Intelligence Scales as the Logarithm of Compute (& Data)

LessWrong explains that a convenient way to gauge intelligence is by the logarithm of effective compute and data. In practice AI benchmark scores tend to follow symmetric logistic curves that saturate over roughly a year, while the amount of effective compute used for training has been rising about an order of magnitude each year. Because a tenfold increase in compute corresponds to roughly fifty points on an IQ‑like scale, the article suggests that AI systems have been gaining about fifty IQ points annually—roughly matching Sam Altman’s description of GPT‑3 as a high‑school student, GPT‑4 as a college student, and GPT‑5 as a PhD‑level expert. The piece also points out that human IQ tests, such as those based on Item Response Theory, show a similar logarithmic relationship to effective training. Finally, LessWrong cites Maxim Lott’s TrackingAI project, which recorded an improvement of fifteen to twenty IQ points per year from early twenty twenty‑four to early‑mid twenty twenty‑6.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton