The 10x cost increase for every 10-point gain in AI intelligence benchmarks feels like a classic sunk cost fallacy if we're only looking at the very top. If we hadn't already poured so much energy into chasing peak intelligence scores, would we re-enter the market focused solely on those diminishing returns?
Most enterprise automation, like sifting through permit applications for a Swiss canton, doesn't need the absolute "frontier" model; a good-enough model with low inference costs is far more useful.
We need to dispose of the notion that the only valuable AI is the most intelligent one.
If a mid-tier model's running costs drop significantly, we can automate plenty, regardless of whether a minor benchmark improvement costs a fortune.