Jared Kaplan and colleagues, 2020

“Performance has a power-law relationship with each of the three scale factors: model size, dataset size, and computation.”

Jared Kaplan and colleagues studied how language-model loss changes as model parameters, training data, and computation increase. Their 2020 paper reported smooth power-law trends across the ranges they tested. That regularity suggested that researchers could forecast performance and allocate resources more deliberately instead of treating each large training run as an isolated experiment. The work helped turn scaling into an engineering discipline with measurable tradeoffs. Its force comes from connecting a technical result to a wider question about how intelligent systems are built, evaluated, and placed in human institutions.

A power law does not say that bigger is always better without limit. It describes an empirical relationship under particular architectures, data, optimization methods, and metrics. The three resources must be balanced: an oversized model trained on too little data or too little compute may waste capacity. Later work revised some optimal ratios, showing that scaling laws themselves improve through new evidence. Their value lies in prediction and planning, not in becoming an unquestionable rule. This interpretation avoids treating intelligence as a single score. Performance depends on the task, data, environment, objective, and people who must live with the outcome.

Scaling analyses now guide expensive decisions about datasets, chips, training duration, and expected returns. They can prevent waste by showing where an additional dollar or unit of compute is most useful. Public policy can also use them to reason about infrastructure concentration and resource demands. A responsible plan includes energy, water, supply chains, and opportunity cost rather than treating compute as an abstract quantity detached from communities and the environment. The quotation offers a lens for current AI debates: ask what changed technically, what remained uncertain, and whether the surrounding organization can manage the consequences of error.

Lower training loss does not automatically imply safer, fairer, more truthful, or more socially valuable behavior. Some risks emerge only at deployment, and some capabilities are difficult to forecast from aggregate curves. Teams should pair scaling predictions with capability evaluations, misuse analysis, and decisions about whether the task deserves more scale at all. Efficiency gains should be compared with smaller models, better data, retrieval, and improved system design. Good governance does not reduce ambition; it directs ambition toward results people can trust. Clear ownership, meaningful appeal, and honest reporting turn a clever system into accountable infrastructure. Kaplan’s result made growth more predictable, but prediction is not purpose. The central governance question remains which capabilities merit the resources required and how their benefits and burdens will be shared. The most valuable legacy is a better question, not a larger slogan: what evidence would show that this idea improves human outcomes, and what evidence would tell us to stop? The work continues with care.