AI Is Not Like the Internet. It Helps Build Its Successor
Historical technology analogies are useful, but they may understate how quickly AI can change the work around it.
Part 3 of 3 in The Future of Software Engineering When Code Becomes Cheap.
Whenever people talk about AI and employment, someone eventually reaches for the history of electricity, computers, or the internet. I understand why. Every major technology shift has produced anxiety, every one has displaced some work, and every one eventually created industries and roles that would have been difficult to imagine at the beginning.
That historical argument is worth taking seriously. Panic is usually a poor operating model. Organizations rarely absorb a general-purpose technology overnight, and the complementary changes are often slower than the invention itself. Electricity required factories to be redesigned. Computers required new business processes. The internet required logistics, payments, identity, security, content models, advertising models, and habits that took years to develop.
So when someone says AI may take longer to absorb than the loudest forecasts imply, I think that is a valid concern. Institutions are slow. Procurement is slow. Regulation is slow. Training is slow. Trust is slow, especially inside organizations where a wrong answer can create financial, legal, safety, or reputational consequences.
What I am less convinced about is whether the older analogies fully capture what is different this time. Electricity did not help invent better electricity. The internet did not independently redesign the next generation of networks. AI is increasingly involved in the research, engineering, coding, testing, evaluation, and operational work required to improve AI.
That does not make the future automatic. It does make the feedback loop different.
Earlier technologies mostly amplified physical power, calculation, communication, and information distribution. AI increasingly amplifies cognitive work, including the work of technology creation. It can help write training and inference code, generate synthetic data, evaluate model behaviour, analyze experiments, search research literature, optimize infrastructure, design chips, discover algorithms, and coordinate research workflows.
The technology is becoming one of the inputs used to improve the technology. That is the part I think historical comparisons often miss, and it is why I am cautious about assuming AI will follow the same absorption curve as earlier platforms.
To be clear, I am not arguing that we have runaway self-improvement. I am not arguing that current systems can independently redesign themselves without human direction, physical infrastructure, capital, energy, manufacturing, or approval. What we have evidence for is a weaker but still important loop: AI helps people improve AI.
That distinction matters. Strong recursive self-improvement would mean an AI system can meaningfully set its own research goals, redesign its own architecture, obtain the needed resources, validate the result, and continue the cycle without substantial human control. That is not what appears to exist today. Current systems remain dependent on human researchers, human-defined objectives, specialized hardware, data centres, energy, training data, manufacturing supply chains, and institutional decisions.
But a weaker loop can still change the pace of the transition. Better models improve software-development productivity. Better software improves training, evaluation, deployment, and monitoring. Faster experiments produce better models. Better models then accelerate more research and engineering work.
Even if humans remain in control of the objectives and approvals, the cycle time can compress. A research group that can run more experiments, evaluate more failures, and improve more tooling may move faster than one relying entirely on human labour. An infrastructure team that uses AI to optimize systems may reduce costs or increase throughput sooner. A software team that uses agents to build evaluation harnesses may discover regressions faster and ship improvements more confidently.
This is not magic. The visible product may be a chat window, but underneath it is an industrial system. Frontier training runs, data-centre construction, Nvidia accelerators, TSMC fabrication, high-bandwidth memory, power generation, cooling, network infrastructure, capital investment, export controls, and national industrial strategy all sit behind the interface.
That physical reality is important because it prevents the argument from drifting into science fiction. The feedback loop is cognitive, but it remains constrained by hardware, energy, manufacturing, supply chains, capital, regulation, and human organizations. A model can help design an improvement, but someone still has to build, power, finance, approve, and operate the infrastructure that makes the improvement real.
There are other constraints as well. High-quality data may become harder to obtain. Energy supply may become a limiting factor. Hardware bottlenecks may matter more than model ideas. Reliability problems may slow adoption. Regulation or safety incidents may interrupt deployment. Economic limits may force companies to charge prices that change usage patterns. Public resistance may grow if the benefits and costs are distributed badly.
This is why I do not find confident single-year predictions very useful. The next five years could include very rapid progress, but it will likely be uneven. Capability may improve faster than reliability. Model quality may improve faster than enterprise governance. The tools may change faster than procurement, security review, hiring models, education, or operating processes can absorb.
That last point may be the most practical one for organizations. Earlier enterprise technologies often gave companies long periods to adapt. With AI, the tool can change while the organization is still deploying the previous version. A model release can alter the economics of last year’s architecture. A new agent capability can make a process redesign obsolete before the process has stabilized. A governance framework can lag behind the behaviour it is supposed to govern.
The challenge may not be adopting AI once. It may be repeatedly reorganizing around a technology that changes faster than institutions do.
For workers, that creates a different kind of transition risk. Previous technologies automated tasks over long absorption periods. AI may automate a workflow, help redesign the workflow, and then automate pieces of the redesigned workflow before the labour market has fully adjusted to the first change. New jobs may emerge, but not necessarily at the same rate, in the same places, or at the same skill level as the work being reduced.
I am careful about this because there are two lazy versions of the argument. One says AI is just another tool and everyone worrying about it is overreacting. The other says every exponential projection continues forever and mass displacement is inevitable. I do not think either view is serious enough.
The next five years are better understood as a range. By approximately 2031, it is plausible that agents will complete multi-hour or multi-day digital work, smaller teams will produce much more software, AI will assist with model research and system optimization, cycles between model generations will shorten, governments will treat frontier AI more like regulated infrastructure, and enterprises will redesign processes around AI-native operations. It is also plausible that reliability, infrastructure, regulation, energy, or economics will slow parts of that path.
The main uncertainty is not whether AI improves. The uncertainty is whether reliability, infrastructure, governance, and institutional adoption keep pace with capability. In the enterprise, that distinction matters more than benchmark movement. A model can be impressive and still not be trustworthy enough for consequential work without controls around identity, data, permissions, evaluation, audit, and human accountability.
For individuals and organizations, I think the right response is neither panic nor complacency. Engage early. Stay model-agnostic. Invest in fundamentals. Build evaluation and governance capability. Avoid rigid five-year technology bets that assume one provider, one model, or one architecture will remain dominant. Design systems that can change underneath you, because they probably will.
This also means preserving human accountability rather than pretending automation removes it. If an AI system recommends, writes, routes, approves, scores, summarizes, or acts, the organization still needs to know who owns the outcome, how the decision can be reconstructed, what data was used, what permissions were granted, and where the exception path lives. Those are not optional details once AI becomes part of real work.
AI may ultimately resemble earlier general-purpose technologies by creating more work and prosperity than it destroys. I hope it does. But it may reach that destination through a faster, less stable, and more self-reinforcing transition than organizations are used to managing.
We should learn from the history of electricity, computing, and the internet. We should not assume AI is required to repeat it.
This article is part of The Future of Software Engineering When Code Becomes Cheap, a three-part series on how AI changes implementation cost, engineering apprenticeship, and the pace of software work.
Read the series: