Generative AI Will Not End Work Directly; It Rearranges Labor and Expands Demand First

In 1930, Keynes had a famous judgment that, as productivity continued to rise, humans could only work 15 hours a week at the beginning of the twenty-first century. Looking back today, this projection is only half. There has been a significant increase in productivity, but labour has not disappeared proportionately. Instead of converting the ability to release technology into leisure, we have reinvested it into more consumption, higher quality, faster response and more complex services.

The questions in this article can also be addressedLLM brings technology equity? Does LLM Bring Equality?、AI Agent's future is not a full autonomy, but a manageable commission.How the concept of a relatively close read together is developed in different contexts.

This is also the starting point of the generation of AI today. It will certainly impact jobs, reduce old tasks, and even blur the boundaries of some jobs, but I do not think it will directly destroy the fabric of society as a whole.AI is more like a job shift than an end to it; it is more likely to magnify demand than to stop it.

Why is Keynes only half right?

Keynes underestimated the technical advances themselves, but rather the elasticity of human desires and the ability of societies to expand their own division of labour. When productivity increases in a society, it does not automatically slide towards a steady pattern of people starting to work less. More often, the result is that what is expensive is cheaper, that services that are scarce are more common, and that people start demanding higher standards and take them for granted quickly.

Air conditioning, automobiles, smartphones, take-out, instant communications have all experienced similar processes. They were not originally essential for survival, but were slowly absorbed into everyday life by society after costs fell and infrastructure improvements were completed. Technology does not shrink desire, but keeps raising the default configuration of life. Thus, labour is not simply replaced, but is re-organized into new industries, new service chains and new systems of comparison.

Why efficiency gains often do not reduce total workload

There are some classic explanations, but they can be summed up in one sentence:When efficiency gains lower supply costs, societies are often not content with “doing less” but instead “doing more”.

There is Jevons Paradox in economics, which speaks of efficiency gains that may lead to an expansion of aggregate demand; there is a sense of pleasure in psychology, where people quickly adapt to new conveniences and use it as a new baseline; and in modern consumer societies, quality, speed, personalization and sustained response themselves are redefined as new demand. Thus, the productivity unleashed by technological progress does not automatically become rest time, but rather continues to be absorbed by society.

That is why automation has not completely eased modern organizations. The water lines improve manufacturing efficiency, but do not end management, coordination, design, marketing and marketing; the Internet reduces the cost of information transmission, but does not reduce information work, but makes content production, platform operation, search screening and attention competition more intense. Efficiency gains often do not remove people from the system, but rather push the system towards higher density.

Generating AI is a continuation of the same logic

Put this line on the generated AI, the problem will be much clearer. The generation AI has reduced the marginal costs of partial cognitive labour, expression of labour and content production. It makes writing, translation, retrieval, design, programming, passenger service, analysis of some of these tasks cheaper, faster and more standardized. But this does not mean that human demand for content, services and judgement will decline simultaneously.

On the contrary, if it is cheaper to generate, society probably will need more. While the last report was written for three days, it may now be available in half a day, but the next change is not necessarily that the enterprise leaves its employees free for two and a half days, but rather that it requires faster iterativeity, more versions, more disaggregated audiences, more timely feedback, and more robust validation and delivery. Coding Agent can magnify 10 times the development efficiency, so software engineering may be able to customize software for everyone; this will require more engineers and a lot of token, and this development will shift from one side of the enterprise to the other. The content will become more numerous, the interaction will become more interactive, the flow of information will be faster, individualized services will become more sophisticated and, ultimately, even the assessment of what is real and what is credible will become more heavy.

Several major technological leaps in history have been similar. The industrial revolution did not end the work, but rather reorganized the jobs; automation did not eliminate organizational collaboration, but rather created new management and interface jobs; and the Internet, instead of reducing information work, magnified the need for information production, distribution and screening. Generating AI is more like moving this logic to the cognitive labor level: it automates a set of old tasks (which we thought only human intelligence could do) and opens up new services that were expensive and are now becoming viable.

AI Coding is strong enough to reorder division, but not enough to replace advanced engineering. Division

Software engineering is also appropriate to observe this change because it also contains automated repetitive tasks, system judgements that are difficult to replace, and clearly visible layers of capability. This logic of “re-loading, re-valueing” would be more specific if the perspective were further narrowed to programmers. Today's AI Coding is not just a few demo toys, it's already able to read code, move, lay out a whole replacement, mass patches, and even produce a complex system that can run. The problem was not whether it had the capacity to do something, but what it had made, often a short distance from mature work.

