We keep calling it a boom. The “AI boom,” the “AI bubble,” the breathless question on every earnings call about whether the whole thing is overhyped and about to pop. It’s a comforting frame, because booms end. You can wait out a boom. You can keep your head down, let the hype burn off, and get back to normal once the bubble bursts.
I don’t think this one bursts. A boom is a market event, and what’s happening underneath is not a market event. Intelligence is moving into everything, the way electricity once did, and electricity never popped. It just became the floor everyone stood on. The right frame isn’t a bubble to survive. It’s a migration to live through, and then a new condition to live inside.
So I want to retire the boom question and ask a different one. Not “is this overhyped?” but “how do we live alongside something that can increasingly do what we do?” That question is less thrilling and far more important, and almost nobody is sitting with it honestly. This essay is an attempt to sit with it, from the silicon all the way up to what it does to a single human life.
If you want to know how serious and how permanent this is, don’t look at the chatbots. Look at the silicon.
What Computex revealed
At Computex 2026, the most revealing thing about the future of AI wasn’t said by an AI lab. It was said, again and again, by the companies that make the metal. On Qualcomm’s stage, Cristiano Amon declared 2026 the “Year of the Agent” and described a move from user-driven computing to agent-driven computing: machines that act on their own instead of waiting to be asked. In a two-hour keynote, Nvidia’s Jensen Huang said the same thing in his own dialect, calling tomorrow’s agents sophisticated users of tools and putting the Vera Rubin platform into production to run what he called “agentic AI factories.” A day later he shared a stage with Marvell’s Matt Murphy under one banner, “The Future of AI Depends on Connectivity,” because agents, distributed by nature, are now what drives the demand. Even AMD, not on the keynote stage this year, sent Lisa Su to Taipei with a processor that runs a 300-billion-parameter model entirely on a laptop.
These are rivals who agree on almost nothing. This year they agreed on the noun. Not chips that run faster. Chips built to run agents.
That convergence is the signal worth watching, and it’s easy to miss if you only follow the app layer. Hype lives in demos. Conviction lives in roadmaps. A silicon roadmap is a three-to-five-year bet, measured in billions of dollars and in fabrication capacity that cannot be unwound. Chipmakers do not re-architect for a fad. When all of them turn toward the same idea at once, the ground underneath has already moved.
Look closer and they agree on the shape of it too. Amon showed an agent finishing a task by splitting the work between the device and the cloud, reaching the same result with thirty percent fewer tokens at a quarter of the cost. Nvidia paired its hyperscale Vera Rubin with the RTX Spark, a personal superchip built, in its own words, for the era of personal AI agents. AMD’s headline was a frontier-sized model running with no cloud at all. The architecture is the same everywhere: small, fast, private models on the device, handing the heavy reasoning up to larger models in the cloud. Local for reflex, cloud for thought. Lisa Su named the scale of what’s coming plainly: the world will need on the order of a hundred times more compute as the number of AI users climbs from one billion toward five. AI stops being a destination you visit and becomes ambient, something simply there, the way electricity stopped being a product and became a wall socket.
It’s worth being precise about what that word, agent, actually changes, because it isn’t mainly about intelligence. It’s about who runs the loop. A chatbot waits. You prompt, it answers, and then it stops and waits for you again. You are the engine: you set the goal, break it into steps, choose the tools, judge the output, and stitch the pieces into something whole. An agent holds the goal instead. It breaks the goal into steps on its own, calls its tools, watches what comes back, corrects, and keeps going until the thing is finished. And then it acts: it sends the email, writes to the database, moves the money, files the ticket. Words become consequences.
Describe it plainly and the structural change is unmistakable. The human moves from inside the loop to outside it. We used to orchestrate while the machine assisted; now the machine orchestrates while we supervise. That sentence is the hinge of everything that follows, because once the machine owns the loop, the first thing at risk isn’t a job. It’s the way human beings become capable in the first place.
What it means for humankind
Start with the thing we say least often, because it sounds inefficient: judgment can’t be downloaded. It has to be accreted.
