“Scaling laws in AI dictate that achieving linear gains on benchmarks requires exponentially more cost in compute and data.”