How many parameters does GPT-3 have?

How to get close to this number

Anchor on the biological comparison people reach for. A human brain has tens of billions of neurons and vastly more connections, and language models of this era were routinely compared to it. The model sits in the same order of magnitude as the neuron count, which is where the comparison came from — and which makes the figure easy to place even if you have never read the paper.

The jump is the striking part rather than the absolute number: more than a hundred times its immediate predecessor, in a single generation. Scaling steps that large are rare in any engineering field, and the result was not a hundred-times-better version of the same thing but a system that did qualitatively different things — which is why the paper mattered.

The explanation names what happened next, and it is the part worth carrying. Parameter counts stopped being the headline almost immediately: newer models are often smaller and trained on far more data, and most laboratories no longer publish the figure at all. A metric that defined a field for two years became commercially sensitive and then obsolete, which is a reminder that the interesting number changes as fast as the technology.