Plurai unveils vibe training

- Plurai described 'vibe training', a method that distills small specialized models (SLMs) via adversarial agent swarms to speed up domain-specific evaluations and guardrail tuning. - They claim vibe-trained SLMs deliver roughly 8x faster evals and about 50% better accuracy for domain failure modes compared with large LLM judges. - Coupled with Claw-Eval findings that LLM judges miss many safety and robustness issues, teams are shifting toward continuous, specialized evals and deterministic checks. (x.com) (x.com) (x.com)