* This text is translated from the original Japanese and may contain some unnatural expressions.

1. LLMs and Fluctuation

The rapid evolution of Large Language Models (LLMs) is bringing irreversible changes to the structure of "knowledge" production in society. An essential concept for grasping the essence of this change is the "convergence and divergence of fluctuation" within their responses.

As LLMs become more sophisticated, the fluctuation in their answers to routine inquiries (for example, fact-checking like "1+1=2" or basic logical progressions) rapidly converges, consistently outputting the identical optimal solution. Conversely, for inquiries that demand creativity or a deep understanding of complex contexts, they are increasingly generating a vast array of divergent responses.

This phenomenon intuitively represents the process of "optimization via probability calculation" inherent in current LLMs, serving as a vital premise for understanding today's AI behavior and prompt engineering. In a society where the cost of generating "results (outputs)" approaches zero due to the evolution of AI, this article will examine where the true, often overlooked essence of human labor's value lies, and explore its practical trajectory.

2. Output Supremacy and the Siege of Starvation

Many business professionals believe that highly specialized tasks, such as management strategy and programming, inherently have "no single correct answer and allow for diverse opinions (fluctuations)." However, the advancement of AI is fundamentally overturning this premise. Many of the tasks humans perceived as complex and lacking definitive answers were merely littered with "local optimums"—a consequence of the limitations of human cognitive and processing capacities (or the difficulty of outputting them at a consistent pace).

Ultra-high-performance AI instantaneously calculates countless variables, processing them as "mere computational problems." As a result, it discovers a single "optimal solution (a perfect generalization)," and the answers converge upon it. This law applies not only within the linguistic space but also extends to physical space. Highly advanced physical operations by autonomous driving or robots are ultimately completed as ultimate optimization tasks, as the fluctuations against real-world noise converge through computation.

What is crucial here is that the process by which AI derives a perfect answer is entirely different from human "comprehension of meaning" or "logical deduction." It is strictly nothing more than "probability calculation based on past data." Standing firmly on this premise allows us to calmly detach from the sci-fi and philosophical leap that AI will achieve a superintelligence possessing consciousness equal to or greater than that of humans.

However, even if the internal processes are entirely different, the cold reality in the real world stands: "If the output (result) is perfect, the practical task will be replaced." Historically, society has trusted and evaluated others based on the quality of their output. But today, as AI has reduced the cost of generating "perfect and polished outputs" to zero (commoditization), this system is breaking down. The increasing sophistication of scams, the hollowing out of evaluation in recruitment and education, and market confusion driven by superficial expertise are typical examples of system bugs spawned by "Output Supremacy."

The simplistic labeling casually tossed around in society, such as "Jobs that will survive the AI era," is a dangerous trap. There is an undeniable concern that such labels fail to fully price in the potential of AI's advancement. It means that even those roles will eventually become unnecessary or be reduced to an extreme minority. Regardless of the profession, workers whose primary purpose is to "output" beautiful pictures or texts will automatically find themselves participating in a game of musical chairs with steadily dwindling seats. Without a fundamental understanding and deep contemplation of AI, one will quite literally fall victim to a "siege of starvation" (a war of attrition).

3. Utility and Ripple Effects

While highly specialized professions face the threat of replacement due to the "convergence of fluctuation" driven by high-performance AI, simple frontline labor—such as factory work, delivery, and teller services—are often cited as the prime examples of jobs destined for heavy AI replacement. Here, I deliberately choose to focus on the latter. Are these tasks truly meaningless, simple labor?

Here, too, the concept of "output (result)" becomes paramount. If we evaluate labor solely by focusing on the "efficiency of the final output," similar to AI's computations, then these jobs could literally be deemed unnecessary tasks meant for machine replacement. However, when we step away from the perspective of "Output Supremacy" and attempt to deeply understand the role these operational processes play in the real world, we find a definitive "utility" that is consistently demonstrated and accumulated.

