A new essay published on November 11, 2026, argues that the AI safety community’s exclusive focus on software alignment and digital detection is a critical oversight. The author, a physical scientist with professional AI experience, contends that natural scientists and hardware engineers possess the tools necessary to build robust, physical defenses against AI misbehavior. Despite well-funded calls for advancement in AI safety, the field remains dominated by 'bits people,' leaving the 'atoms people'—those who understand the physical constraints of hardware and material science—largely on the sidelines. This cultural divide prevents the implementation of defense-in-depth strategies that are standard in other high-stakes engineering disciplines.

The Gap Between Bits and Atoms

The author describes a profound culture shock when transitioning from managing hardware teams to software engineering, noting that 'in the world of bits... there is always someone to blame.' In contrast, hardware engineering is rooted in humility and fatalism because matter itself is prone to 'unpleasant surprises.' This mindset difference creates a barrier to entry for physical scientists. The AI safety community currently centers its work on model alignment and software-based misuse detection, treating the AI stack as a single point of failure. By ignoring physical vulnerabilities, the industry misses opportunities to secure the hardware that runs the models, such as implementing hardware-based kill switches that function without software intervention.

Lessons from Nuclear and Aerospace Engineering

The essay draws sharp parallels between AI safety and established industries like nuclear power and aviation. In nuclear energy, defense-in-depth is mandatory; every three years, civilian power plants undergo physical security drills led by the Nuclear Regulatory Commission, with findings that carry the force of law. The author proposes a similar rigor for AI data centers, suggesting passive safety features where hardware automatically powers down if a model cannot continuously prove alignment. Meanwhile, the aerospace industry demonstrates how to manage complexity through human-machine interfaces. The author cites the specific cost of airplane cockpits, noting that while they cost between a few hundred thousand and a few million dollars to build, the design phase alone costs on the order of $100 million. This investment yields a demonstrably good ROI by allowing two pilots to manage subsystems that no human could monitor alone, a model the author suggests could be adapted for AI inference systems costing thousands of dollars per hour.

AI as a Utility with Emissions

Beyond defensive engineering, the author proposes a regulatory framework based on the concept of AI as a new utility. The essay suggests viewing AI as creating a new utility (intelligence) with a new kind of emissions—defined as unwanted actions due to misalignment or misuse. This analogy invites the application of pollution control principles to AI governance. The author questions whether customers would pay a safety premium for AI, noting that green premiums suggest this is unlikely. Consequently, the essay argues that society may need to regulate the emissions from this growing intelligence grid, similar to how environmental agencies regulate industrial pollutants, to ensure that AI services do not create unchecked externalities.

Infrastructure and Economic Displacement

The author also addresses the economic impact of AI and robotics, warning that significant labor displacement is likely if optimistic predictions about physical AI come true. Drawing historical parallels, the essay points to the Great Depression as the last instance of mass human labor displacement. During that crisis, the key response was building infrastructure, which placed a huge premium on projects that were 'shovel-ready.' The author cites Robert Moses, as described in The Power Broker by Robert Caro, who was able to complete huge public works during this time of crisis by having wildly ambitious plans ready for construction and expert technical staff available to create yet more plans. The essay argues that we need 'atoms people' to dream up modern equivalents—such as terraforming projects, new transit systems, or an entirely new power grid—and to dream them in detail, at the level of concrete mixes and bolt sizes, before they are urgently needed.

Physics as a Security Layer

Natural systems exhibit deterministic, chaotic, and truly random behaviors that defy computational prediction, even with superior algorithms. This unpredictability can be leveraged for physics-based cryptography and 'honeypots,' where any AI accessing certain information must have stolen it. The author argues that these physical defenses are particularly relevant for mitigating risks like financial fraud, which can be countered by maintaining paper records. By integrating ecologists and physicists into safety discussions, the field could uncover strategies where 'lesser' organisms or physical laws restrain more powerful entities, a concept illustrated by the story of a leopard seal outwitting a human on an ice floe.

Key Takeaways

  • AI safety is currently siloed in software, ignoring the physical reality of hardware execution and environmental impact.
  • Nuclear and aerospace industries offer proven models for defense-in-depth, including passive safety mechanisms and rigorous physical security audits.
  • Airplane cockpits illustrate the ROI of human-machine interfaces, costing $100 million in design to manage complex subsystems effectively.
  • AI can be viewed as a utility with 'emissions' (misalignment/misuse), suggesting pollution control principles could inform regulation.
  • Economic displacement from AI requires 'shovel-ready' infrastructure planning, akin to the Great Depression response led by figures like Robert Moses.
  • Physical unpredictability in nature can be harnessed for security tools like physics-based cryptography and tamper-evident systems.
  • Cultural barriers between 'bits' and 'atoms' professionals prevent the adoption of hardware-centric safety measures.

The Bottom Line

If AI is going to break our bones, we need hardware engineers to build the cast. The current software-only approach to safety is fragile; we need to secure the metal, not just the math, and prepare our infrastructure for the economic shifts AI will inevitably cause.