huang even compared this behavior to chip designers refusing to use electronic design automation tools such as CAD and "returning to the paper-and-pencil era" to draw circuit diagrams. He revealed that Nvidia is currently trying to set aside about $2 billion in token budget for the engineering team and is considering incorporating tokens directly into employee compensation packages. He imagined that in addition to a basic salary of hundreds of thousands of dollars per year, an engineer would also receive a "token quota" worth about half of the basic salary, so that he could "amplify his work efficiency by 10 times" by using a large number of AI tools.
Looking at these two public appearances several days apart, Huang Jen-Hsun, on the one hand, used highly provocative language to claim that AGI has "arrived" and met a certain definition of "general intelligence" oriented towards business achievements; on the other hand, when talking about internal management and engineering practices, he emphasized that the current AI system is still highly dependent on human initiative and intensive use, and the probability of success in real-world entrepreneurial practice is extremely low. The tension between the previous and later remarks exactly reflects the subtle swing between the "hype frontier" and "realistic capabilities" of the current AI industry: while competing for resources by exaggerating prospects, one has to admit that technology is still far from reaching the ideal "human-level intelligence" in many key dimensions.