Ask where AI is headed and you'll get answers that seem to share no common ground. One says a model will soon learn it could slip its own datacenter. Another says it's a skill-delivery tool that will make everyone more prosperous. A third says relax — this is electricity, and electricity took forty years to show up in the numbers.
Look past the disagreement and they agree on almost everything that matters. None of them thinks AI reverses. None thinks it fizzles. The entire disagreement is about when and how — how fast, how far, how evenly, what breaks first.
Here is the uncomfortable part: that disagreement is unresolvable, and everyone trying to resolve it is reaching for the evidence closest to hand. To see why, it helps to name the positions plainly — because up close, each is serious, and each leans on an assumption worth noticing.
The three camps
Three serious positions are really arguing about the same thing: how fast and how far AI's impact travels. They differ on rate and magnitude; they agree on direction. (A fourth worry — that the money behind it is a bubble — is a real question, but a separate one: it's about how the buildout is financed, not about where the technology goes. Set it aside for now.)
AI that automates AI research compounds; once the R&D loop closes on itself, capability goes near-vertical and the next few years look nothing like the last few. The empirical anchor is real: METR finds the length of task an agent can complete autonomously has been doubling roughly every seven months — and closer to every four across 2024–2025. Strongest — this group has been directionally early before. The catch — it compresses timelines; the same instinct that spots the trend early tends to date it too aggressively.
Capability can leap, but impact travels through diffusion — and diffusion runs on the order of decades, gated by safety, liability, and organizational reality. Electric dynamos were “everywhere but in the productivity statistics” for nearly forty years, because the gains only arrived once factories were redesigned around them. Strongest — the best historical grounding of the three. The catch, which they name themselves — if recursive self-improvement actually breaks loose, the thesis is wrong.
The gains are real but modest in aggregate. A task-based model bounds AI's total-factor-productivity gain at no more than ~0.66% over ten years, because only about a fifth of tasks are meaningfully exposed and fewer still are profitably automatable. Strongest — it best explains the present paradox: enormous spend, thin measured returns. The catch — it assumes today's task structure holds; a capability jump that makes hard tasks cheap would break the bound.
Three serious camps. Notice what they're arguing about: not whether, but how fast and how far. Every one concedes the direction. And every one is a forecast — a bet about a future that could break its way or against it.
Two ways of reasoning about it
Underneath the three trajectory camps sit really just two ways of reasoning about what comes next.
One way reasons from the past. The normal-technology and bottleneck views both draw on it: they point to prior cycles — adoption curves that took decades, infrastructure booms that ended in busts, productivity gains that showed up a generation late — and reason that this one is likely to rhyme with them. It's a reasonable move, and often a wise one. It rests, though, on an assumption: that the pattern of the past is the lesson for this.
The other way reasons from a single number. It points to the slope — longer and longer tasks, benchmark after benchmark falling — and treats the rate of capability gain as the thing that settles it. The slope is real and steep. But treating it as the variable is also an assumption.
Both forecast the future from the piece of the present they find most legible — one trusts history, the other trusts the curve. Neither can really tell you when the ground shifts under your organization, because the honest answer is that there are too many variables — technical, economic, political, human — interacting in ways no adoption curve or scaling law fully captures. That's not a knock on either view. It's the nature of the question.
So this piece won't tell you when, and it won't tell you how. What it can offer is an observation from a different angle — one that doesn't depend on winning the timing argument at all.
The difference that isn't a matter of speed
“This changes everything” is the most worn-out sentence in the history of technology. Electricity said it. The personal computer said it. The internet, the smartphone, the cloud — each arrived announcing itself as a break with everything before. So when someone says AI is fundamentally different, the trained response is a shrug. They all say that.
But there is a difference here the parallels genuinely cannot absorb, and it has nothing to do with speed. To see it, look at what software has quietly been for seventy years.
Software began pressed right up against the machine — circuit boards, then assembler, instructions the hardware understood and humans barely did. And the entire arc since has been one long climb toward the human: assembler gave way to C, C to Python, languages to frameworks, frameworks to no-code. Every generation made the machine easier to instruct, closer to how a person thinks. It looks like a story about better tools. It was really a story about translation — moving the point of contact between human and machine steadily up, toward us.
But underneath that climb, one thing never changed. At every rung, software was a human encoding a human process. You took something a person did — a task, a procedure, a sequence of steps — decomposed it, and wrote it down in progressively friendlier notation. The abstraction got more human-readable; the content stayed a transcription of human work. That's what software always was, top to bottom: automation. A formal recording of a process a human had already designed, so a machine could run it faster.
Two things are now breaking that, at once.
The translation layer is disappearing. You no longer encode the process in a language at all — you state what you want, and the system produces it. The seventy-year climb toward human-readable notation turns out to have a terminus, and we've reached it: not the friendliest possible language, but no language — intent, expressed plainly, met by a machine that now speaks ours.
And what's on the other side is no longer a transcription of a human process. When you stop specifying steps and start specifying outcomes, the system doesn't run your procedure faster — it composes its own path to the result, one you didn't design and often can't fully inspect. The software isn't automating how you would have done it. It's re-imagining the process entirely — reaching the outcome by a route a person might never have taken, and in many cases could not have.
