"Sparks of Artificial General Intelligence: Early experiments with GPT-4" is a report posted to arXiv on 22 March 2023 by a Microsoft Research team led by Sébastien Bubeck, who had tested an early version of GPT-4. Its central sentence: given the breadth and depth of its capabilities, GPT-4 "could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system" (Bubeck et al., 2023). The best-remembered example: asking it to draw a unicorn in TikZ, a LaTeX graphics language, and watching the drawing improve over the course of training.
The title worked as a slogan. For enthusiasts, it confirmed that the scaling hypothesis was paying off; for doomers, it was the alarm that the clock was ticking (AGI). The song turned it into a pickup line: "I see sparks of AGI in your eyes" (osmarks, n.d.).
The criticisms were methodological and rhetorical. The work was not peer reviewed, it evaluated a closed model whose training data nobody could audit (so data contamination could not be ruled out), and it was signed by a company with huge commercial stakes in that model. Moreover, as its definition of intelligence it adopted that of a 1994 editorial signed by 52 psychologists, "Mainstream Science on Intelligence," published in response to criticism of The Bell Curve; after objections, the reference disappeared in the report's fourth version (12 April 2023). For philosophy, the case is useful for discussing what counts as evidence of generality: a collection of impressive demonstrations, chosen by the authors themselves, is not a measurement, and Chollet had already warned that evaluating skill at separate tasks does not measure generality (Chollet, 2019, sec. II.1.1) (see measuring intelligence and benchmark).