Think about how anyone actually becomes good at cognitive work. The junior arrives knowing the theory and almost nothing else. We hand them the grunt work: the boilerplate, the first-pass analysis, the dull debugging, the thousandth small ticket. They do it, badly at first, then less badly. And somewhere in those thousands of repetitions, without anyone able to point to the moment it happened, the pattern-recognition forms. They begin to smell the bug before they find it. They develop taste. One day they can look at a piece of work and know, fast and without quite being able to explain why, that something is off. We call that seniority and treat it as a credential. It isn’t. It’s sediment. The senior’s judgment was paid for, in advance, by years of doing exactly the work we are now handing to agents.
That’s the paradox, and it has teeth. Agentic AI is best at precisely the well-scoped, repetitive, entry-level work that has always been the training ground. So we automate the bottom rung of the ladder, even as we need, more than ever, the judgment at the top to supervise the agents and catch them when they go confidently wrong. We keep the requirement for expertise and quietly demolish the only road that ever led to it.
We are sawing off the bottom rungs while standing on the top.
This isn’t a new shape of problem; it’s the largest instance of an old one. Aviation learned it the hard way. As autopilot grew more capable, pilots flew less by hand, and their stick-and-rudder instinct thinned, until the rare crisis handed control back and the skill that should have been there had quietly eroded. In 1983 the researcher Lisanne Bainbridge named the pattern: the ironies of automation. The better the automatic system, the more it strips away the routine practice that keeps the human sharp. Then it leaves that same human responsible for exactly the hard cases the machine can’t handle, with exactly the skills the machine let rust. The guilds understood the truth from the other side, which is why mastery took years of doing and not a manual. So does medicine: see one, do one, teach one. The doing was never decoration. It was the mechanism.
And I’ll say what I see, doing this work and leading people who do it: the erosion is already here, and it doesn’t arrive looking like decline. It looks like productivity. Capable people ship more, faster, leaning on the machine, and they are genuinely more productive today. What’s harder to see is the practice that isn’t happening underneath: the wrestle that never occurs because the answer arrived first, the scar that never forms because the friction was removed. Taste is downstream of scar tissue. Take away the friction and you keep the output while losing the formation.
Now the honest objection, the one you’re already forming. We’ve heard this before. Every wave of automation killed some rung and grew new ones. Weavers feared the loom; the calculator didn’t end mathematics; spreadsheets didn’t end finance, they made more of it. The ladder didn’t vanish. It moved up. The optimists were right every time, and they may well be right again: perhaps the apprenticeship simply relocates, and tomorrow’s juniors learn to direct and judge fleets of agents instead of doing the steps by hand.
That’s the strongest version of the counterargument, and I want to concede most of it. The ladder probably does move up. But here is the discontinuity I can’t argue my way out of. Every previous wave automated the doing and left us the judging. The loom took the weaving and left the design. The spreadsheet took the arithmetic and left the analysis. In each case the machine absorbed the labor, handed the human the higher task, and left intact the lower rungs where that higher judgment was built. Agentic AI is the first wave that automates the judging-by-doing itself. It doesn’t only take the task; it takes the apprenticeship to the task. Previous automation changed what we practice. This changes whether we practice. That isn’t the same animal in new clothes. It’s a new animal.
This can sound like a philosopher’s worry, abstract and far off. It isn’t. Follow it one step out of the mind and into the org chart, and it stops being philosophy and becomes a hiring plan.
What it means for companies and workers
If you automate the bottom rung, the consequences aren’t abstract. They show up in hiring plans, org charts, and the arc of individual careers. Let me keep the scope honest. I’m talking about cognitive work: software, law, finance, analysis, design, the knowledge trades where agents bite first. The same logic radiates outward from there.
Start with the coldest link in the chain. Why train a junior for two years to reach competence when an agent produces the junior’s output now, cheaper, instantly, without holidays or attrition? Asked one hire at a time, the answer is merciless and obvious, and you can already watch firms quietly arriving at it: graduate pipelines thinning, “do more with less” hardening from slogan into policy.
Run that logic across a whole organization and the shape it produces is a barbell. A thin band of senior people at the top (the ones with judgment, with taste, with the standing to be accountable), supervising fleets of agents below them, and almost nothing in between. The old pyramid, with its wide base of juniors narrowing through a thick middle to a few at the peak, collapses. The base is automated. The middle, where people used to spend a decade climbing, thins toward nothing.
