Every few months, a politician stands behind a podium, stares earnestly into a television camera, and demands that the United States and China hold hands and sign a treaty to "pause AI" and "ban superintelligence."
It sounds noble. It plays brilliantly on cable news. And if you understand anything about how silicon, software, and geopolitics actually work, you know it is pure, unadulterated theater.
The lazy consensus sweeping through Washington and mainstream news outlets treats artificial intelligence like enriched uranium. The logic goes like this: if a US Senator simply asks the White House and Beijing to agree on a global moratorium, both superpowers will politely power down their data centers, lock up the weights, and wait until international regulators write a sensible rulebook.
This worldview is not just naive. It is actively dangerous. By framing the debate around an unenforceable global ban on "superintelligence," lawmakers are distracting the public from real, solvable engineering problems while hand-delivering an asymmetric advantage to actors who do not care about Senate subcommittees.
The Physics Of Software Will Never Obey A Treaty
To understand why a "pause" is a fantasy, you have to look at the physical reality of what we are trying to regulate.
When the world regulated nuclear weapons, inspectors could look for massive, unmistakable physical footprints. Centrifuges take up acres of space, require specialized alloys, and draw absurd amounts of power from dedicated grids. You cannot hide a Manhattan Project in a basement.
Artificial intelligence is math running on commercially available silicon.
While frontier training runs currently require tens of thousands of high-end GPUs, the efficiency curve of machine learning is collapsing the compute threshold required to reach high capability. Algorithmic breakthroughs, better data curation, and inference-time compute scaling mean that what required a billion-dollar cluster two years ago can increasingly be approximated on a fraction of the hardware.
I have sat in boardrooms where executives burned tens of millions of dollars trying to build proprietary moats around models, only to watch an open-weights release from a small research lab match their performance six months later on consumer-grade hardware. If private companies with limitless capital cannot stop the diffusion of model capabilities, a bipartisan committee in Washington certainly cannot stop it with a piece of paper.
You Cannot Ban A Word Nobody Can Define
Ask ten AI researchers to define "superintelligence" and you will get twelve different answers.
Is superintelligence a system that can pass every standardized human exam? We already have models doing that. Is it a system that can autonomously discover novel physics? Or is it simply an arbitrary threshold of floating-point operations (FLOPs) used during training?
Regulators love hard numbers because hard numbers are easy to write into law. That is why early executive actions attempted to draw a line in the sand at $10^{26}$ FLOPs.
The problem is that compute thresholds are a terrible proxy for capability or danger. You can spend $10^{26}$ FLOPs training a massive, bloated model on garbage data and end up with an expensive random-text generator. Conversely, a lean, well-architected system trained on high-signal synthetic reasoning data can outperform a giant model while flying completely under the statutory radar.
When politicians call to "ban superintelligence," they are asking the government to outlaw a ghost. Worse, any legal definition rigid enough to enforce will immediately be bypassed by researchers changing their training architectures, while any definition broad enough to catch everything will criminalize basic scientific research.
The Prisoner's Dilemma Nobody Wants To Admit
Let us run a thought experiment. Imagine a scenario where the President of the United States and the General Secretary of the Chinese Communist Party actually sign a binding agreement tomorrow morning to halt all frontier AI development.
What happens at 12:01 PM?
Neither side trusts the other. And neither side can verify compliance without total, invasive surveillance of every server farm, university lab, and defense contractor inside the other's borders.
In game theory, this is a classic Prisoner's Dilemma with catastrophic payoffs. If both sides cooperate and pause, progress stops. If one side secretly defects—masking training runs behind cloud gaming infrastructure, disguising GPU clusters as weather-simulation centers, or routing compute through shell companies in third-party nations—the defecting party gains an insurmountable economic and military lead.
Beijing knows this. Washington knows this. Even the engineers building the models know this.
Openly advocating for a unilateral or bilateral pause assumes that geopolitical rivals view AI through the lens of shared existential safety rather than national survival. In reality, both nations view advanced AI as the defining general-purpose technology of the century—equivalent to electricity or the internal combustion engine. Asking them to hit the brakes is like asking 19th-century industrial powers to ban steam engines because boilers occasionally exploded.
Why The "Pause" Crowd Is Asking The Wrong Question
When people search for whether we should pause AI or ban superintelligence, they are usually driven by a legitimate fear: What happens when systems become smarter than us and we lose control?
