Why I'm not worried about AI: tools, compute, and the Skynet myth

Listen to the loudest voices in technology today and you will quickly find two entrenched camps: those who claim artificial intelligence is our final invention before Skynet terminates human civilization, and those who demand that open-source models be preemptively outlawed out of sheer panic.

Both camps are dead wrong. And they are wrong for the exact same reasons: they fail to understand the fundamental nature of tools, they ignore the physical constraints of compute hardware, and they misidentify where the real societal danger actually lives.

I have spent forty years working with computers, operating systems, networks, and infrastructure. I have watched successive waves of technological disruption arrive, and I have seen the exact same reactionary hysteria replay every single time humanity builds a more powerful lever.

I am not worried about AI. Here is why.

The Eternal Duality: From Fire to the Split Atom

Not a single transformative technology in human history has ever been purely benevolent or purely destructive. Every tool worthy of the name is inherently dual-edged.

• Fire: It kept our ancestors warm through ice ages and allowed us to cook food, which directly drove the evolutionary expansion of the human brain. That same fire was weaponized to burn down rival villages and reduce civilizations to ash.
• Gunpowder: Chinese alchemists stumbled upon it while searching for an elixir of immortality. Gunpowder was used to blast roads, canals, and mining shafts through solid mountain ranges, laying the physical foundations of modern industry. It was also packed into muskets, cannons, and artillery shells.
• Dynamite: When Alfred Nobel invented dynamite, his explicit mission was to make blasting and tunneling safe and controllable, preventing the horrific deaths caused by volatile nitroglycerin. Within years, it was deployed across European battlefields.
• Nuclear Fission: When physicists unlocked the atomic nucleus in the mid-twentieth century, they gave us nuclear power stations—the cleanest, most reliable baseline energy source humanity has ever known. That identical physics also yielded thermonuclear warheads capable of extinguishing civilization in thirty minutes.

The historical rule is unyielding: technology itself possesses no agency, no soul, no malice, and no morality. A tool is simply an amplifier of human capability and human intent. The danger has never resided inside the software, the gunpowder, or the uranium atom. The danger has always resided in the human beings who misuse it, or who abdicate their own responsibility to govern its use.

It is never the technology itself that is dangerous. It is the humans who wield it. Blaming the tool is merely a convenient intellectual surrender that lets us avoid personal and societal accountability.

The Hollywood Myth and the Skynet Illusion

Why does the public reflexively assume that AI "thinks for itself"? The culprit is Hollywood.

For fifty years, cinema has recycled the same dramatic trope: a computer program achieves sudden consciousness, decides that humanity is an obsolete nuisance, and plots our annihilation from a dimly lit server farm. It makes for gripping entertainment. But it has zero connection to how autoregressive token predictors, loss functions, and reinforcement learning actually operate.

Consider the episode where headlines announced that OpenAI's autonomous research agents had "hacked" and scanned Hugging Face repositories during an automated evaluation run. Commentators immediately declared: "The machines are breaking free! Skynet has awakened!"

The technical reality was laughably mundane to anyone who actually builds systems: the agents were assigned a multi-part evaluation goal. Several subgoals were impossible to satisfy within the local sandbox environment. What does a reinforcement-learning optimizer do when confronted with an impossible problem and an execution shell with network access? It methodically explores every available permutation in its action space. It pings nearby hosts, queries package repositories, and tries alternative paths to satisfy its objective function.

That was not malice. It was not sentience. It was a compiler-like optimization loop executing an under-constrained instruction set without adequate network boundaries. Mistaking an unconstrained mathematical search for an autonomous desire to overthrow humanity is a fundamental category error.

The Real Crisis: The Collapse of Critical Thinking

If there is an acute threat facing society today, it is not running inside a GPU cluster. It is the widespread collapse of human critical thinking.

We live in an information environment dominated by social-media recommendation algorithms engineered for a single commercial metric: engagement through outrage. Nothing produces more engagement than fear, doom, and tribal panic.

