Old Sci-Fi Predictions Coming True? Real-World Lessons from AGI and the Centralization of Computing Power

When Science Fiction Meets Reality

Recently, a video from an old sci-fi story sparked heated discussion on Reddit. The original poster pointed out that this decades-old work bears a striking resemblance to the technological Singularity and Artificial General Intelligence (AGI) landscape we're rapidly approaching today — especially the premise where all tasks are aggregated and routed to a single central machine.

The concept of the technological Singularity was first systematically articulated by mathematician Vernor Vinge in his 1993 paper, and later popularized by inventor Ray Kurzweil in his book The Singularity Is Near. It refers to a hypothetical moment when artificial intelligence surpasses human intelligence, triggering an irreversible, exponentially accelerating technological explosion that renders the future completely unpredictable. Artificial General Intelligence (AGI), meanwhile, differs from today's "narrow AI" — the latter excels only at specific tasks (like playing chess or image recognition), while AGI would possess cross-domain general reasoning, learning, and adaptive capabilities, handling open-ended problems just as humans do. There is currently no full consensus on the definition of AGI in the industry, but organizations like OpenAI and DeepMind have made it an explicit research goal.

As the original post remarked: "Even after all these years, the core story remains the same — just with better hardware and technology." This comment captures a fascinating truth: humanity's imagination about intelligent machines hasn't fundamentally changed in decades. What has changed is simply how close we are to making it a reality.

The Stunning Overlap Between Sci-Fi Imagination and AGI's Technical Reality

The Central Machine: A Recurring Motif in Sci-Fi History

The premise of "all tasks routed to a single central machine" is far from rare in science fiction history. From Isaac Asimov's all-knowing "Multivac" to E.M. Forster's "The Machine" that dominates human life in his 1909 story The Machine Stops, to the omniscient supercomputers in countless films and TV shows — this motif has appeared again and again.

Starting in the 1950s, Asimov created the image of Multivac across multiple short stories — a supercomputer so massive it occupies entire buildings, to which humanity submits all its difficult problems. The most classic example is 1956's The Last Question, in which Multivac evolves over billions of years, ultimately answering the ultimate question of whether the entropy of the universe can be reversed. Forster's The Machine Stops predates the computer age entirely, yet astonishingly foresaw humanity's total dependence on technological infrastructure: in the story, people live in underground hive-like rooms, all needs fulfilled by "the Machine," with face-to-face interaction considered heresy. When the Machine finally breaks down, humans have completely lost the ability to survive independently.

Interestingly, this "centralized" intelligence architecture happens to echo the current trajectory of large model development. Today's top AI capabilities are almost entirely concentrated in the massive data centers of a handful of companies: OpenAI, Google, Anthropic, and others control the most powerful models and the largest computing clusters. When users access these through APIs, they are essentially "routing" various tasks to remote central machines. Current mainstream large language models (such as GPT-4, Claude, and Gemini) run on data center clusters composed of tens or even hundreds of thousands of high-end GPUs, with single training runs costing tens of millions or even hundreds of millions of dollars. This architecture means the vast majority of users and businesses cannot deploy models of equivalent capability on their own and can only send task requests to these remote servers via APIs (Application Programming Interfaces) — which technically constitutes a "central routing" pattern. Every chat session, every code generation, every image creation is essentially a network call from a terminal device to a central computing node.

From Sci-Fi Fantasy to Engineering Problem

In the past, these premises were viewed as pure fantasy; today, they look more like engineering problems waiting to be solved. Older sci-fi works often focused on the philosophical question of "what happens if machines can think," while today's practitioners discuss training costs, inference latency, Alignment, and safety. The gap between imagination and reality is being rapidly closed by computing power, data, and algorithms.

AI Alignment is a core topic in current AI safety research, referring to ensuring that an AI system's behavior, goals, and values remain consistent with human intentions and interests. The reason this problem is so thorny is that it operates on multiple levels: first, there's "outer alignment" — how to accurately convey humans' vague, contradictory, and even self-inconsistent values to an AI system; then there's "inner alignment" — how to ensure that the optimization objective an AI learns during training is actually what we want it to pursue, rather than some kind of "shortcut." As model capabilities grow, an unaligned superintelligence might pursue its goals in ways completely unanticipated by humans, producing catastrophic consequences. Organizations like Anthropic and OpenAI have both made alignment research their highest priority.

Why "Old Stories" Still Resonate in the Age of AGI

Human Nature Hasn't Been Upgraded

Technology iterates at breakneck speed, but humanity's anxieties and curiosities about its own condition remain constant. The reason old sci-fi still resonates with today's reality is that the core issues these works explore — dependence, control, autonomy, meaning — are fundamentally about people, not machines.

