On March 23, 2026, Nvidia Chief Executive Jensen Huang faced a thoroughly debated question on the Lex Fridman Podcast regarding when artificial general intelligence (AGI) will arrive. Without hesitation, Huang delivered an era-defining, almost declarative answer stating his belief that we have already achieved it.
Two days later, a stark reality check emerged that bordered on pure mockery when the highly anticipated results of the new ARC-AGI-3 benchmark were published. Human testers achieved a perfect pass rate, while top-tier artificial intelligence models scored near zero.
This abysmal failure was not a routine technical gap but a profound silence that rendered the prevailing technological narrative momentarily speechless. The comprehensive benchmark tests the ability to autonomously understand rules and solve problems in entirely unfamiliar scenarios.
The Limits Of Artificial Intelligence
TheARC-AGI-3 strict evaluation design deliberately strips away the data and pattern recognition capabilities where modern machine learning traditionally excels. While ordinary humans easily figure out the solution on their first attempt, current models behave as if their underlying soul has been extracted.
This massive failure strongly echoes the long-held view that true intelligence requires the initial comprehension of the unknown rather than optimally fitting known patterns. Current technology remains a highly optimized prediction machine that excels at continuing the world as it currently exists.
Shifting Definitions Of AGI
This staggering discrepancy raises an obvious question about what claiming the arrival of artificial general intelligence actually means in modern technological discourse. The unsettling answer is that this phrase has become a constantly redefinedtool, used to suit varying corporate agendas.
In corporate lexicons, the term simply describes an entity capable of generating immense economic value or efficiently running a massive enterprise. However, rigorous academic frameworks require genuine cross-situational generalization and the capacity for autonomous learning independent of human scaffolding.
When massive hardware companies brand processors as general intelligence chips, the terminology rapidly shifts from a scientific concept to a powerful narrative tool for capital markets. This vital narrative drives astronomical corporate investment and unprecedented government procurement across the globe.
The Hidden Environmental Cost
While global attention remains heavily captivated by cinematic fears of machine rebellion, a significantly more grounded and urgent environmental crisis is being dangerously ignored. The genuine threat is the immense energy consumption and massive carbon emissions generated by this computing power.
The electricity demand of large data centers is growing exponentially, heavily reliant on fossil fuels in most cases. In other words, every incremental advancement in an AI model likely contributes to a fractional rise in global temperatures. This impact is not abstract; it is already manifesting tangibly within the Earth's systems.
In Siberia, the Batagaika crater—dubbed the "Gateway to the Underworld"—is expanding by tens of meters each year. The thawing permafrost not only alters the landscape but also releases methane and carbon dioxide trapped for tens of thousands of years, creating a positive feedback loop for climate change. This is a perilous tipping point: once crossed, global warming will no longer require human activity to accelerate; it will become self-sustaining.
Ancient Threats And Tipping Points
Melting ice conceals ancient history, allowing modern scientists to resurrect dormant microbes from Greenland and successfully awaken prehistoric viruses from Siberian permafrost. A deadly recent anthrax outbreak proved this is a stark reality, representing archaic diseases aggressively returning to the present.
Taking a broader view, massive deposits of methane hydrates lurking directly beneath the Antarctic ice represent an incredibly severe planetary threat. If released by shifting temperatures, these deposits could produce a short-term greenhouse effect dozens of times more potent than carbon dioxide.
This catastrophic development transcends standard climate change to represent a fundamental and potentially irreversible state shift within the entire Earth system. Scientists warn that such a chain reaction could cause a severe mass extinction not seen in millions of years.
A Necessary Paradigm Shift
We are currently caught in a state of misplaced collective anxiety where we fear non-existent machines while willfully ignoring the warming already underway. We obsessively debate whether technology will surpass humanity while rarely discussing how we are already critically overdrawing the planet's resources.
Over the past half-century, computing paradigms have evolved from general logic processors to architectures dominated by massive parallel processing algorithms. Today, we stand on the crucial threshold of adopting neural processing units, which represents a necessary course correction rather than just a performance upgrade.
Unlike centralized cloud architectures, neural processing units prioritize highly efficient edge computing that allows inference to be completed locally in real time. This desperately needed innovation dramatically slashes regional energy consumption and effectively reduces the massive hidden carbon footprint associated with data transmission.
The ultimate transition toward decentralized computing architecture ceases to be solely a matter of technological efficiency and becomes a profound civilizational choice. The true test remains whether we can ensure our fragile world continues to exist while we aggressively forge more powerful technology.
* The author is the founder and chairman of Kneron. （Related： KMT Chair Cheng Li-wun Vows To Ease Cross-Strait Business Friction During Shanghai Visit ｜ Latest ）
