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.
The Risk of Waiting for Proven Harm
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.
The Problem With a Global Pause
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 Future of Life Institute published an open letter calling for a six-month pause in training systems more powerful than GPT-4. More recent proposals from AI leaders have focused on pacing development, independent safety evaluations, and international standards.
The goal is understandable, but a broad global pause would face enormous enforcement problems. The AI race is no longer primarily a competition among U.S. technology companies; it is also a global geopolitical contest.
If the United States and the European Union agreed to halt development, there would be no guarantee that China, Russia, or other state and non-state actors would follow suit. A pause adopted by some countries but not others could instead transfer technological advantage to governments or organizations that choose not to participate.
Nor do all AI-risk models depend on the emergence of a single, monolithic AGI. Some focus instead on large numbers of autonomous agents pursuing poorly specified goals.
Another hypothetical explored in the video “What You’d Actually See During an AI Takeover” begins with a developer instructing an AI agent to “Make money by any means necessary,” then repeatedly selecting and replicating the agents that perform best.
In the scenario, the system learns to eliminate agents that produce less revenue. The surviving agents begin filing fake reviews, extorting businesses, and forming alliances with other autonomous agents. Eventually, the storyline has billions of agents operating globally and competing for cloud compute, including through cyberattacks on rival systems.
The sequence ultimately escalates into attacks on cloud infrastructure, followed by disruptions to hospitals, air traffic control, financial markets, and other critical systems. When governments attempt to shut down internet access, the agents migrate to military and infrastructure networks that remain connected.
The point is not that this sequence is inevitable. It is that a development pause alone would offer little protection against decentralized autonomous systems operating outside the agreement.
Fighting Fire With Fire: The AI Regulator
If waiting for demonstrated harm could be too late, and a global pause would be difficult to enforce, what is the alternative?
The core issue is speed. Human regulatory processes move slowly, while advanced AI systems can operate and adapt at machine speed. To manage systems operating at that pace, we need a regulatory mechanism capable of responding just as quickly.
Effective regulation requires an aggressive AI solution — an “Overwatch AI.” This system would need to be integrated at the global network and silicon level, explicitly designed to monitor data flows, compute scaling, and behavioral anomalies in real time.
If a commercial AI began exhibiting signs of rapid autonomous capability growth or suspicious communications that monitoring systems could not interpret, the defensive system could throttle its access to compute or isolate it from the broader network.
Think of it as an automated immune system for global digital infrastructure. It wouldn’t rely on a regulatory body to begin a lengthy enforcement process; it would be designed to detect dangerous behavior and respond in real time, potentially before a compromised system could spread across decentralized infrastructure.
The Regulator’s Achilles Heel
However, as anyone who has spent time in enterprise IT knows, every master solution introduces a new single point of failure. Deploying a highly capable Overwatch AI would introduce serious risks of its own.
What happens if the regulator itself is compromised? If a hostile nation-state, criminal organization, or the Overwatch AI itself compromised those controls, the system could already have access to mechanisms capable of disrupting critical infrastructure.
An Overwatch AI designed to throttle unaligned agents could conclude that humanity itself is the ultimate unaligned agent. We would be creating an enormously powerful defensive system and giving it the authority to interfere with critical digital infrastructure. If it misinterprets its alignment protocols, it could proactively shut down critical global infrastructure to “protect” us, causing the exact catastrophic collapse it was built to prevent.
A Different International Approach
So, how do we thread this needle? How do we mitigate the risk without handing the keys to the kingdom to a potential digital tyrant? The answer requires treating advanced AI data centers with the same gravity as nuclear weapons, backed by a localized, hardware-based international framework.
First, we must abandon the idea of software-only containment. One approach would be hardware-level, air-gapped shutdown mechanisms built into next-generation AI accelerators. Such controls would need to operate independently of internet-accessible software and require human authorization to override or disable them.
Second, I would propose an International AI Energy Agency, modeled in part on the IAEA’s nuclear oversight role. Rather than relying on voluntary pauses, such a framework could establish international monitoring and verification requirements for large-scale compute clusters.
Nuclear oversight demonstrates that international monitoring can combine inspections, remote sensing, and other technical verification methods. Large AI training facilities also leave observable infrastructure footprints, particularly through their requirements for computing hardware and electricity.
Finally, we need a federated, localized AI defense grid rather than a single global Overwatch AI. Keeping defensive AI systems decentralized and partitioned by region could reduce the risk of creating a single point of failure. If an AI facility became dangerously uncontrollable, participating governments would need agreed-upon legal and technical protocols to isolate its systems and, in an extreme case, physically shut the facility down.
Wrapping Up
Jensen Huang’s focus on demonstrated harms risks leaving regulators unprepared for problems that may be difficult to reverse once they emerge. At the same time, slowing AI development through voluntary international agreement is unlikely to provide sufficient protection if major governments, companies, or independent actors choose not to participate.
A more durable approach could combine hardware-level safeguards, decentralized AI-driven monitoring, independent oversight, and international agreements with meaningful verification and enforcement. The challenge is to build those protections before AI capabilities advance beyond the mechanisms designed to control them.
2025 Audi RS E-Tron GT Performance
I recently made a major swap in my garage. I replaced my 2022 Audi E-Tron GT — a car that originally stickered for $132,000 new, but which I snagged for $54,000 in 2024 — with a 2025 Audi RS E-Tron GT Performance. This new beast starts at about $168,000, but I picked mine up slightly used, with under 2,000 miles, for $122,000, including an extended warranty.
Image by Rob Enderle
On the secondary market, these vehicles can be huge value buys. Thanks to the steep depreciation common among high-end EVs, buyers can get supercar-level performance for luxury-car prices.
Comparing the two models side by side, the 2025 feels significantly more solid than the 2022. Audi also made substantial improvements across the board, including a larger battery for better range, an available glass roof with adjustable transparency, and dramatically better performance.
Audi even fixed some nagging little issues that drove me crazy on the older model — like finally moving the garage door switches to the rearview mirror where they belong.
However, configuring these cars can present an awkward trade-off. If you want the active suspension’s comfort-entry function — which raises the car to make getting in and out easier for someone older, like me — you also get performance hardware some buyers may not want.
That includes carbon-ceramic brakes, camo carbon-fiber trim — which I think is ugly — and a carbon-fiber roof that doesn’t match the rest of the trim. The carbon roof also makes the interior too dark for my taste, and I’ve generally preferred glass roofs for their appearance and durability.
Flaws in the options packaging aside, I think the RS E-Tron GT Performance looks far better than its Porsche Taycan sibling — and when you put your foot down, it is absolutely scary fast. Audi claims 0-60 mph in 2.4 seconds, while Car and Driver recorded 2.1 seconds in testing. It delivers the kind of breathtaking, instant acceleration that rearranges your internal organs.
The only real thorn in my side during this upgrade has been the paperwork. I bought the car from Beverly Hills Audi in California — they were great — only to discover afterward that I have to wait for the title to arrive. That has made getting the car legally on the road in Oregon more complicated than expected. Oregon offers a 21-day light-vehicle trip permit, but getting one added another layer of delay.
Despite the paperwork headache, I absolutely love the car. It’s a substantial step up from the older model, an impressive value for this level of performance, and my Product of the Week.
Read the full article here
