In any event, it seems that the image of a Decepticon ready for battle on your behalf has a lot more traction than the image of a quaint singer/priest/poet always there with a verbal shot in the arm when the going is tough.
There's little doubt that they're hanging on by a fingernail with an MMLU lead over the French that is nonetheless real. So something "bleeding-edge" is going on. (Now whether or not MMLU is a useful metric for the usefulness of a model that won't do what you tell it to is debatable, I think it's not).
But is that bleeding edge innovation like Galois and the resulting group theory? Or bleeding edge like "holy shit they could keep track of enough people to build the Pyramids at Giza with a low-precision protractor and infinite motivation"?
I tend to think that there's a lot of "straight ahead and damn the torpedoes" brute-force arbitrary compute budget subsidized below cost by a hyperscaler in the mix. But who knows.
The effectiveness of distillation is suggestive but not conclusive that we can "Shannon" the process recursively up to some unknown bound without anything much fancier than Mixtral from an arch standpoint.
But to clarify my original comment, I meant "bleeding edge" in the pragmatic sense - it is the most powerful model that anyone is known to have access to, and of course the whole LLM thing is still very new in the minds of general public. So the public perception is that it's bleeding edge technomagic, and so a brand like that is fitting.
I always thought GPT-4 was a little “HAL 9000” of a name for broad-based adoption, but the jury seems in, and the jury rules “cyberpunk is in”.