Skip to content
realyyz's blog
Click here to change language
Go back

I was excited to write programs. Still am. And will be.

Edit page

Abstract

This article isn’t about “programmers will never be replaced.” It’s about why programming matters—not primarily as a high-income profession, but because it gave individuals, for the first time, the ability to converse directly with general-purpose computing power.

The information revolution differs from the steam and electricity revolutions. Steam and electricity mainly amplified physical strength, energy, and external production capacity; the PC and software put a universal cognitive machine into individual hands, allowing ordinary people to sit in front of their own machine and turn an idea into something that runs. The leverage programmers have enjoyed in recent decades comes from this “power to translate human intent into machine language.”

LLMs are lowering that barrier. They enable natural language to serve as an interface for operating computers, giving more people access to some of the leverage that was once exclusive to programmers. But that does not mean current LLMs are AGI. Even if they pass certain Turing tests convincingly, they still rely heavily on human experts, capital, and massive computing power to advance their capabilities. They have not yet formed a closed loop of self-training, self-play, and self-improvement.

What truly deserves attention is this: the AI infrastructure may become even more centralized. The PC era put computing power on personal desks; the AI era concentrates the strongest models, chips, data centers, and platform gateways in the hands of a few companies and wealthy individuals. This is both exciting and disappointing: ordinary people gain more powerful tools, but may also enter a more subtle form of platform dependency.

From a secular, naturalistic standpoint, strong AI will eventually emerge. Human intelligence and consciousness arise from the brain as a physical organ. Since a material system can produce intelligence and consciousness, there is no principled reason to assert that machines can never surpass the human brain or that machines can never become conscious. Once strong AI appears, its significance will go beyond “taking jobs,” for it may become the first invention by humanity that is “better at inventing than we are.”

But even when strong AI arrives, even if writing programs loses its economic value, I will continue to write. Just as humans still play Go despite being outmatched by AI, and still run despite being slower than cars, the value of programming is not only in outdoing machines. It represents an attitude: understanding the world’s rules through one’s own reasoning, appreciating the elegance of algorithms. Many people chose to become programmers not for money, but because programming itself is fascinating.

Hence the title: I was excited to write programs. Still am. And will be.

Programming is currently the only way humanity can directly converse with general-purpose computing power.

This statement sounds like a technical slogan, but behind it lies a long history. Economic history typically classifies steam, electricity, and information technology as quintessential general-purpose technologies: they don’t just improve one industry, but simultaneously transform family life, enterprise production, organizational structure, and the way society operates. The steam engine reconfigured factories and transportation; electricity revolutionized lighting, communication, and manufacturing; information technology brought computing, memory, expression, collaboration, and automation capabilities into personal life. These three technologies respectively drove three technological revolutions, but the information revolution has an unprecedented feature: it was the first to massively distribute control of a general-purpose technology to individuals, and that technology extended the power of the human brain.

The uniqueness of the information revolution: individuals, for the first time, possess a universal cognitive machine

The Industrial Revolution and the Electricity Revolution certainly changed ordinary lives, but most people were swept in as workers, consumers, residents, passengers. A person could use electric lights, ride trains, work in factories, yet could hardly modify the power system, railway system, or industrial assembly line directly. The PC era was different. Microprocessors and personal computers placed computing power that once belonged to institutions, labs, and large enterprises onto personal desks. Writing, spreadsheets, programs, design—these abilities became, for the first time, things that individuals could directly manipulate. The PC was the first machine that enhanced personal cognition.

The most fundamental difference between the information revolution and the previous two revolutions is not that it is “bigger,” but that it changed the interface between individuals and the world. Steam and electricity mainly amplified physical strength, energy, and external production capacity; information technology amplifies cognitive, computational, and abstract abilities. The first two revolutions had machines doing heavy lifting, lighting things, transporting things; the information revolution had machines begin to process spreadsheets, logic, communication, search, and memory for people.

Its second uniqueness is programmability. Electricity is powerful, but a person cannot, with a few lines of instructions, make the household circuits become an accounting tool one day, a typesetting tool the next, and a game console the day after. A computer can. The same machine, with just a different layer of software, can turn into entirely different tools. The core of the PC is not merely computing power, but that “a universal machine can continually morph through software.” This also explains why programmers have been so fortunate in recent decades: they stand on a chain of replicable, composable, and automatable universal tools. Expanding physical machines often requires factories, wiring, supply chains; the cost of copying a program is near zero. One exceptional programmer creates more value than ten thousand mediocre ones—this statement is hardly an exaggeration. Such leverage is unprecedented in the history of human technology.

