The Window

"Human history becomes more and more a race between education and catastrophe."
— H. G. Wells

My son said his first words as ChatGPT launched. I used to say that I felt like I was watching them learn to speak at the same time. In the time my son has gone from speaking his first words to reading "cat", ChatGPT has gone from stitching together sentences to outscoring most PhDs in most disciplines. And the curve is only picking up. Hundreds of billions of dollars are being poured into exponentially larger data centers. Engineering talent is gushing in. Data is multiplying. Every major nation and every major company is racing to build the most capable models possible.

My son is learning at the pace of a child. The machines are learning at the pace of machines. Those two clocks are running on the same wall.

This chapter is about the clock that matters more than people think. Not the one inside the labs. The one outside them. The pace at which the world actually changes when a new technology arrives. Here is the thing the doom version misses, and it is really important: even in the worst case, the window is wider than the panic suggests.

We Have Time

In 1950, Alan Turing proposed what became known as the Turing Test: if a human judge cannot reliably distinguish between a machine's responses and a human's responses in conversation, the machine can be said to demonstrate human-like intelligence. In 2024, GPT-4.5 flew past the Turing Test, and the achievement barely made the news.

In 2007, as AI researchers made advances in narrow forms of AI, AI researcher Ben Goertzel popularized the term "artificial general intelligence" (AGI), defining it as "the ability to solve general problems in a non-domain-restricted way, in the same sense that a human can." By that definition, we've arrived. Today's AI models score 83-92% on GPQA Diamond, a test of PhD-level scientific reasoning. Human PhD experts average 69.7%. These models resolve real-world software engineering problems at 77% accuracy on tasks professional developers face daily. They pass graduate medical exams 20 percentage points above the threshold. These numbers will be long-since left behind on the exponential curve when you read this.

Another definition of the key event horizon into AGI was when AI has the ability to improve itself. From I.J. Good's original paper in 1965: "an ultraintelligent machine would be capable of redesigning itself faster and more effectively than human engineers could, thereby triggering a positive feedback loop of accelerating improvement. Each iteration would produce a machine better able to design the next, yielding a runaway process that would culminate in intelligence 'far beyond the human level in all respects.'" We've crossed this threshold as well. Anthropic says that Claude is now writing 90% of its own code.

When the open source AI agent OpenClaw was released, it took the internet by storm. The tool allows a computer complete access to your email, calendar, Notion account, even password manager if you give it to it. And it is designed to be completely unfettered in its autonomy. Business Insider described that, "one founder asked it to make him dinner reservation; when it couldn't complete the task via OpenTable, it used its ElevenLabs skill to call the restaurant." This founder wrote, "AGI is here and 99% of people have no clue."

All this, and yet, society hasn't transformed overnight. We wake up each day like we did the day before. In Sam Altman's words, "AGI kinda went whooshing by. It didn't change the world that much."

In his piece Machines of Loving Grace, Dario Amodei talks about the counterintuitive truth: intelligence does not automatically translate into immediate real-world change. Even superhuman intelligence runs headlong into constraints imposed by reality itself. The physical world moves at fixed speeds. Biology takes time, experiments must be run sequentially, materials must be fabricated, infrastructure must be built, and humans must be coordinated, persuaded, and trained. In many domains, intelligence is not the bottleneck. Data collection is. No amount of reasoning can substitute for measurements that take time to collect. Human institutions impose further friction through laws, regulations, norms, and legitimacy constraints. Many are designed to slow adoption. And beneath all of it sit unbreakable physical limits: energy costs, thermodynamics, semiconductor density, and the speed of light itself. AGI may accelerate thinking, but it cannot compress time, bypass physics, or instantly rewire society. Progress, in reality, even in an age of superintelligence, still unfolds at the pace of atoms, institutions, and humans.

As we look ahead to what happens when we reach increasingly ambitious definitions of AGI, we can learn something from these past few years. Humanity changes at the pace of humans. Technology rolls into society much slower than we expect.

