STIML (Scaling Theory Informed Machine Learning)
STIML — Scaling Theory Informed Machine Learning — is West's late-career methodological synthesis, developed to address the principal limitation of scaling theory: that it predicts ensemble averages but cannot predict individual trajectories. The framework uses scaling laws to capture the 'trend-driven predictability' — the average trajectory that any given company, city, or organism will follow based on its size and network topology — while deploying machine-learning methods to model the 'fluctuation-driven predictability' — the residual deviations that separate individual cases from the average. The framework is an epistemological compromise, acknowledging that scaling laws describe the central tendency but not the specific outcome, and that understanding any individual system requires both the structural prediction from theory and the data-driven modeling of its specific deviations. STIML is also an implicit admission that where individual fate diverges from ensemble average — where specific leadership choices, cultural factors, or accidents of timing matter — the mathematics provides a map but not a route.
In the AI Story
The STIML framework emerged from a tension that had been present in West's work from the beginning. The scaling laws are extraordinarily successful at predicting averages — the mortality curve for the entire population of publicly traded companies, the growth dynamics of cities as a class, the metabolic rate of species across the size spectrum. But they cannot, by construction, predict which specific company will survive, which specific city will thrive, which specific organism will be an outlier.
This limitation became particularly pressing as West's work attracted attention from business audiences who wanted actionable predictions about their company. The framework could tell them that half of all publicly traded companies die within ten years — but not whether they would be in the dead half or the surviving half. The gap between ensemble prediction and individual prediction became the central conversation in applying scaling theory to real-world decisions.
STIML addresses this gap not by eliminating it but by formalizing it. The trend, captured by scaling laws, is genuine and powerful — it provides the structural prediction that any individual case will have to overcome or conform to. The fluctuations, captured by machine learning, are where individual decisions, leadership quality, and idiosyncratic factors live. Neither layer is sufficient alone; both together provide the fullest predictive framework available.
For the AI transition, STIML's implications are honest rather than optimistic. The framework predicts that most AI-adopting organizations will follow the sublinear mortality curve, accelerated by their increased metabolic rates. A small minority will deviate — will undergo genuine topological transformation and achieve city-like persistence. The framework cannot predict which specific organizations will be the exceptions. That depends on factors — leadership judgment, cultural courage, institutional design — that live in the fluctuations, outside the scaling laws.
Origin
STIML was developed by West and colleagues in work published in the early-to-mid 2020s, motivated by the need to bridge ensemble-level scaling predictions with individual-case prediction that machine learning excels at. It represents a methodological synthesis characteristic of mature complexity science: combining analytical theory with data-driven methods.
Key Ideas
Two layers of predictability. Scaling laws capture trend-driven predictability (ensemble averages); machine learning captures fluctuation-driven predictability (individual deviations).
Scaling as structural prediction. The scaling law tells you what the average trajectory looks like and why — the structural constraint any individual case must navigate.
Machine learning as residual modeling. ML methods capture the specific factors that cause individual cases to deviate from the average.
Honest about limits. STIML formalizes the acknowledgment that neither approach alone predicts individual fate — both are needed, and even together they leave uncertainty.
Individual choice lives in the fluctuations. Leadership decisions, cultural factors, and accidents of timing are captured by the fluctuation layer, not the trend layer.
Appears in the Orange Pill Cycle
Further reading
- Geoffrey West and colleagues, recent papers on integrated scaling-and-learning approaches (2020s)
- Geoffrey West, Scale (2017), epilogue on individual vs. ensemble prediction