Source item

Is ~log(15)x improvement in these benchmarks worth it for ~15x cluster scaling? In the end users will decide— by voting with their feet, and the markets— by voting to give Grok dollars It’s clearly a SOTA model, and folks who threw shade at Grok research team were mistaken https://t.co/Gz3YfnaaIK Still feels like open field with many vectors of potential improvement: expert data SFT, reasoning/RL and creative (more general) ways to generate synthetic data, efficiency in use of cluster (which will matter more the bigger the cluster is!), and user facing product Striking how quickly this was done, and with how few people. “You can’t start a new lab, it’s too late, the oligopoly is set” narrative is in shambles And of course all that ignores alternative or orthogonal approaches (eg SSI)