Jon Stokes: frontier gains come from bespoke post-training data, not runaway generalization

ZeroStateReflex · x · 2026-11-19

Jon Stokes argues that public data showing most frontier-lab spend going to post-training confirms his earlier thesis: frontier model capability comes from high-effort, bespoke data collection rather than runaway generalization. Meta and xAI failing to keep pace — and Chinese open-weights models lagging months behind — suggests OpenAI, Google and Anthropic's edge is their massive live user sessions for RL training. The resulting intelligence isn't smooth and general but a 'point cloud of special cases' that big labs densify with more data.

Original post →

More from AGI Musings

AGI Musings channel →