AI Is Compressing the Timeline of Technological Change

If you follow artificial intelligence closely, there is a peculiar feeling that has become increasingly difficult to shake.

Everything is speeding up.

A major model launches on Monday. By Friday, developers have built products around it, researchers have tested its limits, competitors have responded, and someone is already leaking benchmarks for whatever comes next.

A capability that seemed remarkable six months ago becomes a checkbox. An AI coding workflow that felt experimental last year becomes normal enough that developers complain when it briefly stops working. Research papers increasingly read like dispatches from slightly different technological eras, despite being published months apart.

Some of this is undoubtedly perception. AI has attracted extraordinary attention, money and competition. Millions of people are watching every incremental improvement in a way few people watched advances in steam turbines or semiconductor manufacturing.

But there is also something measurable happening beneath the noise.

According to Stanford's 2026 AI Index, generative AI reached roughly 53% population-level adoption within three years, which Stanford says is faster than either the personal computer or the internet. AI use across surveyed organizations has reached 88%, with generative AI being used in at least one business function at 70% of organizations.

That is an extraordinary diffusion curve. And I think understanding why it is happening explains a lot about why technological change suddenly feels so damn fast.

Think about the major technological revolutions that preceded AI.

The Industrial Revolution transformed civilization, but industrial technology had a serious distribution problem.

You could invent a better steam engine, but someone still had to manufacture it. Factories had to be constructed or redesigned. Machines had to be installed. Transportation networks had to move raw materials and finished goods. Workers had to relocate, retrain and reorganize around entirely new systems of production.

Even economically superior technologies could take decades to spread.

Research on 19th-century steam adoption, for example, has found that high fixed costs and the difficulty of switching existing factories from waterpower significantly slowed adoption. One NBER analysis estimated that removing those switching barriers would have pushed steam adoption to a 30% rate more than two decades earlier.

Transformative did not mean instantaneous. Computing accelerated the process, but computers still had to physically arrive.

In 1984, only about 8% of U.S. households had a computer. By 2003, that number had reached roughly 62%. That is spectacular growth, but it still took nearly two decades for home computing to make that journey.

Then came the internet.

Here the curve starts getting noticeably steeper. Only 18% of American households had internet access in 1997. By 2003, roughly 55% did. Part of the reason was simple: the internet did not have to build civilization from scratch.

Computers already existed.

Telephone and cable networks already existed.

Semiconductor manufacturing already existed.

Software ecosystems already existed.

The internet inherited an enormous technological foundation constructed during previous waves of innovation.

AI essentially inherits all of it.

Consider what somebody needs to adopt generative AI today.

Usually, they need a browser, a mobile phone, or a personal computer.

That’s about it.

They do not need to purchase a steam engine, construct a factory or install a mainframe. They often don’ even need to download software.

The world's computing infrastructure was already waiting when ChatGPT arrived.

By 2021, 95% of U.S. households had some form of computer and 90% had broadband internet. Smartphones were present in 90% of households. Meanwhile, businesses had already spent decades digitizing documents, workflows, communications, databases and software.

Then cloud computing centralized enormous amounts of processing power behind APIs accessible from basically anywhere. AI arrived on top of all of that.

Which brings us back to the chart.

When I talk about broad adoption, I am not suggesting that the Industrial Revolution was precisely 35% adopted in some particular year or that AI will mathematically follow a predetermined curve.

The chart is conceptual. Broad adoption describes the transition from a technology being experimental, specialized or confined to early adopters into something routinely used by ordinary people and businesses. That transition appears to be compressing.

Industrial technologies frequently required generations.

Personal computing required decades.

Internet adoption accelerated substantially.

Generative AI reached a majority-scale global adoption estimate within three years. Each technological wave helped construct the distribution system for the next one. AI happens to be arriving very late in that sequence.

That matters enormously.

There is another important difference. Most AI capabilities are software.

Suppose OpenAI, Anthropic or Google discovers a meaningful improvement to a frontier model.

Once that model is deployed, millions of users can potentially gain access without replacing anything sitting on their desks.

No trucks arrive. No factory needs to retool. No technician comes to the house.

A new capability simply appears inside an existing interface.

The physical infrastructure behind that experience is enormous, of course. Data centers, GPUs, networking equipment and power plants most definitely exist in the real world.

But users experience the final capability as software. That creates a strange mismatch. Building the intelligence remains brutally physical.

Distributing the intelligence can happen at software speed.

That alone would make AI adoption unusually fast, but it still doesn’t explain the entire sensation of acceleration.

Because something else is happening simultaneously.

The technology itself is improving rapidly while adoption is spreading.

Stanford's 2026 AI Index found dramatic gains across several measures of AI capability. On SWE-bench Verified, a widely followed software engineering benchmark, frontier performance went from roughly 60% to nearly 100% in a year. On OSWorld, which tests agents performing real computer tasks, success rose from around 12% to roughly 66%.

