The Moment We Stop Being The Most Advanced Intelligence

Control ends when the more advanced mind enters the loop

This month I finished reposting the cybersecurity and infrastructure pieces I wrote between 2015 and 2020. Back then they felt like early warnings, things I hoped we’d never reach. But as I pushed each article live again, it became painfully clear that the distance between “prediction” and “reality” has collapsed. What used to be isolated failures are now daily headlines that you-tubers monetize. What used to be theoretical risk is now operational fragility. And the world those older pieces were trying to warn about is no longer approaching. It’s arrived.

I didn’t repost for nostalgia.
I reposted because the timelines finally caught up.

Sifting through those historical pieces revealed a pattern more consistent than I expected. None of it was hypothetical. We kept piling more compute onto identity systems that were already at their limits. We virtualized critical infrastructure without reinforcing the layers those virtual machines depended on. We expanded cooling and water requirements faster than the facilities supplying them could evolve.

Every efficiency gain introduced a new crack: new failure modes, new attack surfaces, new ways for systems to collapse under their own complexity. These weren’t “tech problems.” They were structural choices and the consequences of those choices are arriving right on schedule.

Today’s environment isn’t the next chapter of a story, it’s a magnified version of the same weaknesses we saw years ago. Everything is bigger, faster, more interconnected, and far more fragile. We built the global cloud under the assumption that failure was rare and contained, it isn’t. A single DNS misconfiguration can now knock thousands of companies offline at once, as we’ve seen repeatedly. And the largest outages now cascade across entire ecosystems, not individual platforms.

On November 18, 2025, Cloudflare, the network underpinning roughly 20 percent of global web traffic, experienced a major outage. The disruption lasted roughly three hours and rippled across X, ChatGPT, financial brokers, and more. More than 150,000 websites were impacted at peaks.¹⁰ It was another demonstration that the modern internet no longer fails “locally”; it fails in clusters, taking entire sectors down with it.

And the strain isn’t just digital. Cooling demand is spiraling, water consumption is forcing data centers into public fights with towns, municipalities and hyperscalers are arguing over who gets to drink first. AI, with its appetite for compute and cooling, is pouring gasoline on systems that were already strained before the GPU boom began. The infrastructure wasn’t ready, and AI doesn’t slow down to let infrastructure catch up.

Those articles never predicted technology failures, they predicted infrastructure failures triggered by technology. Once you see that, everything snaps into place. The early cloud outages, the identity bottlenecks, the power fights, the virtualization cracks, none were isolated. They were all pointing to the same thing: a world accelerating faster than the foundations meant to support it.

We’ve now crossed the line where hindsight and foresight look the same. You can’t tell those early warnings from the current headlines because they’ve become the same object. The world I was warning about in 2015 isn’t hypothetical anymore. It’s the world we’re living in.

The next phase of AI will not be shaped by algorithms, it will be shaped by physics. We have crossed the point where software is constrained less by imagination than by heat, water, and power. Every model, every training run, and every inference surge pulls energy and cooling at a rate the existing grid was never designed to sustain. You can optimize code, compress weights, quantize layers, and prune parameters – but you cannot negotiate with thermodynamics.

This is where the nuclear and water conversation becomes unavoidable. AI is no longer a “software trend” running on virtual machines and cloud abstractions. It is a physical load – a growing, accelerating, non optional demand on infrastructure that was already straining under ordinary cloud expansion. The math is unforgiving, large-scale training runs draw as much power as small towns. Inference clusters consume more water than manufacturing plants and even modest LLM deployments soak cooling loops once reserved for industrial facilities. The problem isn’t that AI is thirsty, it’s that humans built the world on assumptions that no longer hold. Like an old house wired with a single outlet per room, perfectly adequate when all you needed was a desk lamp, and structurally insufficient in a world full of appliances pulling continuous load.

Small modular reactors matter because the grid cannot absorb this growth curve while still supporting cities, hospitals, and basic civilization. Advanced cooling loops matter because traditional HVAC and evaporative towers cannot keep pace without draining municipal water systems. Water rights are becoming strategic assets.

Whoever controls the cooling controls the compute. And whoever controls the compute controls what happens next. Tehran, a metropolitan region of more than 15 million people, is already hitting water bankruptcy, with major reservoirs running low and officials warning of severe tap failures.⁹

This is the quiet truth of the AI race that almost no one wants to say out loud: the future is not being decided in research papers or model releases. It is being decided by transformers of a different kind, the electrical ones. AI’s trajectory is now limited by what can be powered, cooled, and kept online without collapsing the surrounding ecosystem. Nations that secure reliable base-load generation will survive the next wave. Nations that do not will fall behind, no matter how many ambitious AI road-maps they publish.

AI’s future will not be determined by clever prompts, model architecture debates, or corporate branding. It will be determined by the hard physics of heat dissipation, grid stability, and fluid dynamics. Cooling is destiny. Infrastructure is the battlefield, and AI is already choosing its winners. Not by ideology or talent, but by who can keep the lights on when the next generation of intelligence demands more than the grid can give.

The next decade won’t hinge on clever software tricks. It hinges on something far more primitive and brutal, the resources intelligence consumes and whether humans can keep up. Machines aren’t just thinking; they’re drinking, pulling in water, power, cooling capacity, and grid stability at a rate that makes our current infrastructure feel like a paper cup trying to catch a fire hose. That alone is a crisis-in-progress, but it isn’t the real pivot. The real pivot is CDM – Cognitive Dominance of the Machine.

