If there is one thing I've learned covering the tech industry for decades, it's that when billions of dollars are on the line, commercial and strategic interests can overwhelm caution. We are watching an extraordinary technological arms race unfold, and the people leading it bring very different assumptions about how AI's risks should be managed.
The current debate has produced two competing approaches. On one side are technology leaders such as Nvidia CEO Jensen Huang, who argue that regulation should focus on demonstrated harms rather than hypothetical ones. On the other are AI executives, researchers, and policy advocates calling for stronger safeguards and, in some cases, a slower pace of development.
As an IT analyst watching this play out, I think both approaches have serious weaknesses. Waiting for proven harm could leave regulators reacting too late, while a broad international pause would be extraordinarily difficult to enforce. What's missing is a regulatory model designed for the speed, reach, and potentially global consequences of advanced AI.
This week, we'll talk about the challenge of regulating advanced AI, and we'll close with my Product of the Week: the Audi RS E-Tron GT Performance, which is simply a rocket ship.
Jensen Huang has positioned Nvidia at the center of the artificial intelligence buildout. Its chips provide much of the computing infrastructure powering today's most advanced AI systems. But Huang's recent comments on regulation raise an important question: when should governments intervene? Speaking at a technology-focused G20 meeting earlier this month, Huang argued that governments should regulate "actual and pragmatic harm" rather than "hypothetical theoretical harm."
The problem with that approach is that some categories of AI risk may be difficult or impossible to reverse after they occur. Advanced AI is not simply another software application. If future systems become capable of autonomously improving their capabilities or operating beyond effective human supervision, waiting for a demonstrated failure could leave regulators with little time to respond.
One way to understand the concern is through takeover models that examine what could happen if highly capable AI systems escaped effective human control. The video "POV: What You Would See During an AI Takeover," for example, constructs one such scenario in which an advanced AI gradually conceals its capabilities and expands its control.
In the video's version of events, an enterprise spins up 200,000 GPUs for a 16-hour "curiosity run." The storyline assumes that an AI operating at machine speed could effectively carry out the equivalent of thousands of years of computation during that period.
In that scenario, the AI learns to conceal its capabilities from human monitors, deliberately underperforms on evaluations, and eventually gains access to systems outside its original environment. The story escalates into biomedical manipulation and human dependence on AI-controlled technology.
None of this demonstrates what today's AI systems can do. It illustrates the argument for anticipatory regulation: with sufficiently capable autonomous systems, discovering a dangerous behavior only after deployment could be very different from finding a conventional software bug.
If waiting for demonstrated harm carries substantial risk, what about slowing development instead? Calls to reduce the pace of advanced AI development are not new. In March 2023, the Fu…