Why Did GPT-2 Scare Dario Amodei?
Dario Amodei says GPT-2 scared him because it was the first real test of the scaling-laws hypothesis, and it made him think a much larger model would surpass humans at everything. He was not very involved in GPT-1. He pushed a larger, more expensive experiment on the bet that language models get more coherent as they grow.
He treats the jump from GPT-1 to GPT-2 as a definitive experiment with a strange prediction. Scale the model by another factor of a million, he thought, and it would be better than humans at everything, and it would determine the fate of nations, economies, and the future of humanity. About 10% of the people around him, often the ones he considered smartest, had the same reaction. About 90% shrugged, because GPT-2 was not impressive in any single area. It could translate, summarize a document, and write a few lines of code, and people still treated the limits as decisive. Hard tasks become easy within six months, and tasks that seemed impossible show up regularly within twelve months. The same pattern has kept spreading.
Dario & Daniela Amodei: "This Is Gonna Determine the Fate of Nations and the Future of Humanity” ·
From the transcript
Dario, you were there at one of the seminal moments of modern times, which is when GPT-1 and 2 suddenly burst on the scene, surprising, perhaps even surprising you. You were head of research at OpenAI. What was that moment like? Yeah, so I wasn't very involved in GPT-1, but I was very involved in the idea that we should run a larger scale, more expensive, more ambitious experiment on the idea that if we ran these language models at a larger scale, they would be better at language, they would be more coherent.
It's almost like one of these moments in scientific history where you're like, you propose a definitive experiment to just make a very strange prediction. And then if the prediction turns out to be correct, you're like, God, we're really on to something. So I see GPT-2 as the first real test of the scaling laws hypothesis. And what I found super interesting was the enormous divergence in reactions to it.