Construct · 6 fields · 16 citations
One name. Different field. Different construct. One map.
● jingle · one name, six constructs Nearest construct by meaningscalecompare the meaningsEvery construct on the map is compared with every other by the meaning of its definitions, and the closest one is surfaced here. The nearest construct to scaling is scale.
Hold this as a prompt, not a verdict. Two constructs sitting close in meaning is a reason to read them side by side and decide for yourself whether they are one idea under two names. It is not proof that they are. Compare the meanings in the Collider.
Spokes are fields; dashed arcs are shared lineage. Tap a field for its definition, unit, and measures.
Drag to watch the word get claimed. Each field appears at the publication date of the earliest source cited for it. Tap a year to see which.
The empirical regularity that a language model's test loss falls as a power law in the number of parameters, the size of the training set, and the compute spent, over several orders of magnitude, with architecture mattering little by comparison. Kaplan and colleagues fixed the laws in 2020 and drew the practical conclusion that compute is best spent on larger models trained on relatively few tokens. Hoffmann and colleagues revised the recipe in 2022 with Chinchilla: at a fixed compute budget parameters and tokens should grow together, so earlier models were undertrained. Sutton's 2019 essay gave the programme its creed, general methods that use more computation beat methods built on human knowledge, and researchers have learned this lesson repeatedly and reluctantly. The construct is a fitted exponent, and its use is deciding how to spend the next budget.[1][2][3]
The ability of a system to handle more work when given more hardware, and the limits on that ability. Amdahl argued in 1967 that the serial fraction of a program caps the speedup any number of processors can deliver; Gustafson replied in 1988 that in practice the problem grows with the machine, so the serial fraction shrinks and speedup can stay near linear. Hill's 1990 note observed that the field used scalability without a definition and offered one in terms of efficiency as processors are added. Practice splits the word again: vertical scaling adds capacity to one machine, horizontal scaling adds machines. The unit is a workload on a system, and the criterion is throughput or latency as a function of resources.[4][5][6]
The growth of an organisation's activity, reach, or impact at a rate that outpaces the growth of its resources, and the organisational changes that make such growth survivable. Shepherd and Patzelt's 2022 call for research fixed the academic construct: scaling is distinct from growth in that it concerns how the organisation replicates and spreads what works, and it applies to social impact as well as revenue. Hoffman and Yeh's 2018 Blitzscaling is the practitioner version, prioritising speed over efficiency in the face of uncertainty, accepting operational chaos to capture a market before rivals do. The unit is a firm, the clock is years, and the measure is customers or revenue per employee, not loss per FLOP.[7][8]
The assignment of numerals to objects or events according to a rule, and the question of which arithmetic the resulting numbers will bear. Stevens fixed the vocabulary in 1946: nominal, ordinal, interval, and ratio scales, each defined by the transformations that leave its information intact and each licensing a different set of statistics. Thurstone's 1927 law of comparative judgment showed how to build an interval scale from pairwise choices, and Likert's 1932 summated ratings gave the survey item its standard form. Kruskal's 1964 multidimensional scaling recovers a spatial configuration from judged similarities so that distances reproduce the data as well as a monotone fit allows. Scaling here is measurement itself (nothing gets bigger) and the field's disputes are about whether a sum of ratings is interval or only ordinal.[9][10][11][12]
The relationship between an organism's size and its other properties, usually a power law with an exponent that is not one. Kleiber found in 1932 that metabolic rate across mammals rises as body mass to the three-quarter power, not the two-thirds that surface area would predict, and the exponent has held from mice to elephants. West, Brown, and Enquist derived it in 1997 from the geometry of branching transport networks (a fractal supply system that fills space and minimises the energy of distribution) and predicted quarter-power scaling for heart rate, lifespan, and growth. The unit is a species or an individual on a log–log plot, and the exponent is a fact about physiology that a theory has to earn.[13][14]
The behaviour of a system that looks the same at every magnification, so that its properties near a critical point follow power laws whose exponents depend only on dimension and symmetry, not on the material. Kadanoff's 1966 block-spin argument gave the picture: group neighbouring spins into blocks, treat the blocks as spins, and at the critical temperature the description does not change. Wilson turned the picture into the renormalisation group and his 1979 account for a general audience explains why fluctuations at every length scale must be handled together and why unrelated systems share exponents. Scaling here is a symmetry, the exponents are universal, and a scaling law is a statement about what happens when the ruler changes.[15][16]
Next
See how scaling sits against neighbouring constructs in the Collider, or export every reference on this page in one click.
Open in ColliderExport BibTeXReport this collisionCited on this page?
If your work is cited behind one of these entries, you know better than we do whether we captured the meaning right. Confirm it, suggest an edit, or dispute it, and your name goes on the entry as an author who has checked it, linked to your ORCID.
Claim your construct