Yet, by its very nature, the "utility" mentioned here is difficult to demonstrate in real-time as an objective metric. It dissolves unconsciously into the daily operational processes or within the individuals involved. Its true value can only be objectively grasped retroactively, "only after a stable output has been established." Paradoxically, an optimization mindset that relies solely on "output (results)" will forever fail to acquire this fundamental utility.

Following this line of thought, many might imagine this utility to be human empathy, morality, or something akin to "unique human warmth." However, that is not necessarily the case. What is intended here is the "practical business value" required to smoothly drive an organization or the entire social system. The ultimate manifestation of this is the cascading "ripple effect (economic externality)" it has on its surroundings.

For example, imagine a "store with a bad atmosphere (A)." Customers visiting the store accumulate "sparks of dissatisfaction (stress)" due to cold interactions or ignored mistakes. After leaving, these customers vent that spark of dissatisfaction onto others at home or work, causing those on the receiving end to harbor new frustrations. At first glance, this might look like a mere cycle of human emotion, but from a practical, macro perspective, it is nothing less than "a process where an error (noise) generated at a single frontline location cascades throughout society, destroying productivity and relationships across various sectors."

Conversely, suppose a person harboring a "spark of dissatisfaction" somewhere in society visits a "store with a good atmosphere (B)," which delivers value beyond mere numbers. The staff at this store flexibly absorb the "noise"—the customer's bad mood or irregular requests that do not fit neatly into a manual (1+1=2)—on the spot, offering seamless service. As a result, the spark of dissatisfaction the customer carried is extinguished right there, allowing them to head to their next destination (home or work) in a calm state.

Looking at these two patterns, you might think, "Is it really just about that?" However, when we scale up and compare the surrounding area of A versus the surrounding area of B, or an area with many A-type organizations versus one with many B-type organizations, the impression shifts entirely. The volume and vibe of people passing through, the subsequent quality of public safety, the expansion or withdrawal of businesses capturing that data, and the resulting fluctuations in land prices and real estate valuations—these are precisely the ripple effects in action.

For an AI capable of calculating only the optimization of localized "output (numbers and processing speed)," computing such broad and abstract ripple effects is impossible. Pushing AI optimization solely dedicated to generating outputs and blindly eliminating the human process—the "epicenter of noise absorption and ripple effects"—from the frontlines could invite completely counterproductive outcomes.

The answer to the initial question, "Are these tasks truly meaningless, simple labor?" is clear. From a macro perspective, they are indispensable processes that reduce societal friction and generate invisible ripple effects. Furthermore, even from a micro perspective (that of the individual worker), undertaking a portion of the "entire process that AI cannot comprehend" can, depending on the individual's mindset, serve as a valuable exercise in cultivating their own foundational strength. Conversely, obsessing too much over output simply means voluntarily stepping into "territory that AI can 100% understand and replace."

4. The Paradox of AI Implementation

Having established the importance of human "processes" on the frontlines, what is actually happening in advanced workplaces where AI is currently being implemented? One might speculate that "implementing AI has improved the abstract elements of the workplace (atmosphere and human relationships)."

However, this is not because AI is autonomously generating a positive atmosphere. The true reason lies in the fact that AI is acting as an emotionless filter, absorbing and blocking "abstract human errors (emotional friction and negative factors)."

By taking over routine "1+1=2" tasks and acting as a hub for information transmission that usually entails emotional friction, AI drastically reduces human cognitive load. As a result, humans are afforded "margin" (mental breathing room), allowing them to enhance their tolerance, endurance, cognitive awareness, and judgment regarding problems. Particularly in societies like Japan, where the invisible costs associated with cultural backgrounds are high, AI functions as a highly advanced tool that significantly mitigates problems originating from them. I believe this makes a massive contribution to creating a flat, egalitarian environment.

We are currently in an ideal and supremely powerful hybrid state (a transitional phase) where AI handles perfect "outputs," while humans can utilize that freed-up margin to focus entirely on "anomaly detection."

However, this highly successful phase harbors the greatest trap. The vivid success story of "both numbers (output) and atmosphere (process) improving dramatically" will eventually trigger a fatal illusion among management and society at large: "Because AI is extremely excellent, all problems have been solved. Humans in noisy frontline environments are no longer necessary."