That barbell is a photograph. Play it forward and it becomes something worse. The seniors holding the top together can do that job because they came up the old way: they did the steps, they earned the sediment. But the organization has stopped producing their replacements. We still need seniors; we’ve simply stopped minting them. Judgment turns into a non-renewable resource. For a while you don’t feel it, because the existing experts are right there, more leveraged than ever, more valuable than ever. And then they retire, and you reach for the next generation, and find a generation that began its career outside the loop and never accumulated the thing you’re now asking it to spend. That’s the expertise cliff. The barbell is the picture; the cliff is the picture played forward in time.
Underneath all of it, the work itself changes species. Headcount decouples from output: five leveraged people ship what fifty once did. The job stops being production and becomes orchestration, verification, accountability: pointing agents at goals, judging what they hand back, owning the result when it lands in the real world. For whoever keeps a seat at that table, it is, genuinely, a better job.
But notice who is still in the room. The climb that once carried an ordinary person from the bottom toward something better (income, security, and standing, built up across a working life) turns into a wall. You are either above the line, leveraged and scarce and well paid, or below it, competing against software that is already cheaper than you and getting cheaper. Mobility, the quiet engine that made the whole arrangement feel fair, seizes up. And a society where the ladder becomes a wall isn’t only an economic fact. It’s a political one.
All of which leaves one question that is worth, quite literally, a great deal of money, with a harder one standing right behind it.
The million-dollar question
If the machine can do the doing, where does human value come from, and what are we for?
It’s really two questions wearing one coat. There’s the economic one, sharp and immediate: in a world of capable agents, what is the work that still pays, and who gets to do it? And there’s the older, deeper one underneath it: when the things we did to feel useful can be done without us, what is a human life actually for? Nobody has the full answer. Anyone who tells you they do is selling something.
But “I don’t know” is not a strategy, and the people who come through this well will be the ones who act before the answer is clear. Here is where I’d start, at three levels.
For you
Do the reps on purpose, even when you don’t have to. The friction the agent removes was never pure waste; some of it was the very thing that built your judgment, and now you have to seek it out deliberately instead of having it forced on you. Use agents to go further than you could alone, not to skip the part where you become someone. Stay in the loop on the work that forms you, and hand off only the work that doesn’t. And pour yourself into the things that sit above the automatable line: judgment, taste, the nerve to decide under uncertainty and to own what happens next. Learn to direct agents well, because that is the new literacy, but never let directing fully replace doing in the craft you most want to be trusted in.
For companies
Treat the apprenticeship as infrastructure, not overhead. The firm that automates its juniors away wins this quarter and loses the decade, because it is eating its own seed corn, and the bill arrives precisely when its current seniors retire. So design the ladder back in on purpose. Rotate people through the doing even when an agent could do it faster. Protect some of the bottom rungs as a training investment rather than a cost to be optimized away. Make review and mentorship the new classroom, the place where judgment is transferred now that it can’t be absorbed by sheer volume of grunt work. And measure the capability you are building, not just the output you are shipping, because only one of those two numbers tells you whether you’ll still exist in ten years.
For society
When headcount decouples from output and the climb becomes a wall, mobility stops being a private career problem and becomes a public one. Education has to shift from teaching the doing toward teaching judgment and the art of working alongside agents, earlier than feels comfortable. And we will have to be honest about a harder thing: when productivity no longer requires most people, the link between work and worth, which has organized modern life for two centuries, comes under real strain. Holding a society together through that will take more than retraining slogans. It will take policy, a serious safety net for the transition, and a story about human dignity that doesn’t rest entirely on economic usefulness.
Notice that all three levels are the same instruction wearing different clothes: don’t let the machine do the becoming for us. Let it take the toil, not the formation. Let it run the loop where the loop is only labor, and keep our hands on the loop where the loop is how we grow.
We called this an AI boom because boom is an easy word, and easier to wait out. But you don’t wait out a new condition. You learn to live inside it, and how you live determines who you become while living there. The companies have already taken their stance, and cast it in silicon. Ours is still open. The question was never really whether the machines can do the work. It’s whether, once they can, we still choose to become the kind of people who could have.