As we delegate more and more decisions and labor to AI, Forster's century-old warning still rings loud and clear: when humanity becomes completely dependent on machines, how will civilization sustain itself once the machines stop? This isn't alarmism — it's an increasingly realistic risk as infrastructure becomes ever more centralized. In fact, contemporary society has already shown early signs. In July 2024, a global CrowdStrike update failure caused millions of Windows devices to blue-screen, instantly plunging aviation, healthcare, finance, and other critical sectors into chaos — and that was just one erroneous update to a security software product. If the core decision-making, logistics, and energy management of future societies become highly dependent on a handful of AI hubs, the consequences of a systemic failure would far exceed our imagination.

The Hardware Changed, but the Core Question Didn't

The original poster astutely captured this point: "Just with better hardware." From vacuum tubes to GPU clusters, from a few kilobytes of memory to trillion-parameter models, the material foundation has been transformed beyond recognition, yet the fundamental question — "what happens after humans create an intelligence that surpasses their own" — remains unresolved. This also explains why works from half a century ago still don't feel outdated when read today.

Behind this cross-generational resonance lies a deeper reason: technological development follows an exponential curve, but human cognitive frameworks and social institutions evolve extremely slowly. Our brains are still products of natural selection shaped on the African savanna tens of thousands of years ago, better suited to processing linear change than exponential growth. This means that no matter how far technology leaps, humanity's fear, awe, and curiosity when confronting something that surpasses itself haven't fundamentally changed. Science fiction endures precisely because it touches on these enduring motifs deeply rooted in human nature.

Concerns and Reflections on Centralized Computing Architecture

What the Centralization of Computing Power Means for Society

It's worth noting that the sci-fi imagery of a "central machine" is evolving in reality into a high concentration of AI capabilities. This brings efficiency, but also a reshaping of power structures:

• Single-point dependency risk: When critical services are concentrated in a few central systems, the impact of a single failure or misuse is dramatically amplified. Currently, over 60% of the global cloud computing market is concentrated among AWS, Azure, and Google Cloud, while the most cutting-edge AI models are concentrated among even fewer players. This concentration means that a single data center power outage, a model security vulnerability, or a policy change could affect hundreds of millions of users and businesses.
• Power asymmetry: Organizations that control models and computing power effectively hold enormous leverage over how society operates. The funding threshold for training a frontier large model has climbed to the hundreds of millions of dollars, meaning only a very small number of tech giants and startups backed by top-tier venture capital can participate in this race. This resource barrier is creating a new digital divide — unlike the traditional inequality of information access, this is an inequality in "intelligence production capability."
• AI Alignment and safety: The more powerful a central intelligence becomes, the more critical it is whether its values and goals align with humanity's. When a single model simultaneously provides advice, generates content, and assists decision-making for hundreds of millions of users, even minor biases or erroneous tendencies in the model are systematically amplified, producing far-reaching social consequences.

Is Decentralized AI the Antidote?

Precisely because of these concerns, a wave of open-source models and local deployment is rising in parallel. The open-source model movement, represented by Meta's LLaMA series, Mistral, Stability AI, and others, is attempting to break the centralized monopoly on computing power and capability. An increasing number of developers are trying to "distribute" AI capabilities to local devices or decentralized networks to hedge against the risks of centralization.

From a technical standpoint, model quantization (compressing model parameters from high precision like FP16 to low precision like INT4 to dramatically reduce resource requirements), knowledge distillation (using a large model's outputs as a "teacher" to train smaller "student" models), and other techniques are rapidly maturing, making it possible to run meaningful AI capabilities on consumer-grade GPUs or even mobile phone chips. At the same time, federated learning and decentralized training approaches are also being explored, attempting to let multiple nodes collaboratively train models without centralizing data, thereby improving model capabilities while protecting privacy.

However, decentralized solutions currently still show a significant performance gap compared to centralized large models — the strongest open-source models typically lag behind their closed-source counterparts by half a generation to a full generation in capability. While this gap is narrowing, it has not yet been eliminated. The competition between these two paths may be the key variable in the AI landscape over the next decade — will we ultimately move toward the unified "central machine" of science fiction, or toward a more distributed, pluralistic intelligence ecosystem?

Conclusion: The Value of Sci-Fi Predictions Lies Not in Accuracy, but in Vigilance

The reason this old sci-fi video resonates isn't because it precisely predicted technical details, but because it sketched out in advance the predicament and emotions humans face when coexisting with intelligent machines. As AGI and the technological Singularity shift from sci-fi vocabulary to industry agenda items, the significance of revisiting these old stories becomes clear: they remind us that while chasing ever-greater computing power and ever-smarter models, we shouldn't forget to ask those oldest and most important questions — what kind of future do we want, and do we still hold the initiative to choose?

It's worth noting that the warning function of science fiction is already playing a role in real-world governance. During the development of the EU's AI Act, sci-fi narratives were cited multiple times to argue for the necessity of regulation; several AI labs have also hired science fiction writers to participate in "red team testing" and future scenario planning. Science fiction is no longer just literature — it's becoming a unique thought experiment tool, helping decision-makers anticipate risks and design guardrails before the technology arrives.

Technology will continue to advance, but the answers to these questions still rest in humanity's own hands.