The Medium Run Keeps Moving
Meta-paradigmatic thinking and staying oriented
Matt Yglesias made an excellent observation recently that felt like someone finally naming an increasingly noisy background irritation. I’d encourage you to read the whole article which moves through several topics, but I want to pick something specific out and expand on the idea.
Writing medium-run arguments about policy, institutions, or the labor market is getting harder because AI keeps changing the assumptions those arguments rely on. You can build a coherent story about incentives, time horizons, and second-order effects, and then realize that the story depends on what AI will be able to do in three, five, or ten years. That assumption is no longer a minor detail. It determines whether the argument holds together.
His point is not that analysis is pointless or that knowledge is impossible. It is that the “medium run” is starting to feel like the period in which the underlying conditions could change enough to alter the conclusion. The discomfort is not ignorance. It is baseline instability.
That got me thinking about something larger than writing. It is one thing for arguments to become harder to sustain. It is another for a person to live inside the same condition. If the assumptions that support coherent reasoning are increasingly provisional, what does that do to the mind that reasons and to the life built around that reasoning? How do you stay oriented when the conceptual environment changes faster than you can fully absorb?
The first problem is speed. Not just the pace of new products, but the pace at which expectations adjust. The half-life of competence is shrinking. You can learn a tool, start using it, and quickly find that what counted as skill has become expected baseline. Tasks that used to signal expertise become easier to perform and easier to imitate. People describe this as “keeping up,” but keeping up is not only about learning tools. It is also about updating your sense of what is worth investing in and what kind of contribution you can reliably make.
This is the part that Ray Kurzweil gets right in structural terms. His core claim is that technological progress compounds. Capabilities accelerate because each improvement becomes input into the next. You do not have to accept the full “singularity” narrative to accept that compounding is real and that it changes daily life. Where Kurzweil tends to overreach is in what he assumes about human and institutional adaptation. The underlying story often treats adaptation as smooth. In practice, adaptation is slower and uneven.
That friction matters. Stability is not only comforting, it is also the condition that supports coherent professional standards, shared expectations, and durable public debate. If technological change is accelerating, the question is not only whether we can keep producing new capabilities. It is whether we can maintain stable enough frames of interpretation and coordination to make sense of those capabilities and integrate them without constant disruption.
The second problem is the type of change. Sometimes you are not merely learning a new tool. Sometimes the tool changes the categories you use to interpret what is happening. It changes what counts as impressive, what counts as scarce, what counts as expertise, and what counts as uniquely human work.
This is where Thomas Kuhn is useful. Kuhn distinguished between ordinary progress inside a shared framework and paradigm change, where the framework itself changes. In ordinary progress, people refine within the same basic model of what matters and how to evaluate claims. In a paradigm change, the standards shift. Concepts that used to organize perception become less useful. People can end up talking past each other because they are operating with different assumptions about what the problem even is.
AI is not one single Kuhnian revolution. It is a sequence of smaller, repeated shifts in domains that matter for everyday professional life: writing, analysis, synthesis, expertise, authorship, and the boundary between producing text and exercising judgment. Even if you are cautious about long-term AI predictions, you can still see that many categories people relied on are becoming less stable.
This returns to Yglesias. His observation about writing is, in a limited but real sense, a Kuhnian observation. Medium-run arguments depend on background categories staying stable long enough for causal mechanisms to play out. If the categories are moving, the argument becomes fragile even if the logic is sound.
At this point the practical question is obvious: what do you do with this? You cannot freeze the baseline. You also cannot treat every topic as an AI forecasting exercise.
This is where Quine earns his place because he offers a way to think about updating beliefs without losing coherence. Quine described our beliefs as a web. Some beliefs are near the edges and can be revised with limited consequences. Others are central and support many other beliefs. Nothing is immune to revision, but revision is usually manageable because it is partial and structured.
What destabilizes people is not that beliefs change. It is that too much near the center is pressured too quickly. AI puts pressure on central concepts in modern life: intelligence, value, creativity, expertise, and the status attached to “knowledge work.” If you revise those concepts every time new capabilities appear, you will feel unmoored. If you refuse to revise them at all, you will become rigid and defensive.
So the coping problem is not “stay the same” versus “embrace the future.” It is structured revision. It is knowing what can update quickly without destabilizing you and what must update more slowly if you want to remain coherent.
To name that coping posture, I’ll use the term meta-paradigmatic. By this I mean the ability to operate within frameworks while staying aware that frameworks can change. A meta-paradigmatic person treats frameworks as tools rather than as fixed identities. They expect that paradigms will shift, sometimes abruptly, and they build their identity around capacities that are more portable across shifts: judgment, responsibility, and the ability to decide what matters and why. They still learn tools and track facts, but they do not tie their sense of worth to the current difficulty of tasks that machines may soon perform.
This stance is not detachment. It requires distinguishing between the fast layer of life and the slow layer. The fast layer includes tools, workflows, and local definitions of competence. That layer should update quickly. The slow layer includes values and commitments: what you will not outsource, what you regard as legitimate authority, how you define agency, and what kinds of decisions require human responsibility. If you let the fast layer dictate the slow layer, you will lose your center. If you refuse to update the fast layer, you will lose effectiveness and eventually your ability to act.
This is where Yglesias’s observation becomes more than a comment about writing. Baseline instability does not mean medium-run thinking is impossible. It means you have to state assumptions clearly, separate what depends on uncertain technological trajectories from what depends on durable principles, and build arguments that can survive updates rather than collapse when one forecast changes.
It is also a civic skill. A society that cannot stay oriented under moving baselines will outsource judgment, not out of laziness but out of exhaustion. A society that can stay oriented can adapt without surrendering agency. It can treat AI as a powerful and shifting set of tools rather than as a verdict on human worth.
Kurzweil is right about accelerating capability. Kuhn is right that sometimes the framework changes. Quine is right that coherence requires structured updating. Together they point toward a discipline: revise without self-erasure, act inside provisional models without becoming cynical, and keep values stable even when tools change quickly.
If the medium run is now the period in which the substrate can change, then the goal cannot be to wait for stability before thinking. The affirmative response is not denial and not surrender. It is learning how to remain coherent across repeated updates, and making sure that judgment and responsibility are not the parts you outsource first.