How a neuron fires: the 1952 Hodgkin–Huxley model, running live
−65 mVmembrane voltage, live
Press and hold to stimulateSource: Hodgkin & Huxley 1952How a neuron fires: the Hodgkin-Huxley model (1952, squid giant axon) simulated live with the standard parameters. The line is the voltage across the cell membrane over the last 39 milliseconds. It rests at minus 65 millivolts; a short input pulse every 45 milliseconds makes it spike to about plus 40 and fall back. The strip underneath shows the input current. Press and hold, or press the up arrow, to inject a steady current and make it fire repeatedly.
The FFT turns about a million multiplications into 5,120
205×less work at N = 1,024 samples
Drag sideways to change NSource: Cooley & Tukey 1965The fast Fourier transform turns about a million multiplications into 5,120. To find the frequencies in N samples, a direct DFT needs N squared complex multiplications; the radix-2 FFT (Cooley and Tukey, 1965) needs N/2 times log2 N. At N = 1,024 that is 1,048,576 against 5,120, about 205 times less work. The two bars are on a log scale. Drag or use the arrow keys to change N from 8 to about a million. Behind them, a live spectrum of a G major chord, or of your microphone, computed with an FFT every frame.
By 2019, one chip had more transistors than a chimp has neurons
2× every 2.0 yearstransistors per chip. Counts only: a transistor is not a neuron.
Drag to scrub through the yearsSource: Our World in Data, neuron countsTransistors on the largest microprocessor of each year, 1971 to 2021, on a log scale, against neuron counts across animals. A chip passed the neuron count of a fruit fly (150,000) by 1985, a honey bee (960,000) by 1989, a mouse (71 million) by 2002, a dog (2.25 billion) by 2009, a macaque (6.38 billion) by 2016 and a chimp (28 billion) by 2019. The 2021 chip has 58.2 billion, still short of a human brain's 86 billion. The count doubles about every 2.0 years. Counts, not capability. Drag or use the arrow keys to scrub the years. Data: Rupp (2022) via Our World in Data; neuron counts from the studies cited in Wikipedia's list of animals by number of neurons.
AI training compute has grown about 4× a year since 2010
4.3×per year since 2010
Drag to scrub through the yearsSource: Epoch AI · to Sep 2026AI training compute has grown about 4.29 times a year since 2010. The chart plots the computing operations used to train each of 540 notable AI models since 1950, on a log scale where each grid line is a million times more. Since 2010 the trend doubles every 5.7 months, against every 20.5 months before, when it was still called machine learning. Labelled: Perceptron Mark I (1957), AlexNet (2012), GPT-3 (2020) and the largest in the data, GPT-6 Astra (OpenAI, 2026; Epoch rates the compute estimate as likely). Drag or use the arrow keys to scrub through the years. Data: Epoch AI, CC BY 4.0.
US stocks are worth about 65× what they were in 1871, after inflation
65×after inflation, 1871–2026, price only
Hover or drag to read any yearSource: Robert Shiller · to Oct 2026US stocks are worth about 64.6 times what they were in 1871, after inflation. The chart shows the S&P Composite index, monthly since 1871, adjusted to today’s dollars, on a log scale, price only with dividends excluded. Marked: the 1929 peak, the dot-com peak, the 2009 low, covid, and the month ChatGPT launched. Hover, drag or use the arrow keys to read any year. Data: Robert Shiller.