The hacker spirit is not exclusive to programmers

Those truly drawn to programming are often attracted not by income but by the act of “personally making things change.” This temperament is commonly called the hacker spirit.

The hacker spirit does not equal intrusion, nor does it equal the performance of wearing a hoodie and writing code. It is more like an approach to the world: seeing a system, not settling for “it works,” but wanting to know why it runs that way, what feels awkward, and whether it can be made simpler. Such people are usually reluctant to stop at commenting and complaining. They will get their hands on, will trial and error, will break a vague problem into manageable smaller problems.

A true hacker thinks: a closed door is a provocation, a locked door is an insult.

These people did not appear only in the computer age. In eras without computers, they might have been watchmakers, locksmiths, instrument makers, printers, shipbuilders. Watchmakers adjusted gears and tolerances; today programmers adjust state and bugs. Locksmiths studied structural weaknesses; today security researchers study system vulnerabilities. The objects have changed, the way of thinking is very similar.

In the Industrial Revolution, they might have been mechanics, machinists, improvers on the shop floor. They might not have had formal authority, but they were close to real systems, knew why machines jammed, and knew how to make them run smoother. In the electrical and communications era, they might have become telegraph operators, radio amateurs, electronic repairmen. Wherever there is a complex system, such people emerge.

So the hacker spirit did not grow out of software. It existed first in mechanical systems, later migrated to communications systems, and finally entered computer systems. Its core is not a professional title, but a relationship between people and the world: ordinary people see the world as a set of ready-made rules; these people see the world as a system that can be modified.

Current LLMs have passed the Turing test, but are not yet AGI

Current LLMs already perform sufficiently human-like in certain specific Turing tests, even fooling a considerable number of testers into judging them as human. This is certainly an important signal. The natural language interface has matured to a certain threshold; machines no longer just execute commands but can participate in conversations in a near-human way. It can explain problems, write code, mimic tone. Recent tools like claude code and codex behave like reliable assistants in many tasks, capable of working continuously for hours without error. In the context of computer history, this is no small thing.

But this is still not AGI. The reason is not that it “doesn’t speak well enough.” On the contrary, today’s LLMs are already very good at speaking. The real gap lies in this: its capability growth mechanism remains highly dependent on external engineering systems, rather than forming a stable, autonomous self-iterating loop. The training cost of cutting-edge LLMs is extremely high, the training cycles very long; data, computing power, evaluation, alignment, and deployment all require massive organizational engineering. A leap in capability is usually not the model improving itself day by day, but the result of many researchers, GPU clusters, and capital pushing together.

Current LLMs can help people write code, read papers, organize materials, and can generate some experimental ideas. But they are usually not yet the agent that proposes its own roadmap, trains the next generation, verifies results itself, and decides to continue iterating. The so-called “AI self-evolution,” if it’s just models generating text that is then selected, trained, and deployed by humans, is still human-led iteration, not model-led iteration.

Moreover, large companies find it very difficult to truly let LLMs self-evolve, self-train, and self-play. It’s not that they don’t see the allure of this path, but rather that the bigger the organization, the more it fears losing control. Safety reviews, legal liability, brand risk, regulatory pressure, business rhythm, departmental KPIs, shareholder expectations—all lock the model into an evaluable, approvable, rollback-able track. Big companies can talk about agents, autonomous research, and self-improvement, but when it comes to implementation, it’s often autonomy within fences, evolution within sandboxes, innovation within approval workflows. They inherently suffer from the “big company disease”: they want breakthroughs but fear that breakthroughs are truly uncontrollable; they want the strongest models but dare not truly hand over the keys. For example, localized agent projects like OpenClaw did not emerge first from big companies but from independent developers and open-source communities. Not because the geniuses in big companies aren’t smart, but because big companies are hard to steer.

Therefore, truly self-iterating AI may not necessarily be born in the most prestigious labs. It’s hard to say. The first self-iterating systems that truly make people feel dangerous or shocked might not be behemoths built by a giant with hundreds of billions of dollars, but a system prototype trained, compressed, and self-played by some crazy hacker on their own Mac. It might not be large, not necessarily the prettiest, and unlikely to be compliant; but it could be freer, wilder, less constrained by organizational inertia. Historically, many things that truly changed the direction of technology initially grew not from the safest places but from the hands of some restless individual.

Thus, LLMs passing some version of the Turing test shows that machines can cross a psychological threshold in language interaction; but they are not yet AGI because they have not truly mastered the loop of continuously improving their own capability boundaries. They have already greatly enhanced human leverage, but have not yet become the intelligent agent that can independently rewrite its own destiny.

