Synthetic data breakthroughs (AlphaGo Zero-style) will lift the data bottleneck for language models
Amodei identifies data as the current scaling bottleneck but points to self-play and model-generated data — analogous to AlphaGo Zero's superhuman performance without human games — as the solution that will unlock continued scaling.