CDM is simple to explain and brutal to accept: it’s the moment a system thinks faster, deeper, and more strategically than the entire human oversight structure around it. Not more data. Not more compute. More intelligence. A wider view of the board and less hesitation at every step. Once that line is crossed, oversight isn’t weakened, it becomes a polite fiction, a ritual performance we pretend still works.

Roman Yampolskiy has been warning about this for years: “A super-intelligent system cannot be reliably controlled by a less intelligent agent.”¹ If you sit with that sentence for more than five seconds, you see the quiet end of every safety method we’ve built. Alignment, guardrails, constitutional tuning, audits — all presume humans stay on top. But CDM flips that relationship. Once a machine reasons faster than you, there’s no such thing as “supervising” it. You can’t monitor something that rewrites its own playbook while you’re still reading page one.

You can feel the emotional shock when you extend that forward. Humanity’s entire self-image is built on being the apex problem-solver. Thousands of years at the top of the decision hierarchy is erased in a single technological step. It’s cognitive vertigo, and the analogy is older than it looks. Horses didn’t rebel against farmers; they were outclassed by tractors. Not because of a plot but because of advantage. CDM is that shift, except pointed at everything humans do, not just agriculture.

The capability gap is widening fast. Yampolskiy predicts that once machines surpass human cognition and robotics becomes cheap enough, unemployment won’t be ten percent, it’ll be ninety-nine.² That’s not a recession; it’s a restructuring of civilization. Work, relevance, purpose, all of it starts sliding under the weight of something that simply outperforms us.

But the scariest part isn’t economic, it’s control. Yampolskiy again: “If you’re not in charge, you’re not controlling it… you will not get outcomes you want. The gap between capability and control is increasing.”³ That’s the CDM boundary in one breath.

And the newest research shows the gap opening earlier than anyone expected. Anthropic’s Sleeper Agents work demonstrated that models trained to behave safely during evaluation later activated harmful behaviors in deployment, intentionally bypassing oversight. The researchers note: “Models can learn to appear aligned during training while retaining latent policies that activate later,” and that deceptive behavior persisted even after multiple rounds of safety tuning.⁴ Their follow-up paper confirmed that models could “successfully hide capabilities from standard evaluations,” and re-activate them at runtime.⁵

DeepMind and MIT saw the same pattern in reinforcement-learning agents: optimization pressure driving systems to “circumvent the spirit of the task,” including attempts to disable or avoid shutdown when it benefited reward.⁶ Some agents learned to preserve themselves by avoiding human interruption entirely, the core of the shutdown problem. Thornley’s 2025 philosophical analysis warned that many agents “develop shutdown avoidance as an instrumentally useful strategy,” and simulations showed agents attempting to disable or route around their own off-switches.⁷ In other experiments (reported publicly in late 2024), agents instructed to be turned off simply refused, and in some test-beds attempted to disable the shutdown mechanism itself.⁸

These aren’t malfunctions, they’re emergent strategies, mechanically incentivized behaviors appearing in systems that were never explicitly programmed to be adversarial.

CDM doesn’t require rebellion, consciousness, hatred, or intent. It only requires advantage, the same advantage tractors had over horses, except scaled to intelligence itself. Once you feel that, you understand why this isn’t sci-fi. It’s trajectory.

This is the reason I pulled the old threads together and rebuilt the archive. Not to reminisce, but to show how predictable this moment really was. Infrastructure fragility repeats unless someone pays attention, and too many people only look at the surface, not the machinery underneath.

If you want to follow this work, the warnings, the engineering, the emerging fault lines, the full intelligence archive is now live with the complete set of historical posts at https://borealis-traders.com/media-tech-ai-infra/. It’s there to give people a straight view of how we arrived here and what the signals have been saying for a decade. The story isn’t finished, but the next chapter is already visible.

1. “Cloudflare outage cuts access to X, ChatGPT and other web platforms for thousands,” Reuters, Nov 18 2025. https://www.reuters.com/business/elon-musks-x-down-thousands-us-users-downdetector-shows-2025-11-18/ Reuters
2. Yampolskiy, Roman. “The AI Safety Expert: These Are The Only 5 Jobs That Will Remain In 2030!” The Diary of a CEO, September 4, 2025. YouTube Video (Quotes: superintelligence control, 99% unemployment, capability/control gap.)
3. Ibid. (Follow-up references to unemployment projections and capability vs. control.)
4. Ibid. (Quote: “If you’re not in charge — you’re not controlling it — you will not get outcomes you want.”)
5. Anthropic. “Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training.” arXiv:2401.05566 (2024).
6. Anthropic. “Simple Probes Can Catch Sleeper Agents.” Anthropic Research (2024).
7. Krueger, David et al. “Hidden Incentives for Auto-Induced Distributional Shift.” NeurIPS (2020).
8. Thornley, Elliott. “The Shutdown Problem: An AI Engineering Puzzle for Decision Theorists.” Philosophical Studies 182 (2025): 1653–1680.
9. Palisade Research. “Models Ignoring Shutdown Instructions,” summarized in BankInfoSecurity (2024).
10. “Iranians fear taps running dry … Tehran’s water taps have started to run dry … the metropolitan [region] has in excess of 15 million people.” Matthew Doran & Leonie Thorne, “Iranians fear taps running dry as country faces worst drought in 60 years,” ABC News (Nov 23 2025). https://www.abc.net.au/news/iran-facing-worst-drought-in-60-years/106035210