I remember living in Silicon Valley in the 2010s, when everyone was pouring money into self-driving cars. In surveys conducted between 2014 and 2016, one in five people thought that most cars would be able to drive fully automatically by 2020. Business Insider predicted that there would be "10 million self-driving cars on the road by 2020." Today, the technology is excellent. Elon Musk says that, "insurance is half price when Tesla self-driving is activated, because it increases safety so much." And yet, despite that, as of late 2025 there were only 2,500 Waymos on the road in highly geofenced locations in 6 US cities. Even Tesla, with 5 million vehicles equipped with Full Self-Driving capability, has convinced only 12% of owners to use it, some of the time. Yes, this is physical. But so is most of society.

Friction shows up everywhere. Humans adopt transformative technologies slowly. Smartphones took sixteen years to reach half the US population, email took decades to fully displace memos, and videoconferencing existed long before it reshaped work. Organizations move even slower: embedding AI into medical records, financial systems, or global logistics software takes years because mistakes cascade at scale. Regulation compounds the delay, especially in physical industries like construction, healthcare, and transportation, where safety, liability, and jurisdictional complexity dominate. Hospitals, universities, corporations, and legal systems don't just "install" new intelligence. They must renegotiate roles, incentives, workflows, and power structures across thousands of people. Even with AGI, those transformations unfold on a human timescale.

Let's contrast these friction-filled implementations with Facebook, the fastest-growing, stickiest, most viral technology ever created. It's pure software, free, requires no regulatory approval, no new hardware, no installation. It took 7 years from the founding of Facebook to cross the 50% threshold in the US. It took another 5 years, during Trump's 2016 election, before people realized the ways it was fundamentally changing the fabric of society. Even the fastest-moving technologies ripple their way through society relatively slowly.

We cannot conflate technical capability with economic deployment. History shows technology can be ready long before institutions actually restructure. ATMs were invented in 1967; bank tellers increased in number for decades after.

Most industries are full of friction. I was recently speaking with a friend who works in process automation for construction and oil and gas. He described that AI hasn't touched what he does. He wants it to. But their processes are so painstakingly precise, so complex, and so high stakes, that you can't simply allow an AI to rebuild that entire software stack. The majority of industry, logistics, manufacturing, construction, and infrastructure is more like this than the vibe coded app that a high school student can whip up.

As Andrej Karpathy put it in a tweet, "I'm a little bit disappointed (and my timelines are correspondingly slower) with how slowly this progression is happening in the industry overall. 99% of products/services still don't have an AI-native CLI yet…. They give you a list of instructions on a webpage to open this or that url and click here or there to do a thing. In 2026. What am I a computer? You do it. Or have my agent do it."

As Aaron Levie, the founder of Box, put it: "We dramatically underestimate how much change management it is going to take to automate most knowledge worker tasks. Between data being in legacy environments or systems or without good APIs, context missing for doing the task, teams that are less technical, and other factors, there's still a lot of work to drive real AI transformation in an enterprise."

Integration takes time.

If full-blown AGI arrived today (arguably, it already has), here's what would actually happen: Tomorrow, nothing changes. 99% of people see it on their feeds and go about their lives. Within months, a few hundred companies start internal pilots while most businesses form task forces to "explore the implications." By one year, major tech companies and consulting firms deploy it internally, the first wave of junior analysts and content writers lose jobs, but 95% of the economy operates unchanged because integration with existing systems takes time. Kind of like what's actually happening right now.

By 4-5 years, knowledge work has fundamentally transformed. Legal services see huge displacement, consulting restructures dramatically, healthcare diagnostics are revolutionized but delivery remains human, and unemployment rises as wealth concentrates to early adopters. The physical world transforms slower than the cognitive one. Medical care is still majority human. Diagnostics are AI-assisted but doctors see patients and nurses provide care. Construction is maybe 20% automated, held back by robotics limitations. Skilled trades like plumbing and electrical work remain largely human. Shipping and logistics are heavily automated in routing and planning but still require human operators for last-mile and complex situations. But we quickly realize how much of the cognitive world requires the physical world. Advancements in construction techniques are the slowest of all. It requires robotics, and those robotics require enabling technologies like computer vision and world models. Those enabling technologies require their own enabling technologies, like semiconductor progress. Ten years from now, we're still investing in these technologies that unlock the next wave of the cognitive realm.

But then, like manufacturing moving to China, it starts really showing up.

We have time. But we do not have forever.