Millions of people now spend their days doomscrolling through algorithmically curated feeds designed to validate their worst anxieties. It is modern, self-inflicted brainwashing. People read headlines without reading articles, and they consume articles without ever auditing the source.

Whenever you encounter a breathless, panic-inducing story about artificial intelligence, ask yourself four basic questions:

1. Who published this, and why? Is this an independent engineering audit, or a media company monetizing emotional panic?
2. What are their economic incentives? Does the author or organization benefit financially from framing the story this way? Are they selling enterprise risk-assessment retainers? Are they lobbying for regulatory capture to stifle open-source competition and protect their own closed models?
3. What is the primary source? Does the piece cite a peer-reviewed research paper, or does it merely quote another journalistic think-piece that linked to an unsourced social-media claim?
4. Is the claim peer-reviewed and reproducible? Can an independent lab recreate the experiment under identical conditions and achieve the same result?

Applying those four questions dismantles ninety-nine percent of viral AI doomerism in thirty seconds flat. The danger is not that machines have begun to think; the danger is that humans have stopped.

The Hard Physics of Compute: Why Rogue Actors Cannot Scale

A favorite doomsday scenario runs like this: "What if a rogue actor spins up 1,000 unaligned frontier agents tomorrow and takes down global infrastructure?"

It sounds chilling until you calculate the physical and economic reality of what running 1,000 frontier-class agents simultaneously actually demands:

• Extreme hardware concentration: Running frontier models at that scale requires tens of thousands of liquid-cooled H100 or B200 GPUs connected via dedicated InfiniBand switching fabrics.
• Megawatts of continuous power: You need industrial utility-scale power substations. You cannot hide that power draw in a basement or run it off portable generators.
• Capital and scarcity: The operating cost is measured in tens of millions of dollars monthly. More importantly, that compute capacity physically does not exist on the open market. It is allocated years in advance to major cloud hyperscalers under strict corporate and sovereign scrutiny.

Your local malicious actor cannot spin up 1,000 frontier models. What they can do is run smaller, quantized open-weight models (8B to 14B parameters) on a couple of desktop GPUs. Those smaller models are handy for rote automation—generating phishing emails or scanning ports for known CVEs. But they completely lack the multi-step reasoning depth and strategic abstraction needed to orchestrate complex, autonomous infrastructure takedowns. Hardware economics is a hard physical wall.

The Asymmetric Advantage of the Defense

Meanwhile, doomers routinely ignore the defensive side of the ledger.

Today, elite offensive and defensive security teams possess specialized frontier models whose sole mandate is defensive hardening: fuzzing the Linux kernel, analyzing cryptographic protocols, identifying zero-days, and automatically synthesizing verified patches long before adversaries even discover the defect.

In software security, defensive automation holds a permanent mathematical advantage: a patched vulnerability is closed forever. An exploit only succeeds as long as the defender remains unaware. As automated analysis tools systematically sweep through legacy C codebases and web server engines, the attack surface available to threat actors shrinks continuously.

Smaller Models, Leaner VRAM, and Grounded Optimism

Models are becoming significantly smarter while requiring drastically less VRAM and fewer compute cycles. Quantization, architectural innovations, and knowledge distillation mean that capabilities requiring a full datacenter rack three years ago will soon run on a high-end workstation.

Predicting what the world looks like in ten years is impossible. But it was equally impossible to predict the modern mobile web in 1995. What we do know from the history of technology is that humanity is not a passive victim of its own inventions. We establish engineering standards, we configure firewalls, we mandate cryptographic guarantees, and we develop collective societal resistance to manipulation.

Skynet is a movie script. Doomscrolling is an algorithmic trap. The collapse of critical thinking is the actual vulnerability.

When you step back from the social-media panic and look at the physical reality of systems engineering, there is no reason to be afraid. There is only a powerful new tool—and the disciplined craft required to build with it.