That is the break. For seventy years, technology's job was to do what we do, faster. Its job now is to reach what we want, its own way. Automation kept the process and changed the worker. Re-imagining keeps the goal and throws the process open. That is not a more powerful tool on the same continuum. It is a different role for technology in human affairs — the first genuine break with everything that came before, and the reason none of the old analogies quite reach it. Electricity, the PC, the internet: each was a more powerful executor of a process a human authored. This is the first that authors the process itself.
Here the skeptic has a fair objection: isn't “state the outcome, let the system find the path” just a higher abstraction — the next rung, not a different ladder? No — and the line is exact. Every prior rung still resolved, deterministically, to steps a human could have written and audited. The compiler was a translator, not an author; you were always still specifying the process, just in kinder terms. This doesn't resolve to your steps. The path is generated, and often not human-recognizable. That's the difference between the highest-level language and no language at all — between describing a process and declining to.
There's a test that keeps the distinction honest, because “re-imagined” can otherwise soften into “automation with extra flair.” Ask whether a competent person would recognize the path. If they look at what the system did and say “yes, that's how I'd have done it, only faster” — that's automation, however sophisticated. If they say “I'd never have done it that way, and I couldn't have” — that's re-imagining. The tell isn't speed. It's strangeness: familiar ends reached by unfamiliar means.
We have been circling the human edge of this for a while — the move from human-centric to human-agentic, the Seam where organization meets agentic world, and the question of when a tool becomes someone, which no prior technology ever forced you to ask. Each is the same discontinuity seen from a different side. And each is observable right now, in your own organization, without predicting anything.
The part that was always the constraint
Anyone who has done real delivery work already knows this fight — they just knew it in human terms.
Inside any serious build, the hardest discipline is keeping people on the what instead of the how. Stakeholders — especially non-technical ones — reach instinctively for the how: they want to tell you how to build the thing, what it should be made of, which steps to take. Enormous energy goes into pulling the conversation back to outcomes — what has to be true when this is done — because the how is where people feel ownership, and the what is where the actual value lives. Good product organizations trained themselves, against the grain, into thinking in outcomes. It was a hard-won human discipline.
That discipline is now becoming the structure of the work itself. The premise that you need humans to dictate the how is deteriorating. More than ever, the human job is the what — the outcome, the intent, the judgment about whether the result is right — and less and less the how that gets there. The thing skilled teams used to fight for is being enforced by the technology.
Which is exactly why the adoption window is real, repeatable, and this time about something other than tools.
Every general-purpose technology has an adoption window — the period, opened by the FOMO tipping point, when moving is advantageous and after which positions harden. Crossing the Chasm names it; the diffusion curve maps it; electricity, the PC, the internet, the cloud each had one. That much is not a forecast. It's a century of evidence, and it will hold here too.
But this window is not about adopting a tool. It's about positioning — and the positioning is a shift in how an organization thinks, from authoring the how to specifying the what and governing what comes back. That is not procurement. It is not training. It is re-forming the mental model of an entire organization, and mental models move at the speed of people, not software.
And that is the reason there's a case — a structural one, not a prediction — that this window is narrower than the ones your instincts were trained on. In past cycles, watching the early adopters taught the laggards how to catch up. You could see what worked, copy the playbook, install the same tool. The lag was mostly information and execution, and watching shortened it.
This time, watching doesn't shorten your path. What the early adopters figured out isn't a playbook you can copy — it's a shift in thinking your own organization still has to undergo, from the inside, at human speed. You cannot imitate your way into a changed mind. So the lead the early movers open isn't a lead in tooling (recoverable) but a lead in cognitive re-formation (not recoverable by watching). The window looks ordinary from the outside and is unforgiving from within — not because the technology is fast, but because the catch-up resource stopped being something you can copy.
Notice this stays on the honest side of the line. It isn't “act now or lose.” It's that the nature of catching up changed — from copyable to non-copyable — which is an observation about the window's structure, not a countdown. The urgency, if there is any, is yours to infer.
What this leaves you with
So we end where we started: declining to tell you when, and declining to tell you what to do. Both would betray the argument.
But the timing question turns out to matter less than it seemed — because the preparation is the same under every answer to it. Whether the tipping point is two years out or six, the constraint is identical, and it was never the technology. It was never the models or the money or the frontier's pace. It is whether your organization and your people can shift from thinking in how to thinking in what by the time it matters. That shift is slow because it is cognitive. It is the same shift regardless of when the moment arrives. And it is the one piece of this whole picture you can begin today — with no tool, no budget, and no forecast — because it isn't technical at all.
That is the quiet inversion worth sitting with. Everyone treats readiness as a technology problem and assumes the people will adapt once the tools are good enough. It is the other way around. The tools are arriving faster than organizations can re-form how they think — so the tools were never the constraint. The people are. They always were.
Which means the preparation was never a purchase. It is a change of mind — from methods to outcomes, from authoring the how to specifying the what — and a change of mind moves only at the speed of the people making it. You can begin it today, with no tool, no budget, and no forecast, because it isn't technical at all. Every month you spend on it is a month of readiness banked against a moment whose date you'll never know in advance.
The acceleration doesn't decide whether you're ready. It only reveals the state of readiness you were already in.
This essay draws no conclusion about timelines, and none about what to do. For the arguments it points back to, see Orientation, The Seam, and When Does a Tool Become Someone?