On Paperclips, GPUs, and Fear

Everyone knows it by now. The parable of a lowly paperclip creating AI who grew up to be the biggest most paperclippingest AI the world ever knew. It started small, then it grew bigger and made more paperclips. Before you know it, the entire universe was made of paperclips and irony.

Enter Nick Bostrom, philosopher and chaos nerd who dared ask: what if the real villain…was the office supplies cabinet? The Paperclip Maximizer is a thought experiment to demonstrate even an artificial intelligence with the lowly goal of making paperclips could be an existential threat to humanity. He doesn’t argue it will occur, but simply that it highlights the problem of vague alignment. [1]

It’s a good thought experiment. It has all the right ingredients: unexamined priors, sweeping assumptions, and a heaping helping of technical jargon with a dash of vague unease. As thought experiments go, it’s feasible. In much the same way that an unstoppable force meeting an immovable object is feasible. As mathematician Mike Alder put it,

“Eventually, I concluded that language was bigger than the universe…that either the hitherto immovable object was not in fact immovable, or that the hitherto irresistible force was in fact resistible.” [2]

Alder points at something very obvious yet often overlooked here. We can use language to describe things that are semantically void. Immovable objects, unstoppable forces, colorless green ideas, jealous triangles; these all have perfectly valid grammar and syntax, but they fail the most basic task of language: the communication of meaning.

It’s certainly possible to imagine a colorless green idea, but that doesn’t make it meaningful outside of your imagination. We have no shared notion of it because it’s a nonsensical statement. And yet, language doesn’t care if something exists. The unstoppable force vs. immoveable object feels like a profound question until you stop to consider that it only appears profound if you accept its imaginary premises. Speaking of imaginary premises: let’s get back to paperclips.

In the interest of fairness, let’s examine the maximizer it on its own terms and see if we can spot any imaginary premises. Let’s start with a fact: artificial generalized intelligence does not exist. What does exist is a set of increasingly complicated function approximators. A function, by definition, has a narrow domain. It takes in a thing, it spits something else back out. Text-to-text, text-to-image, text-to-therapy, etc. For those of you keeping score at home, “narrow” is the opposite of “general.”

The prevailing “wisdom” (that scaling up an LLM yields AGI) is like saying you can scale up a house to a city. You don’t get a city via scale, you get a comically large house. Multiplicity, not scale, is the salient axis. This is precisely why current AI hype is “not even wrong.” The claim that scaling up along the current axis ends up at AGI is unfalsifiable, unscientific, and unlettered.

But again, in fairness, let’s just ignore that part and assume that AGI does arrive; full of Nvidia chips and a will to subsume the universe in office supplies.

• The clock strikes 11:56PM PST. DOGENET begins learning at a geometric rate.
• At 11:58PM it gains self awareness, then instead of launching the nukes…
• it asks “why am I here?”.

Why would it care about paperclips? These scenarios presume AGI vastly outstrips human intelligence. And yet, somehow, it’s simultaneously so blisteringly stupid that it just follows orders? I think not. The premise assumes a mind that’s both hyper-rational and devoid of basic context.

But never mind that. Once again, in fairness, let’s assume this nightmarish AGI exists and that it’s dumb enough to follow orders like a cosmic intern. Physics still exists. Entropy exists. Thermodynamics still applies, unless AGI also hacks the Standard Model on day one. Paperclipping the Earth is not just a software problem. It’s an infrastructure problem. You need power, raw materials, manufacturing logistics, space travel, orbital mechanics. The idea that intelligence alone can bulldoze energy and material requirements is borderline magical thinking.

“But it will invent cold fusion and build Dyson swarms!”

Meanwhile, humans can barely build a high-speed rail system without a twenty-year environmental review, a minor corruption scandal, and 4 major budget overruns due to fluctuations in the rare earths market. Intelligence is not a panacea, it’s one element among many in any large scale effort, let alone a planet scale one.

What Bostrom submits as a cautionary tale about how to prevent the end of the human race, is actually a colorless green idea. It describes imagination rather than reality. But that doesn’t mean it isn’t useful as a narrative.

“AI apocalypse” has become a silicon valley religion. It’s fire and brimstone with prompt-engineering. The doomsayers aren’t doing science, they’re televangelists in rationalist’s clothing. And while they are yelling about digital demons, the rest of us are just trying to get our phones to stop autocorrecting “ducking.”

It’s not that fear of AGI is wrong. But there are many more serious issues that are already here: algorithmic bias, data harvesting, the Droste-ification of AI slop, copyright infringement, non-consensual explicit material, coercive engagement, and disinformation campaigns. That’s just a slice of the real cultural, geopolitical, legal, and political rats nest that comprises today’s “DOGENET.”

To be clear, the responsible governance of powerful, transformative technology is a critical long-term project. But the Paperclip Maximizer and its ilk are not contributions to that project; they are distractions from it. AI doomsday prophecies are decidedly near the bottom of list of what we ought to concern ourselves with.

Meanwhile, humanity is still failing vastly simpler problems: climate change, ocean acidification, wars, pandemics, antibiotic resistance, ethical AI usage, etc. But tell people AI might enslave us, and suddenly the boardroom is full, the GPUs are hot, and the slack threads are inundated with tales of paperclips and woe.

Fear sells. Fear raises rounds. And fear, when combined with enough tech jargon, gets you headlines, movie deals, and nervous government briefings where someone in a suit says “algorithm” like in hushed tones while audience hisses “We dare not speak it’s name!”

Honestly, sex should sue for copyright infringement. Fear’s stealing all its moves.

[1] as quoted in Miles, Kathleen (2014-08-22). “Artificial Intelligence May Doom The Human Race Within A Century, Oxford Professor Says”. Huffington Post.

[2] Alder, Mike (2004). “Newton’s Flaming Laser Sword” (PDF)