The GPT-3 moment

So what did the largest of the three actually do?

The striking part was how it could be used. A person could describe a task or show a few examples in the prompt. GPT-3 could often continue the pattern without receiving new training for that task.

The same model could attempt translation, question answering, prose, and simple code. Results varied widely. It could also invent facts, repeat biases from its data, or fail a small change to the prompt.

OpenAI then did something that mattered as much as the model. It offered GPT-3 through an , so a developer could send text to it and get a reply back without owning a copy of the model or the machines to run it.

Not to everyone. Access was by waitlist, granted in batches, and the model stayed on OpenAI's own computers. But it meant the people testing what this thing could do were no longer only the lab that built it.

A model that size is not cheap, though, and the shape of that cost is strange enough to be worth a slide of its own.

The GPT-3 moment

So what did the largest of the three actually do?

The striking part was how it could be used. A person could describe a task or show a few examples in the prompt. GPT-3 could often continue the pattern without receiving new training for that task.

A model that size is not cheap, though, and the shape of that cost is strange enough to be worth a slide of its own.