cankun.me
Jun 14, 2026 · 4 min read /

What Survives Scale

What becomes worthless as models scale, what does not, and why you should build the part that scale cannot eat.

A stone survey marker remains anchored to bedrock as a vast tide strips away fragile scaffolding

Most decisions about what to build in AI right now come down to a single question that people rarely ask plainly: as model capability grows, what becomes worthless, and what does not?

Rich Sutton's bitter lesson is still the clearest statement of the threat. Over and over in the history of AI, general methods built on computation, search, and learning eventually overtake methods that encode human understanding of a problem. Researchers can never quite resist pouring their domain insight into a system, because it works in the short term. In the long term, that same insight becomes a ceiling, and a method that simply scales compute walks right past it.

If that is true, the only rational thing to do is to build the part that scale does not eat. That requires knowing how to tell the difference, and the obvious way to draw the line fails in an instructive way.

The first cut, and why it fails

The tempting distinction is between a method and the structure of the world: encoding a method gets eaten; encoding the structure of the world does not. A workflow that says "first do literature review, then generate hypotheses, then critique" encodes a method, a human prior about how the work should proceed, and a stronger model will dissolve that workflow by decomposing the task itself. By contrast, the fact that conclusions rest on evidence, that evidence has a source, and that state changes over time is world-structure, so it stays.

This is a good heuristic, but it is not a law, and it fails exactly where you need it most. The trouble is that world-structure itself gets eaten. The history of science is a graveyard of confidently held ontologies. Phlogiston, the ether, and vital force were all treated as world-structure, accounts of what kinds of things exist and how the world is carved at the joints, and every one of them was overturned. Saying that "ontology survives, method dies" claims too much, because ontologies die all the time.

The better cut

What actually survives is narrower and stranger than "world-structure": it is the class of epistemic constraints. These survive not because they describe the world correctly, but because a stronger model does not make them unnecessary.

Take the rule that you must not let something unverified look verified. A model a hundred times more capable does not retire that rule, because it still cannot collapse the gap between a map and the territory; like everything else, it lives on the map side of that gap. Stronger models make better maps, yet a better map does not make the distinction between "agreed" and "true" obsolete. That distinction is about the fact that the map is a map, not about how good the map is.

The useful distinction, then, is between a domain shortcut and an epistemic constraint rather than between method and world-structure. A domain shortcut encodes how some problem is currently solved, and it gets absorbed. An epistemic constraint encodes the irreducible structure of not-yet-knowing (provenance, the difference between consensus and truth, and an honest accounting of what rests on what), and it does not get absorbed because no amount of capability removes the condition it responds to.

Drawing the line here also explains why the first cut failed. Ontologies get eaten when they are really domain shortcuts wearing the costume of world-structure: phlogiston was a particular era's best guess about combustion, dressed up as the nature of fire. The constraints that survive are the ones that encode the gap between any model and the world, rather than any particular model of it.

The unstable corner

Even this cleaner cut rests on a bet in one place, and honesty requires naming it.

Epistemic constraints survive because the map-territory gap is permanent, and the gap is permanent because some verifiers are irreducible: there exist questions whose ground truth requires the real world, in real time, at irreducible complexity, to answer. If the territory could ever be fully simulated, so that a good enough model could always stand in for reality's reply, the gap would slowly close. "Agreed" would creep toward "true," and the surviving constraints would become as quaint as a discarded ontology.

Everything here therefore hangs on a single empirical wager: that the territory contains a residue no map can fully replace. There is reason to believe it does. Any simulation is itself a map, so using simulation to stand in for the territory just uses one map to verify another, which is circular. Still, that is a structural argument from a premise about the world, not a theorem, and in some specific domain it could be overturned by a single empirical fact.

The discriminator works, and what it protects stays durable, as long as the territory keeps that residue. It is worth betting that it does, provided we hold the bet as a bet.

Why this matters for what you build

In practice, this means the most defensible thing to build is rarely the most impressive-looking thing. A system that autonomously does the whole job is almost always a stack of domain shortcuts, and those shortcuts look necessary right now precisely because they fill a gap in general capability. That gap is closing, and those shortcuts are nearing their expiration date.

The durable part is the scaffolding that a more capable model would still need: an honest record of where things stand, a refusal to let the unconfirmed pass as confirmed, and a structure that lets work accumulate without corrupting itself. That work is less exciting, but it grows more valuable rather than less as the model gets stronger, because a stronger model produces more work that needs to be kept honest.

Seen this way, the deepest version of the bitter lesson may go beyond the observation that scale beats hand-coded knowledge: with each capability jump, it also reveals which of the things we called structure were only scaffolding we built because we did not have enough compute to do without it.