kwj.ai · acquisition inquiries from >$999view prospectus →
The Domesday Book ofKWJ · AI

Era II · 2018–2020

The Scaling-Laws Era

By C.W. Jameson · Published 19 May 2026 · Last reviewed 19 May 2026

Between 2018 and 2020 the field stopped tinkering with architectures and started buying compute. Kaplan, McCandlish, and a small team at OpenAI published the scaling-laws paper in January 2020 that gave the trend a name. GPT-3, released that summer, validated it. The lesson was that capability rose smoothly and predictably with parameters, training data, and compute.

GPT-2

Released in stages through 2019 — 124M, 355M, 774M, 1.5B parameters. OpenAI initially withheld the 1.5B weights citing misuse risk. The risk did not materialise; the weights were released in November 2019. The capability shocked the field. The headline demo was the unicorn-in-the-Andes story.

Scaling laws

Kaplan et al., 'Scaling Laws for Neural Language Models', arXiv 2001.08361, published 23 January 2020. The paper described power-law relationships between loss and compute, model size, and data. The practical implication was that the right way to improve capability was to spend more, not to invent.

GPT-3

175 billion parameters, released June 2020. The API was invitation-only for the first year. The few-shot-prompting demonstration in the technical report changed what most people thought a language model was.

Signature models of the era

  • GPT-2 (1.5B)
  • GPT-3 (175B)
  • T5
  • Megatron-LM

Technical shifts

  • Few-shot prompting replaces task-specific fine-tuning for many use cases
  • Compute budgets become the primary lever
  • The first $1M+ training runs

Market shifts

  • OpenAI's API launches; commercial LLMs become a product category
  • Anthropic founded in 2021 by Dario Amodei and team departing OpenAI

Authentication — is the document from this era?

TellMeaning
Reference to 'few-shot prompting' as a noveltyDocument dates to 2020 or shortly after.

Primary sources

  1. [1] Kaplan et al.: Scaling Laws2020-01-23
  2. [2] Brown et al.: GPT-32020-05-28

From the Almanac shop

The AI Eras — Pocket Field Guide

Ten eras of AI on a single foldable. The Almanac in your pocket.

$19Coming soon

Back to the timeline