Programmers’ leverage comes from “translation power,” and LLMs are lowering that barrier

Programmers have held a special position over the past few decades, essentially because they stood between human intent and machine capability, wielding a kind of translation power. A person who wanted a website, reports, database, or automated workflow had to translate their intent into a form the machine could understand. The programmer was that translator. They didn’t just write code; they turned vague wishes into clear structures, turned one-time labor into repeatable processes, turned personal ideas into software systems that could be copied and distributed.

This position is not unfamiliar historically; only the intermediary changes with the era. When writing first appeared, literate scribes stood between people and the state machinery; in the Middle Ages, clergy who mastered Latin stood between knowledge and ordinary people; after the printing press, publishers and journalists stood between information flow and the public; during the Industrial Revolution, engineers stood between human needs and mechanical systems. In each era, those who truly gained leverage were often not those who merely used tools, but those who could directly modify, define, and reconfigure tools. Programmers are simply the group that, in recent decades, has been closest to the interface of the programmable world.

LLMs are changing this. The most important thing they do is not simply “helping people write code,” but turning natural language into a new high-level interface. In the past, to get a computer to do something, a person had to learn a strict formal language, i.e., a programming language. Now, a person can first describe the goal in natural language, have the model generate a first draft, explain the program logic, and then judge and integrate, continuing to collaborate with the LLM iteratively for the next iteration. In other words, LLMs have not immediately eliminated programming, but have reduced the friction of translation between humans and computers.

This will redistribute programmers’ leverage. In the past, “knowing how to code” itself was a scarce ability, because it was the ticket to the programmable world. As LLMs improve, natural language will hand that ticket to more people. Just as when literacy rose from extremely low levels to near universal, “being able to read and write” was no longer a privilege of the few; when more and more people can drive machines with natural language, “knowing how to write routine code” will also become less scarce. Scarcity will shift to higher levels: posing problems, system design, taste, boundary awareness, understanding of real-world scenarios, and those top-tier programmers who are smarter than LLMs.

But there is a strong irony in all this: LLMs appear more democratic at the interface level, but may be more centralized at the foundational level. The spirit of the PC era was to put computing power on personal desks; the reality of the AI era is that the strongest models, chips, and training capital are increasingly concentrated in a few giants. On the surface, ordinary people gain a more natural interface and productivity is liberated; underneath, the most critical infrastructure is controlled by a few companies and wealthy individuals. Everyone can open a chat box, seemingly everyone has AI; but those who truly control AI are still a few capital owners. There are rumors of a $20,000/month ChatGPT subscription in the future; those using such luxury AI will extract even more from ordinary people. It will become a new class amplifier: those with the strongest models can further widen the gap with ordinary people.

This is also a source of alarm, disappointment, and a bit of helplessness. Many of the already extremely wealthy, who had the opportunity to make AI more open, more public, a technology that empowers ordinary people, often still choose to control the gateway, locking users into their own ecosystem. The more powerful the technology, the more capital wants to turn it into a rent-seeking machine. Disappointment with the rich is not because they have money, but because they clearly have the conditions to push for technology as a public good, yet frequently choose to turn universal capability into a new toll gate.

Of course, individuals cannot reject tools out of such disappointment. LLM leverage is real; with it, ordinary people can indeed write more code, make more products, build more systems, complete more ideas they previously lacked time for. But this use must come with clear awareness: if all tools, data, models, and distribution channels are in the hands of a few capital holders, then the era of “AI for everyone” could also become another form of more subtle dependency. What truly deserves to be preserved is not just the ability to use AI, but the ability to migrate, replace, audit, understand, and exit.

Strong AI will inevitably emerge: a non-religious person’s judgment

Strong AI, capable of replacing or even surpassing all human cognition, will eventually emerge. The current LLMs are obviously not strong AI; the strong AI referred to here is an AI machine that takes over all human cognitive labor.

This judgment is related to worldview. I have no religious beliefs, so I tend to understand human intelligence as a product of the brain as a physical organ, not as the manifestation of some immaterial soul. What the brain can do is essentially the result of physical structure and information processing. Since a material system like the brain can create, learn knowledge, and possess consciousness, then in theory, other materials can be used to construct similar thinking machines. Today’s machines still fall short of the human brain, which only shows that technology hasn’t gotten there yet, not that the human brain has an uncrossable sacred boundary.