That is a valid concern. But asking "How do we stop people from building it?" is the wrong question. The right question is: "How do we build defensive infrastructure that survives in a world where high-capability AI is ubiquitous?"
Right now, the policy conversation is entirely dominated by two equally unhelpful tribes:
- The Doomers, who demand an impossible global shutdown to prevent sci-fi extinction scenarios.
- The Accelerationists, who argue that any regulation whatsoever is treasonous and that the market will magically sort out catastrophic risks.
Both camps are wrong because both ignore engineering discipline. As Yann LeCun and Andrew Ng have repeatedly pointed out in their critiques of existential panic, treating current autoregressive architectures as nascent digital gods distracts from the actual, mundane ways AI fails today. And as Geoffrey Hinton and Yoshua Bengio rightly warn, ignoring long-term alignment and control problems as systems scale is reckless.
The middle ground is not a pause. It is hardened verification and defensive acceleration.
What Actually Works Instead Of Theater
If lawmakers actually want to protect the public from runaway AI systems, cyberattacks, and engineered pathogens, they need to stop writing open letters to foreign leaders and start funding unglamorous, concrete guardrails.
1. Mandate Cryptographic Provenance For Compute And Content
Instead of trying to ban model weights, track the physical hardware. High-end AI accelerators are manufactured by a tiny, concentrated supply chain—primarily TSMC, ASML, and Nvidia. Embedding hardware-level telemetry and cryptographic verification directly onto datacenter-grade chips allows auditors to verify that massive training runs are adhering to safety evaluations without exposing proprietary model weights or trade secrets.
2. Shift Liability To Deployment, Not Research
Stop trying to regulate the math. Regulate the harm. If a company deploys an autonomous agent that causes financial collapse, breaches critical infrastructure, or facilitates biological harm, hold the deploying entity strictly liable—just as we do with aviation, pharmaceuticals, and chemical manufacturing. When executives face personal and financial ruin for deploying unverified systems, safety stops being a PR talking point and becomes an engineering requirement.
3. Build AI-Powered Cyber Defense Before Offense Scales
If you are worried about superintelligent systems breaking into power grids and banking networks, pausing domestic research leaves you defenseless against adversaries who did not pause. Governments should be pouring capital into formal verification, automated vulnerability patching, and AI-driven cyber defense systems right now. You do not beat a faster adversary by tying your own shoelaces together; you build better armor.
4. Require Standardized Red-Teaming With Teeth
Right now, frontier labs essentially grade their own homework. They hire third-party evaluators, publish sanitized safety reports, and release models on their own timelines. True oversight requires independent, government-backed red teams—staffed by actual offensive security engineers, not policy generalists—with the statutory authority to delay commercial deployment if a model demonstrates autonomous replication or cyber-exploitation capabilities.
The Downside Of My Approach
I am not going to pretend that defensive acceleration and deployment liability come without severe trade-offs. They do.
First, strict deployment liability will inevitably slow down enterprise adoption and crush smaller startups that cannot afford massive compliance teams or liability insurance. It will consolidate market power among the tech giants who already own the compute and the legal armies.
Second, hardware-level telemetry walks a razor-thin line toward total state surveillance of computation. If governments can audit every large cluster, the temptation to use that same infrastructure to stifle political dissent or censor open-source developers will be immense.
Those are real, ugly problems. We should be fighting over how to solve them—how to protect open-source innovation while enforcing liability, and how to audit hardware without building a digital panopticon.
Instead, we are wasting precious legislative bandwidth pretending that a senator's press release can convince rival superpowers to abandon the most consequential technology in human history.
Stop Wasting Time On Fantasies
History is unambiguous on this point: no major technological wave has ever been stopped by a moral plea for a timeout. Not the printing press, not gunpowder, not nuclear fission, and certainly not software.
Every hour policymakers spend debating an unenforceable ban on "superintelligence" is an hour they are not spending hardening critical infrastructure, securing the semiconductor supply chain, or writing liability laws that actually hold tech companies accountable for the products they ship.
You cannot legislate away mathematics. You cannot treaty your way out of game theory. And you definitely cannot pause the future just because it makes you uncomfortable.
Either build the defenses required to survive a high-AI world, or get out of the way of the engineers who are trying to.
