The imagination gap: when execution is cheap, judgement is the divide

Overview

By early 2026, a frontier model could carry, on its own, a piece of expert work that takes a person more than five hours — METR measured Claude Opus 4.5 at a roughly 320-minute task horizon — and the length of job these models can handle keeps doubling. Competent execution is becoming something you buy by the token. You’d expect that to level the field between people. It is doing the opposite.

This weekly is about a gap that most discussion of AI gets backwards. The interesting divide is no longer between you and the machine. It’s between people — and it’s widening.

By the end you’ll have the reframe that matters for your own standing, the reason more capability sharpens the divide rather than softening it, and where professional advantage is quietly migrating.

The content

Start with what’s getting cheap. For three years the hard part was getting a model to do the thing at all, so the skill — and the job titles — formed around execution: the right prompt, the right tool, the careful operation. That era is closing. MIT’s NANDA project, surveying business use through 2025, found AI had “won the war for simple work”: most people now reach for it to draft the email or the first-pass analysis. METR’s measurements show the frontier handling work that runs to hours, not minutes. Getting competent output is turning into a utility — something nearly everyone can buy, like electricity. And a utility is not where advantage lives.

The intuitive next step is that this levels the field: hand everyone the same capable assistant and the gaps between people should shrink. On narrow tasks with a clear right answer, that is roughly what happens — the field experiments tend to find AI lifting the least-experienced the most. But that is the easy half of the work, and it is not where this is going. On open-ended work — the kind where you have to decide what to do, not just how — the same tool cuts both ways. A 2026 review of generative AI in entrepreneurship frames it as a “double-edged sword” that can “both empower and entrap”: the same system that sharpens one person’s thinking can, in the authors’ words, “erode critical thinking, learning, and memory” in another. Whether it empowers or entraps turns less on the model than on the person holding it.

So the binding skill is no longer operating the model. It’s judgement — knowing what the thing should be aimed at, which problems are worth solving, and what “good” looks like when the output lands. Nate B Jones, who writes on AI for a living, hit the new limit and named it from the inside this June: “The wall I hit wasn’t the model running out of ability. It was me running out of work I knew how to hand over.” The constraint moved off the machine and into the person. Two people with the identical model now produce wildly different value, and the difference isn’t how well they prompt. It’s what they can see to do.

This is the imagination gap — and it is worth being exact about what it is not. It isn’t the distance between what you can think and what a machine can do. It’s the distance between people: the difference in their ability to imagine what AI should be doing, and to point it at the work that is actually worth it. It is a judgement gap. And here is the part that should change how you read the next model release — it widens as the models get stronger.

The reason is simple once you see it. When the tool could barely execute, the gap between a sharp instruction and a vague one was small: both produced something mediocre, and you fixed it by hand. When the tool can execute almost anything you can specify well, that same gap becomes the difference between a finished, valuable result and a confident, useless one. Power multiplies the quality of the aim behind it. So every gain in capability stretches the distance between the people who know what to point it at and the people who don’t. A better model does not close the gap between workers. It opens it.

And the people on the right side of that gap don’t just do the old work faster. They reset what the work is. When a capable model makes a whole class of once-too-hard or too-expensive jobs suddenly doable, someone has to imagine that those jobs are now worth doing — and then rebuild the role around them. That is the recalibration, and it is where leadership is moving: the highest-value people aren’t winning by out-executing anyone, they’re winning by changing the target. They decide what a role should now produce in a world where the doing is cheap, and everyone else’s expectations reset to meet them.

Two honest qualifications, because the strong version can mislead. First, judgement here is not a mystical gift — it is the most learnable thing in the whole shift. It is built from depth in a domain, from knowing which problems in your field actually matter, and from the plain ability to tell a good answer from a plausible-sounding wrong one. Without that grounding, “aim” is just opinion, and a stronger model will carry your bad aim further and faster. Second, the levelling on routine work is real and good: if AI lifts the least-experienced on bounded tasks, that is a gain worth having. The widening is specific to the open, judgement-heavy work — which happens to be the work that decides who leads.

Additional reading

• Time Horizon 1.1 — METR (29 Jan 2026) — frontier models now carrying multi-hour work (Claude Opus 4.5 at ~320 minutes), and the doubling continuing.
• Measuring AI Ability to Complete Long Tasks — METR (19 Mar 2025) — the original “50%-task-completion time horizon” and its caveats.
• The “shadow AI economy” is booming — Fortune on MIT NANDA (19 Aug 2025) — AI has “won the war for simple work,” while the work that matters still goes to people.
• Generative AI Use in Entrepreneurship: an Empowerment-Entrapment Framework — Lu, Zhao & Zheng (2 Apr 2026) — GenAI as a “double-edged sword” that can both empower and entrap, depending on how it’s used.
• Claude Fable 5: The Skill for Handing AI Whole Jobs — Nate B Jones (23 June 2026) — the limit moving off the model and into the person; the line about running out of work to hand over (partly paywalled; the quoted line is in the free portion).
• Anthropic Economic Index: Learning curves (24 Mar 2026) — real AI use broadening to a far wider range of people and tasks, as execution becomes ordinary.

Editor’s note

I am a believer that there are people who “do the work”, and people who “pursue the goal”. In a snapshot, the two can look the same, but they are fundamentally different. One is executing a task because they know that the task is supposed to be executed. The other is executing the same task because they know that it serves their ultimate goal. When a task, a workflow, or even an entire operating model, can be performed using an AI assistant, the latter camp are the population that will reshape the way the work is done. That is the imagination gap. It comes down to the ability to judge what is worth the resources, including where the answer is something that didn’t previously exist. Practise this presence of mind in your daily work, and your capacity to create using AI will grow.