Nobody wants to admit that while partial to one company or another, brand and all that, I highly doubt that within 1-2 years we can tell the difference between model output for 90% of our tasks.

Except for the 'vibe' of some of the models with their writing or testing style I am not certain I could tell you which model is which unless asked.

Switching from model to model at the moment has been seamless without extensive 1 to 1 bench-marking but most of my tasks are never 1 to 1.

It could be like the last SOTA lab to run out of money wins

Training is expensive. That seems like an important incentive.

I agree. Which is why I think the bubble is going to pop in the nearish term.