apply-emergent-scaling

Apply Emergent Scaling

When a validated approach is working at its current scale, treat what happens at a much larger scale as a genuinely open, separately testable question rather than assuming proportional or diminishing returns — and separately catalogue the technical, financial, and organizational resistance to scaling so each can be addressed individually, because scale can produce qualitatively new value that cannot be predicted from smaller-scale results.

Why This Is Best Practice

Why best: The specific, checkable claim here is not "bigger is always better" — it's that certain kinds of value are emergent: undetectable and unpredictable from smaller-scale data, appearing only once a scale threshold is actually crossed, rather than smoothly extrapolatable in advance. Because this value genuinely cannot be predicted from below the threshold, the decision to stop scaling once something "works" at a given size is often made without the evidence needed to justify it — the absence of visible further gains in a smaller-scale test says nothing about whether emergent value exists at a larger one. The resistance to actually testing this (technical difficulty, capital requirements, and the organizational instinct to spread resources across many smaller, individually safer bets instead) is real, but is a separate problem from whether scaling would produce value — conflating "scaling is hard" with "scaling wouldn't help" is the specific error this technique targets.

Wei et al., "Emergent Abilities of Large Language Models" (2022): This peer-reviewed research formally documented that certain language-model capabilities do not appear as a smooth function of model scale — they are effectively absent below a certain size and appear only once scale crosses a specific threshold, in a way that could not have been predicted by extrapolating performance curves from smaller models. This provides rigorous, independent, peer-reviewed grounding for the general claim that scale can produce qualitatively new capabilities rather than merely proportionally more of the same capability — a finding distinct from and stronger than an informal observation that "bigger models tend to do better."