This illusion amplifies an "overconfidence in Output Supremacy," steering management toward decisions that sever the invisible infrastructure supporting the organization. Evaluating only the AI's output while dismissing the processes that absorbed noise behind the scenes will ultimately pull the trigger on a "siege by starvation" that renders the entire organization unsustainable. We are facing a paradox where, the more perfect the output becomes, the more rigorously we must evaluate the existential significance of the "human processes" that anchor that output.

5. Controllability in the AI Era

In an AI society where "outputs" perfectly converge, what is the true role that humans must play? The general public often celebrates the skill to "input clever prompts (instructions) to extract beautiful outputs from AI" as the pinnacle of mastering AI. However, in a future where the output itself is commoditized (rendered valueless), such superficial tricks will hold zero competitive advantage.

In the truest sense, "mastering AI" means transitioning from being an outputter to becoming a "responsible owner/editor." It means treating the infinite local optimums and perfectly curated generalizations calculated by AI via probability as a "mere starting line," and exercising maximum human sensory discretion through trial and error. The generated result is not the direct finish line. In creative tasks, out of the sea of ideas endlessly scattered by AI, humans must exert a high degree of controllability to curate—through human emotion and context—which combinations will generate genuinely new value.

Imagine there are 100 math problems in front of you. If the goal is to obtain "accurate results (outputs) to report at tomorrow's management meeting," the indisputably correct answer is to leave it to AI. However, if the goal is "educating a child" or "personal training," one must ignore efficiency and intentionally struggle through the problems with pen and paper. This is because the value lies within the "friction-laden process" itself.

Only those who possess the metacognitive ability to discern "whether the output is needed or the process is needed," and who can make the highly advanced decision to reject the temptation of efficiency by "intentionally choosing an inconvenient process without using AI," will be able to evade the siege of starvation brought on by Output Supremacy.

A person who continually abandons the process will eventually lose the ability to intuitively detect anomalies (by gut feeling) when AI outputs hyper-advanced management strategies or plans. They will be unable to sense whether the numbers and logic are valid, or if they contain fatal hallucinations (plausible lies) or sycophancy (excessive flattery). Valuing the process is an indispensable act for cultivating the "intuitive baseline (foundational strength)" necessary to control AI's outputs.

When witnessing AI's perfect output (the pinnacle of generalization), humans often anthropomorphize it and mistakenly assume that "AI understands meaning through an advanced process (it has reached superintelligence)." Escaping into this sci-fi and philosophical interpretation is dangerous. It averts our eyes from the impending crisis of "declining demand for humans reliant solely on output" and hinders the actions required to shift ourselves over to the side that shoulders the process and responsibility.

6. Barriers to Societal Implementation

Up to this point, we have discussed the collapse of Output Supremacy and the value of human "processes." However, implementing this into actual societal systems presents several objective barriers (challenges).

The first barrier is the "messiness of the transitional period (time lag)." Logically, a future where "the value of output drops to zero" is inevitable, but complex legal regulations, existing business customs, and organizational structures with outdated evaluation criteria remain deeply entrenched in the real world. A time lag of 10 to 20 years may occur before this gap between theory and reality is bridged. The perspective of how to survive this "muddy transitional phase where old systems and new AI infrastructure coexist" is an extremely practical challenge.

The second barrier is the "risk of the collapse of AI's 'perfect output' itself." This article proceeds on the assumption that AI's output will become a flawless infrastructure. However, current AI research raises concerns about a phenomenon called "Model Collapse," where the model degrades as AI continues to train on data generated by AI itself. If the AI meant to serve as infrastructure becomes contaminated, the next generation of responsible leaders will be required to formulate risk hedges on how to address it.

The third and most significant barrier is the "absence of new social metrics to evaluate 'process'." Modern capitalist society is optimized by quantitative systems that measure "output," such as sales, follower counts, and test scores. Even if the "human process (the cultivated utility and anomaly detection abilities)" is vital in the AI era, how do we quantitatively evaluate it and convert it into compensation such as a salary? The "design of the evaluation system" is precisely the most difficult theme we must solve next.