AI models don't need to slow down because things are already extreme

compute is the bottleneck

openai spent millions on compute and cleared a Millennium Prize problem, so a lot may already be possible. there just isn't enough compute

it comes down to how fast hyperscalers can bring more compute online