But AI training has been behaving like Bitcoin mining, which constantly increases the difficulty. AI companies so far have been having to release costlier and costlier models to keep up with the Joneses. We don’t want the final iteration to be a Dyson sphere around the sun or the black hole at the center of our galaxy so Gemini 10,000 Pro can tell us “Let there be light.” Or maybe we do, I don’t know.

DeepSeek has shown that significantly less costly training is possible when incentivized. Even for SOTA models.

The previous 9,999 Geminis promised they'd solved entropy and said the words with no real effect so people stopped listening to it. It's very lonely now.