Benchmarks deserve skepticism. They saturate. They can become contaminated. Success on a benchmark does not magically translate into reliable performance in every real-world environment.

But independent measurements are finding a similar direction.

METR has been tracking the length of software tasks frontier AI agents can successfully complete. Its research has found that the task horizon has historically doubled roughly every seven months, although METR explicitly warns that the measure has substantial uncertainty and varies dramatically across domains.

So we have two curves accelerating at once.

More people can use AI. And the AI those people receive keeps becoming more capable.

That creates a dynamic previous technological revolutions did not experience in quite the same way. Your railroad did not become substantially smarter three months after you started using it.

Your washing machine did not wake up one morning with much better reasoning capabilities.

The internet certainly evolved rapidly, but the network itself was primarily a distribution mechanism. Most innovation happened through the products people built on top of it.

AI is the product, the platform and increasingly part of the machinery used to improve the platform.

That last part may be the most consequential.

We’ve even reached a sort of entry-level "recursive self-improvement." Current models are not autonomously redesigning themselves, building their successors and launching them without human involvement.

But something else very real is already underway. AI is increasingly being used by the people building AI.

Anthropic reported in May that more than 80% of the code merged into its own codebase was authored by Claude, up from the low single digits before Claude Code entered research preview in February 2025.

Google DeepMind's AlphaEvolve provides an even stranger example.

The Gemini-powered system has been used to discover and optimize algorithms involved in Google's data centers, chip design and AI training. DeepMind says some of those improvements have been used in training the large language models underlying AlphaEvolve itself.

Read that carefully.

An AI system helped improve the computational processes used to train the class of AI systems that produced it.

Humans are still choosing objectives, running experiments, evaluating outputs, building infrastructure and deciding what gets deployed. But the loop has changed. AI researchers now have AI research assistants.

Software engineers building AI systems have AI coding agents. Chip designers can use AI systems to optimize chips used for AI.

Scientists can use AI to search literature, write simulations and investigate problems that may eventually improve algorithms, hardware or energy systems feeding back into AI development.

The loop does not need to become fully autonomous before it matters.

It only needs to get faster.

This is where the pace starts to become easier to understand. AI development is benefiting from several acceleration mechanisms simultaneously.

Each layer pushes on the others. Better models create more useful products. More useful products attract more users. More users justify greater investment. Greater investment buys more compute. More compute and better research methods produce stronger models.

Stronger models become better tools for engineers and researchers. Those engineers and researchers accelerate the next cycle.

You do not need to believe in an imminent intelligence explosion to recognize the feedback structure. We can already see pieces of it operating.

Stanford reports that global corporate AI investment more than doubled in 2025, while generative AI investment grew more than 200%.

Capital is chasing capability, capability is driving adoption, adoption is creating demand for more capability.

And intelligence itself is increasingly entering the production process. That’s a formidable flywheel.

There is an obvious objection to all of this. Atoms remain stubborn.

AI may travel through the internet at nearly zero distribution time, but the systems running it require fabs, GPUs, data centers, transformers, transmission lines, cooling equipment, land and enormous quantities of electricity.

Those things cannot be downloaded. This may become one of the defining tensions of the next phase of AI. The software curve wants to go vertical, but the physical world does not.

AI companies are already pouring extraordinary amounts of capital into compute infrastructure. Stanford reports that compute spending and infrastructure investment are reaching record levels alongside AI revenue growth.

There are other brakes, too. Reliability remains uneven. Agents remain much less widely deployed than ordinary generative AI, with Stanford finding single-digit adoption across nearly all business functions.

Organizations take time to redesign workflows.

Governments take time to write rules.

Schools take time to rewrite curricula.

Humans take time to change habits.

The technology may move at software speed, but civilization usually doesn’t. That gap may become increasingly uncomfortable.

Perhaps this is why the AI era feels so different. Human beings have lived through transformative technologies before.

Steam changed where we worked. Electricity changed when we worked. Computers changed how we processed information. The internet changed how information moved.

Each revolution left behind infrastructure, institutions, knowledge and tools that reduced the friction facing the next one.

AI is arriving after two centuries of that accumulation, so it does not have to build the entire road. Much of the road is already paved.

And now, for the first time, the thing traveling down that road can help engineers make the vehicle faster. That doesn’t guarantee every AI forecast currently floating around Silicon Valley will survive contact with reality.

But it does help explain something increasingly difficult to dismiss. The intervals are shrinking.

The Industrial Revolution took generations to remake society.

Computing took decades to become ubiquitous.

The internet compressed the transition further.

AI has barely had time to introduce itself. And already, more than half of the measured global population has encountered generative AI, businesses around the world are integrating it into workflows, and the engineers building the next generation are increasingly using the current generation to help them do it.

Maybe the most important feature of the AI revolution is not simply how powerful the technology becomes.

It may be how little time the world gets between versions of the future.