That's what I'm talking about. Chips and Cheese test of Claude to generate a C compiler It's written in plain white. The article was clearly a sarcastic but technically clear: a compiler driven purely by ant team has produced a real product that can run, compile, produce results; but it still has a significant difference in performance, optimization of maturity, code quality compared to the human engineering that has been grinding for decades. Full autocoding is strong enough to get into the real scope of the project, but not enough to get the senior engineers out of the stage. It's more like a part of a job within a compressor than a whole rewriting of software.

The other side.Articles The blogger says that the government is not going to be able to get a better and more optimistic picture: When AI is not required to invent an independent system from zero, but to enter a complex system that has been well understood by senior engineers, with clear boundaries and clear standards for validation, its productivity is rather magnified. Large-scale interface migration, cross-language mapping, global rule-driven rewriting, rapid absorption of external best practice, which in the past has consumed a large amount of high-level manpower, can now be taken over by angent.

But the scarce parts here are not gone, but are more clearly exposed. AI is best at matching and spreading the large-scale semantics after the rules have been defined; it is still scarce, and it is the specification, boundary determination, standard design for validation, and the final correctness responsibility in complex systems. AI Coding first changes not whether or not the programmer is still there, but who does the job within the programmer. In this sense, it is more like a re-engineering of the division of labour in software engineering than a replacement for the entire career.

The zooming up is not just output, but the faults between the senior and junior engineers.

This is why the impact of AI on the labour market for programmers is likely to fall not on the unemployment of senior engineers but on the growth of junior engineers. Many engineers in the past did not come up to make structural judgements, define specification, design validation; they often began in a real system, starting with small bug repair, retracing, migration, demand for sideways, and spreading repetitive changes, gradually understanding why the system was designed so that it could move, where it could move, and where it could never move. Those that appear to be the most trivial and repetitive are often also the most critical training grounds.

The problem is, these tasks are exactly what is best for the coding parties to take over. For models, work that is clear, repetitive, broad-ranging, patient and consistent is often more friendly than for new human beings; for organizations, a senior engineer plus an agent may be more cost-effective in the short term than a senior engineer with a few junior engineers. The former are almost immediately operational, do not require long-term training, do not leave halfway and are more stable than new recruits in a number of global rewriting tasks. So AI replaces not only low-value duplicated work, but also a part of the low-threshold but highly trained growth path.

That's what I think is deeper risk. What is lacking is not necessarily simply the ability to write codes, but the ability to judge whether codes are correct, to understand why systems are so designed, to know what to measure and what not to change. These capabilities are often modified by long-term participation in real projects, repeated implementation, repeated pitfalls, and feedback from senior engineers. AI makes realization itself cheaper, but does not automatically hand this judgement over to the newer. If access to high-quality real projects continues to decline and organizations continue to treat an individual as an alternative to the application, the impact is not just the number of jobs, but the path of people to grow into high-level workers.

From the sample of programmers, the impact of the generated AI is not just "can it write code?" The more acute problem is that when senior engineers are significantly amplified and the training paths of junior engineers begin to shrink, the industry as a whole will grow up in the coming years to the next generation of people who understand complex systems. It is not a question of emotionalization, but of supply. AI is not just re-ordering, but is re-aligning people as high-level workers.

The real guard is not a collapse, but a transformational mismatch.

This does not mean that AI can be downplayed as creating more jobs automatically or directly leading to total unemployment. It's too simple, no society can be summed up in a single sentence, not to mention a system of 8 billion people connected. The contraction of local jobs is real, the depreciation of skills is real, and the income shocks and occupational anxiety during the transition period are real. For programmers, the shock may even begin with a reduction in junior positions, a narrowing of training paths and a greater willingness on the part of organizations to buy modelling capabilities rather than to develop people. The question is never whether there is a shock, but how fast it will come, and whether society has done re-training, reorganisation and redistribution fast enough.

What we're about to meet isTransition mismatchInstead of unemployment in society as a whole. When technology spreads faster than the adjustment that is buffered by the education system, business organization and system, the pain is concentrated; but this is not the same thing as AI is going to get society out of work. This mismatch is not just a mismatch of jobs after the task has been replaced, but also of capacity structures, growth paths and memory: Old mandates have disappeared, but new certification capabilities, system judgement and collaborative approaches are not necessarily equally rapidly captured by new generations.

The strength of human society lies in its ability to redistribute tasks so strongly that, when old mandates disappear, people redefine values around new tools, new interfaces, new ways of working together. But what is of concern is not whether society will be directly breached, but whether we can catch up with this productivity leap with sufficiently fast-tracked organizational adjustments, educational buffers and institutional arrangements; otherwise, the short-term increased efficiency could well become the cost of a short-term timescale of brain faults and inadequate capacity.

References

Chips and Cheese, Embracing AI with Claude’s C Compiler
Zhao Taku, know what you say,Discussion on Claude Code, maintenance of scientific software and human capacity development