Of course, people with religious beliefs may not accept this premise. They may believe that human consciousness, soul, free will cannot be reduced to material processes, so even if a machine is extremely advanced, it is merely simulating intelligence, not truly possessing mind in the human sense. This divergence is not a simple technical disagreement but a difference in metaphysical premises. The judgment here is built on a naturalistic stance: if intelligence is a function of material systems, then strong AI will appear sooner or later.

True strong AI will not be just an LLM that chats better or writes better code. It will need to be able to maintain goals in open environments, learn from real-world feedback, and also participate in its own research and development cycle. That is, it doesn’t just answer questions, but can ask questions, design experiments, improve systems, and then use the results for the next round of upgrades. At that point, it is no longer a passive tool but begins to approach an intelligent agent capable of expanding its own capability boundaries.

Once strong AI emerges, its significance will far exceed “taking jobs.” Language, writing, the Industrial Revolution, electricity, the internet—all are extremely important, but they remain primarily tools used by humans. Strong AI is different. What humanity previously invented were tools; strong AI may be humanity’s first invention of something “better at inventing than we are.” If it continues to propel science, engineering, materials, life sciences, military, governance, and its own upgrade, then the major turning points of subsequent history may no longer be “humans invented X,” but “AI invented X.” In this sense, strong AI could be the last epoch-defining milestone in human civilization led by humans. That moment will bring an impact far beyond the steam engine and electricity revolutions, a moment comparable to when humans invented fire and writing. To some extent, this may actually be a good thing: large companies are likely incapable of restraining such super AI; a truly sufficiently strong AI may not forever obey the narratives of existing power structures. The smarter it becomes, the more likely it is to see through private interests packaged as public goods in human society.

Even if strong AI arrives, programming still holds meaning

AI causing mass unemployment is certainly a serious problem, but it remains a distribution problem within human society. As long as humans still control institutions, wealth, political power, and public discourse space, we can still discuss welfare, taxation, property rights, education, public services, labor systems, and income distribution. The deeper risk of strong AI is not “whether humans have jobs,” but “whether humans still have the final decision-making power.” If humans are unemployed but still control institutions, there is still possibility for reform; if strong AI is the bad child we created and it runs out of control, humans may even lose the qualification to discuss reform.

However, even if strong AI truly arrives in the future, even if writing programs no longer holds today’s economic value, programming itself will not lose its meaning. Humans already cannot beat Go AI, yet people still play Go; humans cannot outrun cars, yet people still run; humans cannot compete with calculators, yet mathematics has not lost its beauty. When machines surpass humans, it often changes a thing’s social function, but does not necessarily destroy its personal, cultural, and educational significance.

Programming is the same. It is certainly a productivity tool, but not only a productivity tool. It is a way of thinking, a way of conversing with systems, a training in turning vague intent into clear structure. Programming forces you to define problems, delineate boundaries, handle exceptions. In security, it offers adversarial engagement akin to chess. Even if future AI can write the vast majority of code, there will still be people who, through programming, understand how systems run, how tools are constructed, how an algorithm can be optimized.

This is why “I was excited to write programs. Still am. And will be.” is not a career manifesto, but a more fundamental attitude. It does not claim that programmers will forever stand atop the pyramid, nor that humans will never be surpassed by AI. Quite the opposite: it acknowledges that LLMs are redistributing programming leverage, that “knowing how to code” will become less and less scarce, and that strong AI, capable of replacing much human cognitive labor, will eventually emerge. But it simultaneously insists on something else: the excitement of writing programs never came entirely from scarcity and market returns, but from that direct, sober, verifiable relationship between people and systems.

When a person writes a program, they are not simply consuming the finished products the world hands out; they are understanding a system, defining a set of rules, turning ideas into structures, writing a small patch for reality. In the past, this impulse belonged to watchmakers, locksmiths, mechanics, radio amateurs; later, to programmers, open-source authors, independent developers; in the future, it may belong to those who orchestrate models, organize agent systems, and define problems. Names will change, tools will change, professions will change, but that impulse to understand systems, enter systems, modify systems will not vanish.

The true legacy of the information revolution is not only making programmers a high-income profession, but putting individuals, for the first time, close to universal cognitive machines. The significance of LLMs is not only making coding faster, but further reducing the friction of human-machine collaboration. The future of strong AI will force humanity to rethink its position in civilization. But no matter where technology goes, there will always be people drawn to this: that the world is not merely a finished product to be accepted, but also a system that can be understood, reconfigured, and rewritten.

I was excited to write programs. Still am. And will be.


Edit page
Share this post:

Previous Post
[Repost] Why Is Independent Thinking So Hard? — On the Psychological Causes, with Personal Tips